out.log

// -----// IR Dump After TosaToSCF (tosa-to-scf) //----- //
func.func @main() {
  %0 = util.unfoldable_constant dense<1> : tensor<1024x1024xi8>
  %1 = util.unfoldable_constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %c0 = arith.constant 0 : index
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq_const(%4, dense<1024> : tensor<1024x1024xi32>) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TopLevelSCFToCFG (iree-top-level-scf-to-cfg) //----- //
func.func @main() {
  %0 = util.unfoldable_constant dense<1> : tensor<1024x1024xi8>
  %1 = util.unfoldable_constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %c0 = arith.constant 0 : index
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq_const(%4, dense<1024> : tensor<1024x1024xi32>) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Inliner (inline) //----- //
module {
  func.func @main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After TosaMakeBroadcastable (tosa-make-broadcastable) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToArith (tosa-to-arith) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToTensor (tosa-to-tensor) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToLinalgExt (iree-tosa-to-linalg-ext) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaOptionalDecompositions (tosa-optional-decompositions) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaMakeBroadcastable (tosa-make-broadcastable) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToLinalgNamed (tosa-to-linalg-named) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaLayerwiseConstantFoldPass (tosa-layerwise-constant-fold) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaMakeBroadcastable (tosa-make-broadcastable) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaValidation (tosa-validate) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToLinalg (tosa-to-linalg) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToArith (tosa-to-arith) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After TosaToTensor (tosa-to-tensor) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After StripSignedness (iree-flow-strip-signedness) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After LinalgQuantizedMatmulToMatmulPass (iree-linalg-quantized-matmul-to-matmul) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After LinalgQuantizedConvToConvPass (iree-linalg-quantized-conv-to-conv) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After VerifyCompilerTOSAInputLegality (iree-tosa-verify-compiler-input-legality) //----- //
module {
  func.func @main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IREEImportPublic (iree-import-public) //----- //
module {
  func.func @main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ImportMLProgram (iree-import-ml-program) //----- //
module {
  func.func @main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After SanitizeModuleNames (iree-sanitize-module-names) //----- //
module {
  func.func @main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After mlir::iree_compiler::IREE::ABI::WrapEntryPointsPass (iree-abi-wrap-entry-points) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Inliner (inline) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SymbolDCE (symbol-dce) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After DemoteF64ToF32 (iree-util-demote-f64-to-f32) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After DetachElementwiseFromNamedOps (iree-flow-detach-elementwise-from-named-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After LinalgNamedOpConversion (linalg-named-op-conversion) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Convert1X1FilterConv2DToMatmul (iree-flow-convert-1x1-filter-conv2d-to-matmul) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After DetachElementwiseFromNamedOps (iree-flow-detach-elementwise-from-named-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After LinalgNamedOpConversion (linalg-named-op-conversion) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Convert1X1FilterConv2DToMatmul (iree-flow-convert-1x1-filter-conv2d-to-matmul) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After EraseUnusedLinalgOperands (iree-flow-erase-unused-linalg-operands) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After VerifyInputLegality (iree-verify-input-legality) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ExpandTensorShapes (iree-flow-expand-tensor-shapes) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After FixedPointIterator (iree-util-fixed-point-iterator) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After TensorPadToTensorInsertSlice (iree-flow-tensor-pad-to-tensor-insert-slice) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %c0_i32 = arith.constant 0 : i32
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = tensor.empty() : tensor<1024x1024xi32>
    %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ConvertElementwiseToLinalg (convert-elementwise-to-linalg) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After LinalgFoldUnitExtentDims (linalg-fold-unit-extent-dims) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After RaiseSpecialOps (iree-flow-raise-special-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After InterchangeGenericOps (iree-flow-interchange-generic-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ResolveShapedTypeResultDims (resolve-shaped-type-result-dims) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ConvertElementwiseToLinalg (convert-elementwise-to-linalg) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After LinalgFoldUnitExtentDims (linalg-fold-unit-extent-dims) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After FusionOfTensorOps (iree-flow-fusion-of-tensor-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After RaiseSpecialOps (iree-flow-raise-special-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After LinalgDetensorize (linalg-detensorize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After InterchangeGenericOps (iree-flow-interchange-generic-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ResolveShapedTypeResultDims (resolve-shaped-type-result-dims) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CollapseDims (iree-flow-collapse-dims) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SplitReduction (iree-flow-split-reduction-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After InterchangeGenericOps (iree-flow-interchange-generic-ops) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After FormDispatchRegions (iree-flow-form-dispatch-regions) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After FusionOfTensorOps (iree-flow-fusion-of-tensor-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CollapseDimensions (iree-flow-collapse-dimensions) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After LinalgDetensorize (linalg-detensorize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After FormDispatchWorkgroups (iree-flow-form-dispatch-workgroups) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CaptureDispatchDynamicDims (iree-flow-capture-dispatch-dynamic-dims) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CollapseDims (iree-flow-collapse-dims) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SplitReduction (iree-flow-split-reduction-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After InterchangeGenericOps (iree-flow-interchange-generic-ops) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After FormDispatchRegions (iree-flow-form-dispatch-regions) //----- //
func.func private @_main() {
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %c1 = arith.constant 1 : index
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1_1 = arith.constant 1 : index
  %4 = affine.apply affine_map<()[s0, s1, s2] -> ((s1 - s0) ceildiv s2)>()[%c0, %c1024, %c1_1]
  %c0_2 = arith.constant 0 : index
  %c1024_3 = arith.constant 1024 : index
  %c1_4 = arith.constant 1 : index
  %5 = affine.apply affine_map<()[s0, s1, s2] -> ((s1 - s0) ceildiv s2)>()[%c0_2, %c1024_3, %c1_4]
  %c0_5 = arith.constant 0 : index
  %c1_6 = arith.constant 1 : index
  %6 = affine.apply affine_map<()[s0, s1, s2] -> ((s1 - s0) ceildiv s2)>()[%c0_5, %c1, %c1_6]
  %7 = flow.dispatch.region[%4, %5, %6] -> (tensor<1024x1024xi32>) {
    %8 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.return %8 : tensor<1024x1024xi32>
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%7, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CollapseDimensions (iree-flow-collapse-dimensions) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %c0_i32 = arith.constant 0 : i32
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = tensor.empty() : tensor<1024x1024xi32>
  %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %4 = flow.dispatch.region[%c1024, %c1024, %c1] -> (tensor<1024x1024xi32>) {
    %5 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.return %5 : tensor<1024x1024xi32>
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%4, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After FormDispatchWorkgroups (iree-flow-form-dispatch-workgroups) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch.workgroups[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32> =
      (%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
    %c0_i32 = arith.constant 0 : i32
    %3 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %4 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %5 = tensor.empty() : tensor<1024x1024xi32>
    %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.dispatch.tensor.store %7, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    flow.return
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CaptureDispatchDynamicDims (iree-flow-capture-dispatch-dynamic-dims) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch.workgroups[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32> =
      (%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
    %c0_i32 = arith.constant 0 : i32
    %3 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %4 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %5 = tensor.empty() : tensor<1024x1024xi32>
    %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.dispatch.tensor.store %7, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    flow.return
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch.workgroups[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32> =
      (%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
    %c0_i32 = arith.constant 0 : i32
    %3 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %4 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %5 = tensor.empty() : tensor<1024x1024xi32>
    %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.dispatch.tensor.store %7, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    flow.return
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch.workgroups[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32> =
      (%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
    %c0_i32 = arith.constant 0 : i32
    %3 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %4 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %5 = tensor.empty() : tensor<1024x1024xi32>
    %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.dispatch.tensor.store %7, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    flow.return
  } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After InitializeEmptyTensors (iree-flow-initialize-empty-tensors) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch.workgroups[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32> =
        (%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
      %c0_i32 = arith.constant 0 : i32
      %3 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %4 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %5 = tensor.empty() : tensor<1024x1024xi32>
      %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      flow.dispatch.tensor.store %7, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      flow.return
    } count(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After OutlineDispatchRegions (iree-flow-outline-dispatch-regions) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After StripDebugOps (iree-util-strip-debug-ops) //----- //
flow.executable private @_main_dispatch_0 {
  flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  builtin.module {
    func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
      %c0_i32 = arith.constant 0 : i32
      %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %2 = tensor.empty() : tensor<1024x1024xi32>
      %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      return
    }
  }
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After DeduplicateExecutables (iree-flow-deduplicate-executables) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After CleanupTensorShapes (iree-flow-cleanup-tensor-shapes) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CleanupTensorShapes (iree-flow-cleanup-tensor-shapes) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
flow.executable private @_main_dispatch_0 {
  flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  builtin.module {
    func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
      %c0_i32 = arith.constant 0 : i32
      %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %2 = tensor.empty() : tensor<1024x1024xi32>
      %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      return
    }
  }
}

// -----// IR Dump After CSE (cse) //----- //
flow.executable private @_main_dispatch_0 {
  flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    flow.return %x, %y, %z : index, index, index
  }
  builtin.module {
    func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
      %c0_i32 = arith.constant 0 : i32
      %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %2 = tensor.empty() : tensor<1024x1024xi32>
      %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      return
    }
  }
}

// -----// IR Dump After SymbolDCE (symbol-dce) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After VerifyInput (iree-stream-verify-input) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After OutlineConstants (iree-stream-outline-constants) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
  %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
  %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
  %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
  check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  flow.executable private @_main_dispatch_0 {
    flow.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      flow.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg1: !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>, %arg2: !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>) {
        %c0_i32 = arith.constant 0 : i32
        %0 = flow.dispatch.tensor.load %arg0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %1 = flow.dispatch.tensor.load %arg1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %2 = tensor.empty() : tensor<1024x1024xi32>
        %3 = linalg.fill ins(%c0_i32 : i32) outs(%2 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %4 = linalg.matmul ins(%0, %1 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%3 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %4, %arg2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = arith.constant dense<1024> : tensor<1024x1024xi32>
    %cst_0 = arith.constant dense<1> : tensor<1024x1024xi8>
    %0 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %1 = util.optimization_barrier %cst_0 : tensor<1024x1024xi8>
    %2 = flow.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%0, %1) : (tensor<1024x1024xi8>, tensor<1024x1024xi8>) -> tensor<1024x1024xi32>
    check.expect_eq(%2, %cst) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ConvertToStream (iree-stream-conversion) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %c0_i32 = arith.constant 0 : i32
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = stream.tensor.constant : tensor<1024x1024xi32> in !stream.resource<constant> = dense<1024> : tensor<1024x1024xi32>
    %0 = stream.resource.size %cst : !stream.resource<constant>
    %1 = stream.async.transfer %cst : !stream.resource<constant>{%0} -> !stream.resource<*>{%0}
    %cst_0 = stream.tensor.constant : tensor<1024x1024xi8> in !stream.resource<constant> = dense<1> : tensor<1024x1024xi8>
    %2 = stream.resource.size %cst_0 : !stream.resource<constant>
    %3 = stream.async.transfer %cst_0 : !stream.resource<constant>{%2} -> !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %c0 = arith.constant 0 : index
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%8}
    %10 = stream.async.transfer %9 : !stream.resource<*>{%8} -> !stream.resource<external>{%8}
    %11 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%8} -> tensor<1024x1024xi32>
    %12 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %13 = stream.tensor.export %12 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%11, %13) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After VerifyLoweringToTensors (iree-stream-verify-lowering-to-tensors) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %c0_i32 = arith.constant 0 : i32
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %cst = stream.tensor.constant : tensor<1024x1024xi32> in !stream.resource<constant> = dense<1024> : tensor<1024x1024xi32>
    %0 = stream.resource.size %cst : !stream.resource<constant>
    %1 = stream.async.transfer %cst : !stream.resource<constant>{%0} -> !stream.resource<*>{%0}
    %cst_0 = stream.tensor.constant : tensor<1024x1024xi8> in !stream.resource<constant> = dense<1> : tensor<1024x1024xi8>
    %2 = stream.resource.size %cst_0 : !stream.resource<constant>
    %3 = stream.async.transfer %cst_0 : !stream.resource<constant>{%2} -> !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %c0 = arith.constant 0 : index
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%8}
    %10 = stream.async.transfer %9 : !stream.resource<*>{%8} -> !stream.resource<external>{%8}
    %11 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%8} -> tensor<1024x1024xi32>
    %12 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %13 = stream.tensor.export %12 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%11, %13) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
  %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
  %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
  %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
  %4 = util.optimization_barrier %3 : !stream.resource<*>
  %5 = util.optimization_barrier %3 : !stream.resource<*>
  %6 = stream.resource.size %4 : !stream.resource<*>
  %7 = stream.resource.size %5 : !stream.resource<*>
  %8 = stream.tensor.sizeof tensor<1024x1024xi32> : index
  %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%8}
  %10 = stream.async.transfer %9 : !stream.resource<*>{%8} -> !stream.resource<external>{%8}
  %11 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%8} -> tensor<1024x1024xi32>
  %12 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
  %13 = stream.tensor.export %12 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
  check.expect_eq(%11, %13) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
  %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
  %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
  %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
  %4 = util.optimization_barrier %3 : !stream.resource<*>
  %5 = util.optimization_barrier %3 : !stream.resource<*>
  %6 = stream.resource.size %4 : !stream.resource<*>
  %7 = stream.resource.size %5 : !stream.resource<*>
  %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
  %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
  %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
  check.expect_eq(%10, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
  %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
  %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
  %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
  %4 = util.optimization_barrier %3 : !stream.resource<*>
  %5 = util.optimization_barrier %3 : !stream.resource<*>
  %6 = stream.resource.size %4 : !stream.resource<*>
  %7 = stream.resource.size %5 : !stream.resource<*>
  %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
  %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
  %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
  check.expect_eq(%10, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
    %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
    %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
    %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
    %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
    %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
    %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
    %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
    %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After CombineInitializers (iree-util-combine-initializers) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.tensor.sizeof tensor<1024x1024xi32> : index
    %1 = stream.tensor.splat %c1024_i32 : i32 -> tensor<1024x1024xi32> in !stream.resource<*>{%0}
    %2 = stream.tensor.sizeof tensor<1024x1024xi8> : index
    %3 = stream.tensor.splat %c1_i8 : i8 -> tensor<1024x1024xi8> in !stream.resource<*>{%2}
    %4 = util.optimization_barrier %3 : !stream.resource<*>
    %5 = util.optimization_barrier %3 : !stream.resource<*>
    %6 = stream.resource.size %4 : !stream.resource<*>
    %7 = stream.resource.size %5 : !stream.resource<*>
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%4[%c0 to %6 for %6], %5[%c0 to %7 for %7]) : (!stream.resource<*>{%6}, !stream.resource<*>{%7}) -> !stream.resource<*>{%0}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    %11 = stream.async.transfer %1 : !stream.resource<*>{%0} -> !stream.resource<external>{%0}
    %12 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%0} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After EncodeHostTensors (iree-stream-encode-host-tensors) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After EncodeHostTensors (iree-stream-encode-host-tensors) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %1 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %2 = util.optimization_barrier %1 : !stream.resource<*>
  %3 = util.optimization_barrier %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.resource.size %3 : !stream.resource<*>
  %6 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%2[%c0 to %4 for %4], %3[%c0 to %5 for %5]) : (!stream.resource<*>{%4}, !stream.resource<*>{%5}) -> !stream.resource<*>{%c4194304}
  %7 = stream.async.transfer %6 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %9 = stream.async.transfer %0 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%8, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After EncodeDeviceTensors (iree-stream-encode-device-tensors) //----- //
stream.executable private @_main_dispatch_0 {
  stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
    %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
    stream.return %x, %y, %z : index, index, index
  }
  builtin.module {
    func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
      %c0_i32 = arith.constant 0 : i32
      %c0 = arith.constant 0 : index
      %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
      %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
      %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
      %5 = tensor.empty() : tensor<1024x1024xi32>
      %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
      flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      return
    }
  }
}

// -----// IR Dump After MaterializeBuiltins (iree-stream-materialize-builtins) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %1 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %2 = util.optimization_barrier %1 : !stream.resource<*>
    %3 = util.optimization_barrier %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.resource.size %3 : !stream.resource<*>
    %6 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%2[%c0 to %4 for %4], %3[%c0 to %5 for %5]) : (!stream.resource<*>{%4}, !stream.resource<*>{%5}) -> !stream.resource<*>{%c4194304}
    %7 = stream.async.transfer %6 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %9 = stream.async.transfer %0 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%8, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<*>
    %2 = util.optimization_barrier %0 : !stream.resource<*>
    %3 = stream.resource.size %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
    %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<*>
    %2 = util.optimization_barrier %0 : !stream.resource<*>
    %3 = stream.resource.size %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
    %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<*>
    %2 = util.optimization_barrier %0 : !stream.resource<*>
    %3 = stream.resource.size %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
    %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<*>
    %2 = util.optimization_barrier %0 : !stream.resource<*>
    %3 = stream.resource.size %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
    %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After MaterializeCopyOnWrite (iree-stream-materialize-copy-on-write) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After MaterializeCopyOnWrite (iree-stream-materialize-copy-on-write) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ElideAsyncCopies (iree-stream-elide-async-copies) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<*>
    %2 = util.optimization_barrier %0 : !stream.resource<*>
    %3 = stream.resource.size %1 : !stream.resource<*>
    %4 = stream.resource.size %2 : !stream.resource<*>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
    %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
    %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
    %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %10) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After EmplaceAllocations (iree-stream-emplace-allocations) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After EmplaceAllocations (iree-stream-emplace-allocations) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<*>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<*>
  %2 = util.optimization_barrier %0 : !stream.resource<*>
  %3 = stream.resource.size %1 : !stream.resource<*>
  %4 = stream.resource.size %2 : !stream.resource<*>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<*>{%3}, !stream.resource<*>{%4}) -> !stream.resource<*>{%c4194304}
  %6 = stream.async.transfer %5 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %7 = stream.tensor.export %6 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %8 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<*>{%c4194304}
  %9 = stream.async.transfer %8 : !stream.resource<*>{%c4194304} -> !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After RefineUsage (iree-stream-refine-usage) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%6, %8) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
  %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%6, %8) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
  %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%6, %8) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
  %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%6, %8) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%6, %8) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%6, %8) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%6, %8) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%1[%c0 to %3 for %3], %2[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    %8 = stream.tensor.export %7 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%6, %8) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ScheduleExecution (iree-stream-schedule-execution) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ScheduleConcurrency (iree-stream-schedule-concurrency) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ScheduleExecution (iree-stream-schedule-execution) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
    %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    stream.yield %8 : !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
    %8 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg0[%c0 to %3 for %3], %arg1[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
    %9 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
    stream.yield %8, %9 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  } => !stream.timepoint
  %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %6) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ScheduleConcurrency (iree-stream-schedule-concurrency) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
    %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    stream.yield %8 : !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
    %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
      %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
      stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    }
    stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  } => !stream.timepoint
  %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %6) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After PropagateTimepoints (iree-stream-propagate-timepoints) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %11 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %11 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %5 = stream.timepoint.immediate => !stream.timepoint
    %6 = stream.timepoint.immediate => !stream.timepoint
    %7 = stream.timepoint.immediate => !stream.timepoint
    %results_0:2, %result_timepoint_1 = stream.async.execute await(%7) => with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %11:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %12 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %13 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %12, %13 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %11#0, %11#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %8:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %9 = stream.tensor.export %8#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %10 = stream.tensor.export %8#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%10, %9) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
    %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    stream.yield %8 : !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
    %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
      %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
      stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    }
    stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  } => !stream.timepoint
  %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %6) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
    %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    stream.yield %8 : !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
    %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
      %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
      stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    }
    stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  } => !stream.timepoint
  %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %6) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
    %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
    stream.yield %8 : !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
  %1 = util.optimization_barrier %0 : !stream.resource<transient>
  %2 = util.optimization_barrier %0 : !stream.resource<transient>
  %3 = stream.resource.size %1 : !stream.resource<transient>
  %4 = stream.resource.size %2 : !stream.resource<transient>
  %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
    %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
      %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
      stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    }
    stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  } => !stream.timepoint
  %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%7, %6) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %8 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %6) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %8 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %6) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %8 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %6) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %8 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %6) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After VerifyLoweringToAsync (iree-stream-verify-lowering-to-async) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %results, %result_timepoint = stream.async.execute with() -> !stream.resource<transient>{%c1048576} {
      %8 = stream.async.splat %c1_i8 : i8 -> !stream.resource<transient>{%c1048576}
      stream.yield %8 : !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %0 = stream.timepoint.await %result_timepoint => %results : !stream.resource<transient>{%c1048576}
    %1 = util.optimization_barrier %0 : !stream.resource<transient>
    %2 = util.optimization_barrier %0 : !stream.resource<transient>
    %3 = stream.resource.size %1 : !stream.resource<transient>
    %4 = stream.resource.size %2 : !stream.resource<transient>
    %results_0:2, %result_timepoint_1 = stream.async.execute with(%1 as %arg0: !stream.resource<transient>{%3}, %2 as %arg1: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
      %8:2 = stream.async.concurrent with(%arg0 as %arg2: !stream.resource<transient>{%3}, %arg1 as %arg3: !stream.resource<transient>{%4}) -> (!stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}) {
        %9 = stream.async.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%arg2[%c0 to %3 for %3], %arg3[%c0 to %4 for %4]) : (!stream.resource<transient>{%3}, !stream.resource<transient>{%4}) -> !stream.resource<external>{%c4194304}
        %10 = stream.async.splat %c1024_i32 : i32 -> !stream.resource<external>{%c4194304}
        stream.yield %9, %10 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
      }
      stream.yield %8#0, %8#1 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    } => !stream.timepoint
    %5:2 = stream.timepoint.await %result_timepoint_1 => %results_0#1, %results_0#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %6 = stream.tensor.export %5#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %7 = stream.tensor.export %5#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%7, %6) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ScheduleAllocation (iree-stream-schedule-allocation) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After PackConstants (iree-stream-pack-constants) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After PackAllocations (iree-stream-pack-allocations) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After LayoutSlices (iree-stream-layout-slices) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After ScheduleAllocation (iree-stream-schedule-allocation) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %c0_0 = arith.constant 0 : index
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0_0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %c0_1 = arith.constant 0 : index
  %7:2 = stream.resource.alloc uninitialized : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7#0 as %arg2: !stream.resource<external>{%c4194304}, %7#1 as %arg3: !stream.resource<external>{%c4194304}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0_1 for %c4194304] : !stream.resource<external>{%c4194304}
      }
      stream.cmd.fill %c1024_i32, %arg3[%c0_1 for %c4194304] : i32 -> !stream.resource<external>{%c4194304}
    }
  } => !stream.timepoint
  %9:2 = stream.timepoint.await %8 => %7#1, %7#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %11 = stream.tensor.export %9#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%11, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After PackConstants (iree-stream-pack-constants) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %c0_0 = arith.constant 0 : index
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0_0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %c0_1 = arith.constant 0 : index
  %7:2 = stream.resource.alloc uninitialized : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7#0 as %arg2: !stream.resource<external>{%c4194304}, %7#1 as %arg3: !stream.resource<external>{%c4194304}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0_1 for %c4194304] : !stream.resource<external>{%c4194304}
      }
      stream.cmd.fill %c1024_i32, %arg3[%c0_1 for %c4194304] : i32 -> !stream.resource<external>{%c4194304}
    }
  } => !stream.timepoint
  %9:2 = stream.timepoint.await %8 => %7#1, %7#0 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %10 = stream.tensor.export %9#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %11 = stream.tensor.export %9#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%11, %10) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After PackAllocations (iree-stream-pack-allocations) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %c0_0 = arith.constant 0 : index
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0_0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %c0_1 = arith.constant 0 : index
  %7:3 = stream.resource.pack slices({
    [0, 0] = %c4194304,
    [0, 0] = %c4194304
  }) : index
  %8 = stream.resource.alloc uninitialized : !stream.resource<external>{%7#0}
  %9 = stream.resource.subview %8[%7#1] : !stream.resource<external>{%7#0} -> !stream.resource<external>{%c4194304}
  %10 = stream.resource.subview %8[%7#2] : !stream.resource<external>{%7#0} -> !stream.resource<external>{%c4194304}
  %11 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %9 as %arg2: !stream.resource<external>{%c4194304}, %10 as %arg3: !stream.resource<external>{%c4194304}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0_1 for %c4194304] : !stream.resource<external>{%c4194304}
      }
      stream.cmd.fill %c1024_i32, %arg3[%c0_1 for %c4194304] : i32 -> !stream.resource<external>{%c4194304}
    }
  } => !stream.timepoint
  %12:2 = stream.timepoint.await %11 => %10, %9 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %13 = stream.tensor.export %12#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %14 = stream.tensor.export %12#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%14, %13) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After LayoutSlices (iree-stream-layout-slices) //----- //
func.func private @_main() {
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c0 = arith.constant 0 : index
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %c0_0 = arith.constant 0 : index
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0_0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %c0_1 = arith.constant 0 : index
  %c0_2 = arith.constant 0 : index
  %c4194304_3 = arith.constant 4194304 : index
  %c4194304_4 = arith.constant 4194304 : index
  %c8388608 = arith.constant 8388608 : index
  %c8388608_5 = arith.constant 8388608 : index
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608_5}
  %8 = stream.resource.subview %7[%c0_2] : !stream.resource<external>{%c8388608_5} -> !stream.resource<external>{%c4194304}
  %9 = stream.resource.subview %7[%c4194304_4] : !stream.resource<external>{%c8388608_5} -> !stream.resource<external>{%c4194304}
  %10 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %8 as %arg2: !stream.resource<external>{%c4194304}, %9 as %arg3: !stream.resource<external>{%c4194304}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0_1 for %c4194304] : !stream.resource<external>{%c4194304}
      }
      stream.cmd.fill %c1024_i32, %arg3[%c0_1 for %c4194304] : i32 -> !stream.resource<external>{%c4194304}
    }
  } => !stream.timepoint
  %11:2 = stream.timepoint.await %10 => %9, %8 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %11#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After PropagateSubranges (iree-util-propagate-subranges) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c0_0 = arith.constant 0 : index
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %c0_1 = arith.constant 0 : index
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0_1 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %c0_2 = arith.constant 0 : index
    %c0_3 = arith.constant 0 : index
    %c4194304_4 = arith.constant 4194304 : index
    %c4194304_5 = arith.constant 4194304 : index
    %c8388608 = arith.constant 8388608 : index
    %c8388608_6 = arith.constant 8388608 : index
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608_6}
    %8 = stream.resource.subview %7[%c0_3] : !stream.resource<external>{%c8388608_6} -> !stream.resource<external>{%c4194304}
    %9 = stream.resource.subview %7[%c4194304_5] : !stream.resource<external>{%c8388608_6} -> !stream.resource<external>{%c4194304}
    %10 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %8 as %arg2: !stream.resource<external>{%c4194304}, %9 as %arg3: !stream.resource<external>{%c4194304}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0_0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0_0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0_2 for %c4194304] : !stream.resource<external>{%c4194304}
        }
        stream.cmd.fill %c1024_i32, %arg3[%c0_2 for %c4194304] : i32 -> !stream.resource<external>{%c4194304}
      }
    } => !stream.timepoint
    %11:2 = stream.timepoint.await %10 => %9, %8 : !stream.resource<external>{%c4194304}, !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %11#0 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11#1 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After VerifyLoweringToCmd (iree-stream-verify-lowering-to-cmd) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After SCFToControlFlow (convert-scf-to-cf) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SCFToControlFlow (convert-scf-to-cf) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After ElideTimepoints (iree-stream-elide-timepoints) //----- //
module attributes {iree.fixedpoint.iteration = 0 : index} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FixedPointIterator (iree-util-fixed-point-iterator) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c4194304] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseDispatchBindings (iree-stream-fuse-dispatch-bindings) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding, %arg1: !stream.binding, %arg2: !stream.binding, %arg3: index, %arg4: index, %arg5: index) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%arg3] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%arg4] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%arg5] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_0 = arith.constant 0 : index
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%c0, %c0, %c0 : index, index, index) {
          ro %arg0[%c0_0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0_0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0_0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After AnnotateDispatchArguments (iree-stream-annotate-dispatch-arguments) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}, %arg3: index {stream.values = [0 : index]}, %arg4: index {stream.values = [0 : index]}, %arg5: index {stream.values = [0 : index]}) {
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %0 = stream.binding.subspan %arg0[%arg3] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%arg4] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%arg5] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_0 = arith.constant 0 : index
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%c0, %c0, %c0 : index, index, index) {
          ro %arg0[%c0_0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0_0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0_0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After PackDispatchOperands (iree-stream-pack-dispatch-operands) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}, %arg3: i32, %arg4: i32, %arg5: i32) {
        %0 = arith.index_castui %arg3 {stream.values = [0 : index]} : i32 to index
        %1 = arith.index_castui %arg4 {stream.values = [0 : index]} : i32 to index
        %2 = arith.index_castui %arg5 {stream.values = [0 : index]} : i32 to index
        %c0_i32 = arith.constant 0 : i32
        %c0 = arith.constant 0 : index
        %3 = stream.binding.subspan %arg0[%0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %4 = stream.binding.subspan %arg1[%1] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %5 = stream.binding.subspan %arg2[%2] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %6 = flow.dispatch.tensor.load %3, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %7 = flow.dispatch.tensor.load %4, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %8 = tensor.empty() : tensor<1024x1024xi32>
        %9 = linalg.fill ins(%c0_i32 : i32) outs(%8 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %10 = linalg.matmul ins(%6, %7 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%9 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %10, %5, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %c0_i32_1 = arith.constant 0 : i32
    %c0_i32_2 = arith.constant 0 : i32
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%c0_i32, %c0_i32_1, %c0_i32_2 : i32, i32, i32) {
          ro %arg0[%c0_0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0_0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0_0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}, %arg3: i32, %arg4: i32, %arg5: i32) {
        %0 = arith.index_castui %arg3 {stream.values = [0 : index]} : i32 to index
        %1 = arith.index_castui %arg4 {stream.values = [0 : index]} : i32 to index
        %2 = arith.index_castui %arg5 {stream.values = [0 : index]} : i32 to index
        %c0_i32 = arith.constant 0 : i32
        %3 = stream.binding.subspan %arg0[%0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %4 = stream.binding.subspan %arg1[%1] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %5 = stream.binding.subspan %arg2[%2] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %6 = flow.dispatch.tensor.load %3, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %7 = flow.dispatch.tensor.load %4, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %8 = tensor.empty() : tensor<1024x1024xi32>
        %9 = linalg.fill ins(%c0_i32 : i32) outs(%8 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %10 = linalg.matmul ins(%6, %7 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%9 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %10, %5, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_i32 = arith.constant 0 : i32
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1](%c0_i32, %c0_i32, %c0_i32 : i32, i32, i32) {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldUniformOperands (iree-stream-fold-uniform-operands) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0_i32 = arith.constant 0 : i32
        %0 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %1 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %2 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %c0_i32_0 = arith.constant 0 : i32
        %3 = stream.binding.subspan %arg0[%0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %4 = stream.binding.subspan %arg1[%1] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %5 = stream.binding.subspan %arg2[%2] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %6 = flow.dispatch.tensor.load %3, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %7 = flow.dispatch.tensor.load %4, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %8 = tensor.empty() : tensor<1024x1024xi32>
        %9 = linalg.fill ins(%c0_i32_0 : i32) outs(%8 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %10 = linalg.matmul ins(%6, %7 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%9 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %10, %5, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_i32 = arith.constant 0 : i32
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After MemoizeChannels (iree-stream-memoize-channels) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0_i32 = arith.constant 0 : i32
        %0 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %1 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %2 = arith.index_castui %c0_i32 {stream.values = [0 : index]} : i32 to index
        %c0_i32_0 = arith.constant 0 : i32
        %3 = stream.binding.subspan %arg0[%0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %4 = stream.binding.subspan %arg1[%1] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %5 = stream.binding.subspan %arg2[%2] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %6 = flow.dispatch.tensor.load %3, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %7 = flow.dispatch.tensor.load %4, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %8 = tensor.empty() : tensor<1024x1024xi32>
        %9 = linalg.fill ins(%c0_i32_0 : i32) outs(%8 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %10 = linalg.matmul ins(%6, %7 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%9 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %10, %5, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %c0_i32 = arith.constant 0 : i32
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After CSE (cse) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After IPO (iree-util-ipo) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After SymbolDCE (symbol-dce) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After Canonicalizer (canonicalize) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func @main() attributes {iree.abi.stub} {
  call @_main() : () -> ()
  return
}

// -----// IR Dump After SimplifyGlobalAccesses (iree-util-simplify-global-accesses) //----- //
func.func private @_main() {
  %c8388608 = arith.constant 8388608 : index
  %c0 = arith.constant 0 : index
  %c4194304 = arith.constant 4194304 : index
  %c1048576 = arith.constant 1048576 : index
  %c1_i8 = arith.constant 1 : i8
  %c1024_i32 = arith.constant 1024 : i32
  %c1024 = arith.constant 1024 : index
  %c1 = arith.constant 1 : index
  %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
  %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
    stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
  } => !stream.timepoint
  %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
  %3 = util.optimization_barrier %2 : !stream.resource<transient>
  %4 = util.optimization_barrier %2 : !stream.resource<transient>
  %5 = stream.resource.size %3 : !stream.resource<transient>
  %6 = stream.resource.size %4 : !stream.resource<transient>
  %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
  %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
    stream.cmd.concurrent {
      stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
        ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
        ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
        wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
      }
      stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
    }
  } => !stream.timepoint
  %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
  %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
  %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
  check.expect_eq(%13, %12) : tensor<1024x1024xi32>
  return
}

// -----// IR Dump After ApplyPatterns (iree-util-apply-patterns) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FoldGlobals (iree-util-fold-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After FuseGlobals (iree-util-fuse-globals) //----- //
module {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After mlir::iree_compiler::IREE::HAL::AssignTargetDevicesPass (iree-hal-assign-target-devices) //----- //
#executable_target_vulkan_spirv_fb = #hal.executable.target<"vulkan", "vulkan-spirv-fb", {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, Float16, Int16, Int8, StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16, StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8, GroupNonUniform, GroupNonUniformVote, GroupNonUniformArithmetic, GroupNonUniformBallot, GroupNonUniformShuffle, GroupNonUniformShuffleRelative, GroupNonUniformClustered, GroupNonUniformQuad, VariablePointers, VariablePointersStorageBuffer], [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class, SPV_KHR_variable_pointers]>, api=Vulkan, ARM:IntegratedGPU, #spirv.resource_limits<max_compute_shared_memory_size = 32768, max_compute_workgroup_invocations = 512, max_compute_workgroup_size = [512, 512, 512], subgroup_size = 16, cooperative_matrix_properties_nv = []>>}>
#device_target_vulkan = #hal.device.target<"vulkan", {executable_targets = [#executable_target_vulkan_spirv_fb], legacy_sync}>
module attributes {hal.device.targets = [#device_target_vulkan]} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After mlir::iree_compiler::IREE::HAL::VerifyTargetEnvironmentPass (iree-hal-verify-target-environment) //----- //
#executable_target_vulkan_spirv_fb = #hal.executable.target<"vulkan", "vulkan-spirv-fb", {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, Float16, Int16, Int8, StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16, StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8, GroupNonUniform, GroupNonUniformVote, GroupNonUniformArithmetic, GroupNonUniformBallot, GroupNonUniformShuffle, GroupNonUniformShuffleRelative, GroupNonUniformClustered, GroupNonUniformQuad, VariablePointers, VariablePointersStorageBuffer], [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class, SPV_KHR_variable_pointers]>, api=Vulkan, ARM:IntegratedGPU, #spirv.resource_limits<max_compute_shared_memory_size = 32768, max_compute_workgroup_invocations = 512, max_compute_workgroup_size = [512, 512, 512], subgroup_size = 16, cooperative_matrix_properties_nv = []>>}>
#device_target_vulkan = #hal.device.target<"vulkan", {executable_targets = [#executable_target_vulkan_spirv_fb], legacy_sync}>
module attributes {hal.device.targets = [#device_target_vulkan]} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  stream.executable private @_main_dispatch_0 {
    stream.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 workgroups(%arg0: index, %arg1: index, %arg2: index) -> (index, index, index) {
      %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg0, %arg1, %arg2
      stream.return %x, %y, %z : index, index, index
    }
    builtin.module {
      func.func @_main_dispatch_0_matmul_1024x1024x1024(%arg0: !stream.binding {stream.alignment = 64 : index}, %arg1: !stream.binding {stream.alignment = 64 : index}, %arg2: !stream.binding {stream.alignment = 64 : index}) {
        %c0 = arith.constant 0 : index
        %c0_i32 = arith.constant 0 : i32
        %0 = stream.binding.subspan %arg0[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %1 = stream.binding.subspan %arg1[%c0] : !stream.binding -> !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
        %2 = stream.binding.subspan %arg2[%c0] : !stream.binding -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
        %5 = tensor.empty() : tensor<1024x1024xi32>
        %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
        flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        return
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        }
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After mlir::iree_compiler::IREE::HAL::(anonymous namespace)::MaterializeInterfacesPass (iree-hal-materialize-interfaces) //----- //
#executable_target_vulkan_spirv_fb = #hal.executable.target<"vulkan", "vulkan-spirv-fb", {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, Float16, Int16, Int8, StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16, StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8, GroupNonUniform, GroupNonUniformVote, GroupNonUniformArithmetic, GroupNonUniformBallot, GroupNonUniformShuffle, GroupNonUniformShuffleRelative, GroupNonUniformClustered, GroupNonUniformQuad, VariablePointers, VariablePointersStorageBuffer], [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class, SPV_KHR_variable_pointers]>, api=Vulkan, ARM:IntegratedGPU, #spirv.resource_limits<max_compute_shared_memory_size = 32768, max_compute_workgroup_invocations = 512, max_compute_workgroup_size = [512, 512, 512], subgroup_size = 16, cooperative_matrix_properties_nv = []>>}>
#pipeline_layout = #hal.pipeline.layout<push_constants = 0, sets = [<0, bindings = [<0, storage_buffer, ReadOnly>, <1, storage_buffer, ReadOnly>, <2, storage_buffer>]>]>
#device_target_vulkan = #hal.device.target<"vulkan", {executable_targets = [#executable_target_vulkan_spirv_fb], legacy_sync}>
module attributes {hal.device.targets = [#device_target_vulkan]} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  hal.executable private @_main_dispatch_0 {
    hal.executable.variant public @vulkan_spirv_fb, target = #executable_target_vulkan_spirv_fb {
      hal.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 ordinal(0) layout(#pipeline_layout) {
      ^bb0(%arg0: !hal.device, %arg1: index, %arg2: index, %arg3: index):
        %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg1, %arg2, %arg3
        hal.return %x, %y, %z : index, index, index
      }
      builtin.module {
        func.func @_main_dispatch_0_matmul_1024x1024x1024() {
          %c0 = arith.constant 0 : index
          %c0_i32 = arith.constant 0 : i32
          %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
          %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
          %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
          %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
          %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
          %5 = tensor.empty() : tensor<1024x1024xi32>
          %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
          %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
          flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
          return
        }
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        } attributes {hal.interface.bindings = [#hal.interface.binding<0, 0>, #hal.interface.binding<0, 1>, #hal.interface.binding<0, 2>]}
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After mlir::iree_compiler::IREE::HAL::DumpExecutableSourcesPass (iree-hal-dump-executable-sources) //----- //
#executable_target_vulkan_spirv_fb = #hal.executable.target<"vulkan", "vulkan-spirv-fb", {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, Float16, Int16, Int8, StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16, StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8, GroupNonUniform, GroupNonUniformVote, GroupNonUniformArithmetic, GroupNonUniformBallot, GroupNonUniformShuffle, GroupNonUniformShuffleRelative, GroupNonUniformClustered, GroupNonUniformQuad, VariablePointers, VariablePointersStorageBuffer], [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class, SPV_KHR_variable_pointers]>, api=Vulkan, ARM:IntegratedGPU, #spirv.resource_limits<max_compute_shared_memory_size = 32768, max_compute_workgroup_invocations = 512, max_compute_workgroup_size = [512, 512, 512], subgroup_size = 16, cooperative_matrix_properties_nv = []>>}>
#pipeline_layout = #hal.pipeline.layout<push_constants = 0, sets = [<0, bindings = [<0, storage_buffer, ReadOnly>, <1, storage_buffer, ReadOnly>, <2, storage_buffer>]>]>
#device_target_vulkan = #hal.device.target<"vulkan", {executable_targets = [#executable_target_vulkan_spirv_fb], legacy_sync}>
module attributes {hal.device.targets = [#device_target_vulkan]} {
  func.func @main() attributes {iree.abi.stub} {
    call @_main() : () -> ()
    return
  }
  hal.executable private @_main_dispatch_0 {
    hal.executable.variant public @vulkan_spirv_fb, target = #executable_target_vulkan_spirv_fb {
      hal.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 ordinal(0) layout(#pipeline_layout) {
      ^bb0(%arg0: !hal.device, %arg1: index, %arg2: index, %arg3: index):
        %x, %y, %z = flow.dispatch.workgroup_count_from_dag_root %arg1, %arg2, %arg3
        hal.return %x, %y, %z : index, index, index
      }
      builtin.module {
        func.func @_main_dispatch_0_matmul_1024x1024x1024() {
          %c0 = arith.constant 0 : index
          %c0_i32 = arith.constant 0 : i32
          %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
          %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
          %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
          %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
          %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
          %5 = tensor.empty() : tensor<1024x1024xi32>
          %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
          %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
          flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
          return
        }
      }
    }
  }
  func.func private @_main() {
    %c8388608 = arith.constant 8388608 : index
    %c0 = arith.constant 0 : index
    %c4194304 = arith.constant 4194304 : index
    %c1048576 = arith.constant 1048576 : index
    %c1_i8 = arith.constant 1 : i8
    %c1024_i32 = arith.constant 1024 : i32
    %c1024 = arith.constant 1024 : index
    %c1 = arith.constant 1 : index
    %0 = stream.resource.alloc uninitialized : !stream.resource<transient>{%c1048576}
    %1 = stream.cmd.execute with(%0 as %arg0: !stream.resource<transient>{%c1048576}) {
      stream.cmd.fill %c1_i8, %arg0[%c0 for %c1048576] : i8 -> !stream.resource<transient>{%c1048576}
    } => !stream.timepoint
    %2 = stream.timepoint.await %1 => %0 : !stream.resource<transient>{%c1048576}
    %3 = util.optimization_barrier %2 : !stream.resource<transient>
    %4 = util.optimization_barrier %2 : !stream.resource<transient>
    %5 = stream.resource.size %3 : !stream.resource<transient>
    %6 = stream.resource.size %4 : !stream.resource<transient>
    %7 = stream.resource.alloc uninitialized : !stream.resource<external>{%c8388608}
    %8 = stream.cmd.execute with(%3 as %arg0: !stream.resource<transient>{%5}, %4 as %arg1: !stream.resource<transient>{%6}, %7 as %arg2: !stream.resource<external>{%c8388608}) {
      stream.cmd.concurrent {
        stream.cmd.dispatch @_main_dispatch_0::@_main_dispatch_0_matmul_1024x1024x1024[%c1024, %c1024, %c1] {
          ro %arg0[%c0 for %5] : !stream.resource<transient>{%5},
          ro %arg1[%c0 for %6] : !stream.resource<transient>{%6},
          wo %arg2[%c0 for %c8388608] : !stream.resource<external>{%c8388608}
        } attributes {hal.interface.bindings = [#hal.interface.binding<0, 0>, #hal.interface.binding<0, 1>, #hal.interface.binding<0, 2>]}
        stream.cmd.fill %c1024_i32, %arg2[%c4194304 for %c4194304] : i32 -> !stream.resource<external>{%c8388608}
      }
    } => !stream.timepoint
    %9 = stream.timepoint.await %8 => %7 : !stream.resource<external>{%c8388608}
    %10 = stream.resource.subview %9[%c4194304] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %11 = stream.resource.subview %9[%c0] : !stream.resource<external>{%c8388608} -> !stream.resource<external>{%c4194304}
    %12 = stream.tensor.export %10 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    %13 = stream.tensor.export %11 : tensor<1024x1024xi32> in !stream.resource<external>{%c4194304} -> tensor<1024x1024xi32>
    check.expect_eq(%13, %12) : tensor<1024x1024xi32>
    return
  }
}


// -----// IR Dump After TypePropagation (iree-codegen-type-propagation) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %5 = tensor.empty() : tensor<1024x1024xi32>
  %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  return
}

// -----// IR Dump After BufferizeCopyOnlyDispatches (iree-codegen-bufferize-copy-only-dispatches) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
    %5 = tensor.empty() : tensor<1024x1024xi32>
    %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
    flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    return
  }
}

// -----// IR Dump After DecomposeSoftmax (iree-linalg-ext-decompose-softmax) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %5 = tensor.empty() : tensor<1024x1024xi32>
  %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  return
}

// -----// IR Dump After RematerializeParallelOps (iree-codegen-rematerialize-parallel-ops) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %3 = flow.dispatch.tensor.load %0, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %4 = flow.dispatch.tensor.load %1, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x1024xi8>
  %5 = tensor.empty() : tensor<1024x1024xi32>
  %6 = linalg.fill ins(%c0_i32 : i32) outs(%5 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  %7 = linalg.matmul ins(%3, %4 : tensor<1024x1024xi8>, tensor<1024x1024xi8>) outs(%6 : tensor<1024x1024xi32>) -> tensor<1024x1024xi32>
  flow.dispatch.tensor.store %7, %2, offsets = [0, 0], sizes = [1024, 1024], strides = [1, 1] : tensor<1024x1024xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  return
}

// -----// IR Dump After TileAndDistributeToWorkgroups (iree-codegen-tile-and-distribute-to-workgroups) //----- //
hal.executable.variant public @vulkan_spirv_fb, target = <"vulkan", "vulkan-spirv-fb", {spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Shader, Float16, Int16, Int8, StorageBuffer16BitAccess, StorageUniform16, StoragePushConstant16, StorageBuffer8BitAccess, UniformAndStorageBuffer8BitAccess, StoragePushConstant8, GroupNonUniform, GroupNonUniformVote, GroupNonUniformArithmetic, GroupNonUniformBallot, GroupNonUniformShuffle, GroupNonUniformShuffleRelative, GroupNonUniformClustered, GroupNonUniformQuad, VariablePointers, VariablePointersStorageBuffer], [SPV_KHR_16bit_storage, SPV_KHR_8bit_storage, SPV_KHR_storage_buffer_storage_class, SPV_KHR_variable_pointers]>, api=Vulkan, ARM:IntegratedGPU, #spirv.resource_limits<max_compute_shared_memory_size = 32768, max_compute_workgroup_invocations = 512, max_compute_workgroup_size = [512, 512, 512], subgroup_size = 16, cooperative_matrix_properties_nv = []>>}> {
  hal.executable.export public @_main_dispatch_0_matmul_1024x1024x1024 ordinal(0) layout(#hal.pipeline.layout<push_constants = 0, sets = [<0, bindings = [<0, storage_buffer, ReadOnly>, <1, storage_buffer, ReadOnly>, <2, storage_buffer>]>]>) attributes {translation_info = #iree_codegen.translation_info<SPIRVBaseVectorize>, workgroup_size = [8 : index, 2 : index, 1 : index]} {
  ^bb0(%arg0: !hal.device, %arg1: index, %arg2: index, %arg3: index):
    %c32 = arith.constant 32 : index
    %c128 = arith.constant 128 : index
    %c1 = arith.constant 1 : index
    hal.return %c32, %c128, %c1 : index, index, index
  }
  builtin.module {
    func.func @_main_dispatch_0_matmul_1024x1024x1024() {
      %c8 = arith.constant 8 : index
      %c32 = arith.constant 32 : index
      %c1024 = arith.constant 1024 : index
      %c0 = arith.constant 0 : index
      %c0_i32 = arith.constant 0 : i32
      %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
      %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
      %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      %workgroup_id_x = hal.interface.workgroup.id[0] : index
      %workgroup_count_x = hal.interface.workgroup.count[0] : index
      %workgroup_id_y = hal.interface.workgroup.id[1] : index
      %workgroup_count_y = hal.interface.workgroup.count[1] : index
      %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
      %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
      scf.for %arg0 = %3 to %c1024 step %4 {
        %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
        %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
        scf.for %arg1 = %5 to %c1024 step %6 {
          %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [%c8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<?x1024xi8>
          %8 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, %c32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x?xi8>
          %9 = tensor.empty() : tensor<8x32xi32>
          %cast = tensor.cast %9 : tensor<8x32xi32> to tensor<?x?xi32>
          %10 = linalg.fill ins(%c0_i32 : i32) outs(%cast : tensor<?x?xi32>) -> tensor<?x?xi32>
          %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%7, %8 : tensor<?x1024xi8>, tensor<1024x?xi8>) outs(%10 : tensor<?x?xi32>) -> tensor<?x?xi32>
          flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [%c8, %c32], strides = [1, 1] : tensor<?x?xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
        }
      }
      return
    }
  }
}

// -----// IR Dump After FuseTensorPadWithConsumer (iree-codegen-fuse-tensor-pad-with-consumer) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c8 = arith.constant 8 : index
  %c32 = arith.constant 32 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [%c8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<?x1024xi8>
      %8 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, %c32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x?xi8>
      %9 = tensor.empty() : tensor<8x32xi32>
      %cast = tensor.cast %9 : tensor<8x32xi32> to tensor<?x?xi32>
      %10 = linalg.fill ins(%c0_i32 : i32) outs(%cast : tensor<?x?xi32>) -> tensor<?x?xi32>
      %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%7, %8 : tensor<?x1024xi8>, tensor<1024x?xi8>) outs(%10 : tensor<?x?xi32>) -> tensor<?x?xi32>
      flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [%c8, %c32], strides = [1, 1] : tensor<?x?xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After ConvertToDestinationPassingStyle (iree-codegen-convert-to-destination-passing-style) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c8 = arith.constant 8 : index
  %c32 = arith.constant 32 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [%c8, %c32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<?x?xi32>
      %cast = tensor.cast %7 : tensor<?x?xi32> to tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [%c8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<?x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, %c32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x?xi8>
      %cast_0 = tensor.cast %cast : tensor<8x32xi32> to tensor<?x?xi32>
      %10 = linalg.fill ins(%c0_i32 : i32) outs(%cast_0 : tensor<?x?xi32>) -> tensor<?x?xi32>
      %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<?x1024xi8>, tensor<1024x?xi8>) outs(%10 : tensor<?x?xi32>) -> tensor<?x?xi32>
      flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [%c8, %c32], strides = [1, 1] : tensor<?x?xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
        %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
        flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
        %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
        flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After FoldAffineMinInDistributedLoops (iree-codegen-fold-affinemin-in-distributed-loops) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
      %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
      flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After ResolveShapedTypeResultDims (resolve-shaped-type-result-dims) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
        %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
        flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
        %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
        flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
        %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
        flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After SPIRVCreateFastSlowPath (iree-spirv-create-fast-slow-path) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = linalg.fill ins(%c0_i32 : i32) outs(%7 : tensor<8x32xi32>) -> tensor<8x32xi32>
      %11 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%8, %9 : tensor<8x1024xi8>, tensor<1024x32xi8>) outs(%10 : tensor<8x32xi32>) -> tensor<8x32xi32>
      flow.dispatch.tensor.store %11, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After SPIRVTile (iree-spirv-tile) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = linalg.fill ins(%c0_i32 : i32) outs(%extracted_slice : tensor<4x4xi32>) -> tensor<4x4xi32>
          %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
            %extracted_slice_0 = tensor.extract_slice %8[%arg2, %arg6] [4, 4] [1, 1] : tensor<8x1024xi8> to tensor<4x4xi8>
            %extracted_slice_1 = tensor.extract_slice %9[%arg6, %arg4] [4, 4] [1, 1] : tensor<1024x32xi8> to tensor<4x4xi8>
            %14 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%extracted_slice_0, %extracted_slice_1 : tensor<4x4xi8>, tensor<4x4xi8>) outs(%arg7 : tensor<4x4xi32>) -> tensor<4x4xi32>
            scf.yield %14 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
          %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %12 = linalg.fill ins(%c0_i32 : i32) outs(%extracted_slice : tensor<4x4xi32>) -> tensor<4x4xi32>
            %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
              %extracted_slice_0 = tensor.extract_slice %8[%arg2, %arg6] [4, 4] [1, 1] : tensor<8x1024xi8> to tensor<4x4xi8>
              %extracted_slice_1 = tensor.extract_slice %9[%arg6, %arg4] [4, 4] [1, 1] : tensor<1024x32xi8> to tensor<4x4xi8>
              %14 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%extracted_slice_0, %extracted_slice_1 : tensor<4x4xi8>, tensor<4x4xi8>) outs(%arg7 : tensor<4x4xi32>) -> tensor<4x4xi32>
              scf.yield %14 : tensor<4x4xi32>
            }
            %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %11 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
      %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
      scf.for %arg1 = %5 to %c1024 step %6 {
        %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
          %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %12 = linalg.fill ins(%c0_i32 : i32) outs(%extracted_slice : tensor<4x4xi32>) -> tensor<4x4xi32>
            %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
              %extracted_slice_0 = tensor.extract_slice %8[%arg2, %arg6] [4, 4] [1, 1] : tensor<8x1024xi8> to tensor<4x4xi8>
              %extracted_slice_1 = tensor.extract_slice %9[%arg6, %arg4] [4, 4] [1, 1] : tensor<1024x32xi8> to tensor<4x4xi8>
              %14 = linalg.matmul {lowering_config = #iree_codegen.lowering_config<tile_sizes = [[8, 32], [4, 4], [0, 0, 4]]>} ins(%extracted_slice_0, %extracted_slice_1 : tensor<4x4xi8>, tensor<4x4xi8>) outs(%arg7 : tensor<4x4xi32>) -> tensor<4x4xi32>
              scf.yield %14 : tensor<4x4xi32>
            }
            %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %11 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

--- After vectorization ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.transfer_write %cst, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
          %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
            %extracted_slice_0 = tensor.extract_slice %8[%arg2, %arg6] [4, 4] [1, 1] : tensor<8x1024xi8> to tensor<4x4xi8>
            %extracted_slice_1 = tensor.extract_slice %9[%arg6, %arg4] [4, 4] [1, 1] : tensor<1024x32xi8> to tensor<4x4xi8>
            %14 = vector.transfer_read %extracted_slice_0[%c0, %c0], %c0_i8 {in_bounds = [true, true, true], permutation_map = affine_map<(d0, d1) -> (d0, 0, d1)>} : tensor<4x4xi8>, vector<4x4x4xi8>
            %15 = vector.transfer_read %extracted_slice_1[%c0, %c0], %c0_i8 {in_bounds = [true, true, true], permutation_map = affine_map<(d0, d1) -> (0, d1, d0)>} : tensor<4x4xi8>, vector<4x4x4xi8>
            %16 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<4x4xi32>
            %17 = arith.extsi %14 : vector<4x4x4xi8> to vector<4x4x4xi32>
            %18 = arith.extsi %15 : vector<4x4x4xi8> to vector<4x4x4xi32>
            %19 = arith.muli %17, %18 : vector<4x4x4xi32>
            %20 = vector.multi_reduction <add>, %19, %16 [2] : vector<4x4x4xi32> to vector<4x4xi32>
            %21 = vector.transfer_write %20, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
            scf.yield %21 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After peephole optimization ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.transfer_write %cst, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
          %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
            %extracted_slice_0 = tensor.extract_slice %8[%arg2, %arg6] [4, 4] [1, 1] : tensor<8x1024xi8> to tensor<4x4xi8>
            %extracted_slice_1 = tensor.extract_slice %9[%arg6, %arg4] [4, 4] [1, 1] : tensor<1024x32xi8> to tensor<4x4xi8>
            %14 = vector.transfer_read %extracted_slice_0[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : tensor<4x4xi8>, vector<4x4xi8>
            %15 = vector.transfer_read %extracted_slice_1[%c0, %c0], %c0_i8 {in_bounds = [true, true]} : tensor<4x4xi8>, vector<4x4xi8>
            %16 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<4x4xi32>
            %17 = arith.extsi %14 : vector<4x4xi8> to vector<4x4xi32>
            %18 = arith.extsi %15 : vector<4x4xi8> to vector<4x4xi32>
            %19 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %17, %18, %16 : vector<4x4xi32>, vector<4x4xi32> into vector<4x4xi32>
            %20 = vector.transfer_write %19, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
            scf.yield %20 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After folding tensor extract/insert slice ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.transfer_write %cst, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
          %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
            %14 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<4x4xi8>
            %15 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<4x4xi8>
            %16 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<4x4xi32>
            %17 = arith.extsi %14 : vector<4x4xi8> to vector<4x4xi32>
            %18 = arith.extsi %15 : vector<4x4xi8> to vector<4x4xi32>
            %19 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %17, %18, %16 : vector<4x4xi32>, vector<4x4xi32> into vector<4x4xi32>
            %20 = vector.transfer_write %19, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
            scf.yield %20 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After lowering multi_reduction ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.transfer_write %cst, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
          %13 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %12) -> (tensor<4x4xi32>) {
            %14 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<4x4xi8>
            %15 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<4x4xi8>
            %16 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<4x4xi32>
            %17 = arith.extsi %14 : vector<4x4xi8> to vector<4x4xi32>
            %18 = arith.extsi %15 : vector<4x4xi8> to vector<4x4xi32>
            %19 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %17, %18, %16 : vector<4x4xi32>, vector<4x4xi32> into vector<4x4xi32>
            %20 = vector.transfer_write %19, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<4x4xi32>, tensor<4x4xi32>
            scf.yield %20 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %13 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {4, 4, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
JAKUB: #87
JAKUB: #94
JAKUB: #103
JAKUB: #109
JAKUB: #77
JAKUB: #81
JAKUB: lowerToInnerPro: 1
JAKUB: bounds: {1, 1, 4}
JAKUB: nativeSize: {1, 1, 4}
--- After unrolling vector ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %cst = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.extract_strided_slice %cst {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %13 = vector.transfer_write %12, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %14 = vector.extract_strided_slice %cst {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %15 = vector.transfer_write %14, %13[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %16 = vector.extract_strided_slice %cst {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %17 = vector.transfer_write %16, %15[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %18 = vector.extract_strided_slice %cst {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %19 = vector.transfer_write %18, %17[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %20 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %19) -> (tensor<4x4xi32>) {
            %21 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %22 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
            %23 = vector.transfer_read %8[%22, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %24 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
            %25 = vector.transfer_read %8[%24, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %26 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
            %27 = vector.transfer_read %8[%26, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %35 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %36 = vector.transfer_read %arg7[%c1, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %37 = vector.transfer_read %arg7[%c2, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %38 = vector.transfer_read %arg7[%c3, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %39 = arith.extsi %21 : vector<1x4xi8> to vector<1x4xi32>
            %40 = arith.extsi %23 : vector<1x4xi8> to vector<1x4xi32>
            %41 = arith.extsi %25 : vector<1x4xi8> to vector<1x4xi32>
            %42 = arith.extsi %27 : vector<1x4xi8> to vector<1x4xi32>
            %43 = arith.extsi %28 : vector<1x4xi8> to vector<1x4xi32>
            %44 = vector.insert_strided_slice %43, %cst {offsets = [0, 0], strides = [1, 1]} : vector<1x4xi32> into vector<4x4xi32>
            %45 = arith.extsi %30 : vector<1x4xi8> to vector<1x4xi32>
            %46 = vector.insert_strided_slice %45, %44 {offsets = [1, 0], strides = [1, 1]} : vector<1x4xi32> into vector<4x4xi32>
            %47 = arith.extsi %32 : vector<1x4xi8> to vector<1x4xi32>
            %48 = vector.insert_strided_slice %47, %46 {offsets = [2, 0], strides = [1, 1]} : vector<1x4xi32> into vector<4x4xi32>
            %49 = arith.extsi %34 : vector<1x4xi8> to vector<1x4xi32>
            %50 = vector.insert_strided_slice %49, %48 {offsets = [3, 0], strides = [1, 1]} : vector<1x4xi32> into vector<4x4xi32>
            %51 = vector.extract_strided_slice %50 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %52 = vector.extract_strided_slice %35 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %53 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %51, %52 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %54 = vector.extract_strided_slice %50 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %55 = vector.extract_strided_slice %35 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %56 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %54, %55 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %57 = vector.extract_strided_slice %50 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %58 = vector.extract_strided_slice %35 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %59 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %57, %58 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %60 = vector.extract_strided_slice %50 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %61 = vector.extract_strided_slice %35 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %62 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %60, %61 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %63 = vector.extract_strided_slice %50 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %64 = vector.extract_strided_slice %36 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %65 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %63, %64 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %66 = vector.extract_strided_slice %50 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %67 = vector.extract_strided_slice %36 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %68 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %66, %67 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %69 = vector.extract_strided_slice %50 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %70 = vector.extract_strided_slice %36 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %71 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %69, %70 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %72 = vector.extract_strided_slice %50 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %73 = vector.extract_strided_slice %36 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %74 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %72, %73 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %75 = vector.extract_strided_slice %50 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %76 = vector.extract_strided_slice %37 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %77 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %75, %76 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %78 = vector.extract_strided_slice %50 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %79 = vector.extract_strided_slice %37 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %80 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %78, %79 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %81 = vector.extract_strided_slice %50 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %82 = vector.extract_strided_slice %37 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %83 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %81, %82 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %84 = vector.extract_strided_slice %50 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %85 = vector.extract_strided_slice %37 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %86 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %84, %85 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %87 = vector.extract_strided_slice %50 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %88 = vector.extract_strided_slice %38 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %89 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %87, %88 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %90 = vector.extract_strided_slice %50 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %91 = vector.extract_strided_slice %38 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %92 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %90, %91 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %93 = vector.extract_strided_slice %50 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %94 = vector.extract_strided_slice %38 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %95 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %93, %94 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %96 = vector.extract_strided_slice %50 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi32> to vector<4x1xi32>
            %97 = vector.extract_strided_slice %38 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %98 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %96, %97 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %99 = vector.insert_strided_slice %53, %cst {offsets = [0, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %100 = vector.insert_strided_slice %56, %99 {offsets = [0, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %101 = vector.insert_strided_slice %59, %100 {offsets = [0, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %102 = vector.insert_strided_slice %62, %101 {offsets = [0, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %103 = vector.insert_strided_slice %65, %102 {offsets = [1, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %104 = vector.insert_strided_slice %68, %103 {offsets = [1, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %105 = vector.insert_strided_slice %71, %104 {offsets = [1, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %106 = vector.insert_strided_slice %74, %105 {offsets = [1, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %107 = vector.insert_strided_slice %77, %106 {offsets = [2, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %108 = vector.insert_strided_slice %80, %107 {offsets = [2, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %109 = vector.insert_strided_slice %83, %108 {offsets = [2, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %110 = vector.insert_strided_slice %86, %109 {offsets = [2, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %111 = vector.insert_strided_slice %89, %110 {offsets = [3, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %112 = vector.insert_strided_slice %92, %111 {offsets = [3, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %113 = vector.insert_strided_slice %95, %112 {offsets = [3, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %114 = vector.insert_strided_slice %98, %113 {offsets = [3, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %115 = vector.extract_strided_slice %114 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %116 = vector.transfer_write %115, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %117 = vector.extract_strided_slice %114 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %118 = vector.transfer_write %117, %116[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %119 = vector.extract_strided_slice %114 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %120 = vector.transfer_write %119, %118[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %121 = vector.extract_strided_slice %114 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %122 = vector.transfer_write %121, %120[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            scf.yield %122 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %20 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After Jakub cleanup extract/insert ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x4xi8>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %cst_0 = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.extract_strided_slice %cst_0 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %13 = vector.transfer_write %12, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %14 = vector.extract_strided_slice %cst_0 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %15 = vector.transfer_write %14, %13[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %16 = vector.extract_strided_slice %cst_0 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %17 = vector.transfer_write %16, %15[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %18 = vector.extract_strided_slice %cst_0 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %19 = vector.transfer_write %18, %17[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %20 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %19) -> (tensor<4x4xi32>) {
            %21 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %22 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
            %23 = vector.transfer_read %8[%22, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %24 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
            %25 = vector.transfer_read %8[%24, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %26 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
            %27 = vector.transfer_read %8[%26, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %35 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %36 = vector.transfer_read %arg7[%c1, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %37 = vector.transfer_read %arg7[%c2, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %38 = vector.transfer_read %arg7[%c3, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %39 = arith.extsi %21 : vector<1x4xi8> to vector<1x4xi32>
            %40 = arith.extsi %23 : vector<1x4xi8> to vector<1x4xi32>
            %41 = arith.extsi %25 : vector<1x4xi8> to vector<1x4xi32>
            %42 = arith.extsi %27 : vector<1x4xi8> to vector<1x4xi32>
            %43 = vector.insert_strided_slice %28, %cst {offsets = [0, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %44 = vector.insert_strided_slice %30, %43 {offsets = [1, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %45 = vector.insert_strided_slice %32, %44 {offsets = [2, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %46 = vector.insert_strided_slice %34, %45 {offsets = [3, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %47 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %48 = arith.extsi %47 : vector<4x1xi8> to vector<4x1xi32>
            %49 = vector.extract_strided_slice %35 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %50 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %48, %49 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %51 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %52 = arith.extsi %51 : vector<4x1xi8> to vector<4x1xi32>
            %53 = vector.extract_strided_slice %35 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %54 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %52, %53 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %55 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %56 = arith.extsi %55 : vector<4x1xi8> to vector<4x1xi32>
            %57 = vector.extract_strided_slice %35 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %58 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %56, %57 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %59 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %60 = arith.extsi %59 : vector<4x1xi8> to vector<4x1xi32>
            %61 = vector.extract_strided_slice %35 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %62 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %39, %60, %61 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %63 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
            %65 = vector.extract_strided_slice %36 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %66 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %64, %65 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %67 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %68 = arith.extsi %67 : vector<4x1xi8> to vector<4x1xi32>
            %69 = vector.extract_strided_slice %36 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %70 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %68, %69 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %71 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %72 = arith.extsi %71 : vector<4x1xi8> to vector<4x1xi32>
            %73 = vector.extract_strided_slice %36 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %74 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %72, %73 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %75 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %76 = arith.extsi %75 : vector<4x1xi8> to vector<4x1xi32>
            %77 = vector.extract_strided_slice %36 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %78 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %40, %76, %77 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %79 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %80 = arith.extsi %79 : vector<4x1xi8> to vector<4x1xi32>
            %81 = vector.extract_strided_slice %37 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %82 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %80, %81 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %83 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %84 = arith.extsi %83 : vector<4x1xi8> to vector<4x1xi32>
            %85 = vector.extract_strided_slice %37 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %86 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %84, %85 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %87 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %88 = arith.extsi %87 : vector<4x1xi8> to vector<4x1xi32>
            %89 = vector.extract_strided_slice %37 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %90 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %88, %89 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %91 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %92 = arith.extsi %91 : vector<4x1xi8> to vector<4x1xi32>
            %93 = vector.extract_strided_slice %37 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %94 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %41, %92, %93 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %95 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %96 = arith.extsi %95 : vector<4x1xi8> to vector<4x1xi32>
            %97 = vector.extract_strided_slice %38 {offsets = [0, 0], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %98 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %96, %97 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %99 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %100 = arith.extsi %99 : vector<4x1xi8> to vector<4x1xi32>
            %101 = vector.extract_strided_slice %38 {offsets = [0, 1], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %102 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %100, %101 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %103 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %104 = arith.extsi %103 : vector<4x1xi8> to vector<4x1xi32>
            %105 = vector.extract_strided_slice %38 {offsets = [0, 2], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %106 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %104, %105 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %107 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %108 = arith.extsi %107 : vector<4x1xi8> to vector<4x1xi32>
            %109 = vector.extract_strided_slice %38 {offsets = [0, 3], sizes = [1, 1], strides = [1, 1]} : vector<1x4xi32> to vector<1x1xi32>
            %110 = vector.contract {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, affine_map<(d0, d1, d2) -> (d2, d1)>, affine_map<(d0, d1, d2) -> (d0, d1)>], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %42, %108, %109 : vector<1x4xi32>, vector<4x1xi32> into vector<1x1xi32>
            %111 = vector.insert_strided_slice %50, %cst_0 {offsets = [0, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %112 = vector.insert_strided_slice %54, %111 {offsets = [0, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %113 = vector.insert_strided_slice %58, %112 {offsets = [0, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %114 = vector.insert_strided_slice %62, %113 {offsets = [0, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %115 = vector.insert_strided_slice %66, %114 {offsets = [1, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %116 = vector.insert_strided_slice %70, %115 {offsets = [1, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %117 = vector.insert_strided_slice %74, %116 {offsets = [1, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %118 = vector.insert_strided_slice %78, %117 {offsets = [1, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %119 = vector.insert_strided_slice %82, %118 {offsets = [2, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %120 = vector.insert_strided_slice %86, %119 {offsets = [2, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %121 = vector.insert_strided_slice %90, %120 {offsets = [2, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %122 = vector.insert_strided_slice %94, %121 {offsets = [2, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %123 = vector.insert_strided_slice %98, %122 {offsets = [3, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %124 = vector.insert_strided_slice %102, %123 {offsets = [3, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %125 = vector.insert_strided_slice %106, %124 {offsets = [3, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %126 = vector.insert_strided_slice %110, %125 {offsets = [3, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %127 = vector.extract_strided_slice %126 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %128 = vector.transfer_write %127, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %129 = vector.extract_strided_slice %126 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %130 = vector.transfer_write %129, %128[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %131 = vector.extract_strided_slice %126 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %132 = vector.transfer_write %131, %130[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %133 = vector.extract_strided_slice %126 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %134 = vector.transfer_write %133, %132[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            scf.yield %134 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %20 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After lowering size-1 reduction contract ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<1x1xi32>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %cst_2 = arith.constant dense<0> : vector<4x4xi8>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %cst_3 = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.extract_strided_slice %cst_3 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %13 = vector.transfer_write %12, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %14 = vector.extract_strided_slice %cst_3 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %15 = vector.transfer_write %14, %13[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %16 = vector.extract_strided_slice %cst_3 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %17 = vector.transfer_write %16, %15[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %18 = vector.extract_strided_slice %cst_3 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %19 = vector.transfer_write %18, %17[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %20 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %19) -> (tensor<4x4xi32>) {
            %21 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %22 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
            %23 = vector.transfer_read %8[%22, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %24 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
            %25 = vector.transfer_read %8[%24, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %26 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
            %27 = vector.transfer_read %8[%26, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %35 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %36 = vector.transfer_read %arg7[%c1, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %37 = vector.transfer_read %arg7[%c2, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %38 = vector.transfer_read %arg7[%c3, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %39 = arith.extsi %21 : vector<1x4xi8> to vector<1x4xi32>
            %40 = arith.extsi %23 : vector<1x4xi8> to vector<1x4xi32>
            %41 = arith.extsi %25 : vector<1x4xi8> to vector<1x4xi32>
            %42 = arith.extsi %27 : vector<1x4xi8> to vector<1x4xi32>
            %43 = vector.insert_strided_slice %28, %cst_2 {offsets = [0, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %44 = vector.insert_strided_slice %30, %43 {offsets = [1, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %45 = vector.insert_strided_slice %32, %44 {offsets = [2, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %46 = vector.insert_strided_slice %34, %45 {offsets = [3, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %47 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %48 = arith.extsi %47 : vector<4x1xi8> to vector<4x1xi32>
            %49 = vector.extract %39[0] : vector<1x4xi32>
            %50 = vector.extract %48[0, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %cst_1 [0] : i32 into vector<4xi32>
            %52 = vector.extract %48[1, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [1] : i32 into vector<4xi32>
            %54 = vector.extract %48[2, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [2] : i32 into vector<4xi32>
            %56 = vector.extract %48[3, 0] : vector<4x1xi32>
            %57 = vector.insert %56, %55 [3] : i32 into vector<4xi32>
            %58 = vector.extract %35[0, 0] : vector<1x4xi32>
            %59 = arith.muli %49, %57 : vector<4xi32>
            %60 = vector.reduction <add>, %59, %58 : vector<4xi32> into i32
            %61 = vector.insert %60, %cst_0 [0] : i32 into vector<1xi32>
            %62 = vector.insert %61, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %63 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
            %65 = vector.extract %39[0] : vector<1x4xi32>
            %66 = vector.extract %64[0, 0] : vector<4x1xi32>
            %67 = vector.insert %66, %cst_1 [0] : i32 into vector<4xi32>
            %68 = vector.extract %64[1, 0] : vector<4x1xi32>
            %69 = vector.insert %68, %67 [1] : i32 into vector<4xi32>
            %70 = vector.extract %64[2, 0] : vector<4x1xi32>
            %71 = vector.insert %70, %69 [2] : i32 into vector<4xi32>
            %72 = vector.extract %64[3, 0] : vector<4x1xi32>
            %73 = vector.insert %72, %71 [3] : i32 into vector<4xi32>
            %74 = vector.extract %35[0, 1] : vector<1x4xi32>
            %75 = arith.muli %65, %73 : vector<4xi32>
            %76 = vector.reduction <add>, %75, %74 : vector<4xi32> into i32
            %77 = vector.insert %76, %cst_0 [0] : i32 into vector<1xi32>
            %78 = vector.insert %77, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %79 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %80 = arith.extsi %79 : vector<4x1xi8> to vector<4x1xi32>
            %81 = vector.extract %39[0] : vector<1x4xi32>
            %82 = vector.extract %80[0, 0] : vector<4x1xi32>
            %83 = vector.insert %82, %cst_1 [0] : i32 into vector<4xi32>
            %84 = vector.extract %80[1, 0] : vector<4x1xi32>
            %85 = vector.insert %84, %83 [1] : i32 into vector<4xi32>
            %86 = vector.extract %80[2, 0] : vector<4x1xi32>
            %87 = vector.insert %86, %85 [2] : i32 into vector<4xi32>
            %88 = vector.extract %80[3, 0] : vector<4x1xi32>
            %89 = vector.insert %88, %87 [3] : i32 into vector<4xi32>
            %90 = vector.extract %35[0, 2] : vector<1x4xi32>
            %91 = arith.muli %81, %89 : vector<4xi32>
            %92 = vector.reduction <add>, %91, %90 : vector<4xi32> into i32
            %93 = vector.insert %92, %cst_0 [0] : i32 into vector<1xi32>
            %94 = vector.insert %93, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %95 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %96 = arith.extsi %95 : vector<4x1xi8> to vector<4x1xi32>
            %97 = vector.extract %39[0] : vector<1x4xi32>
            %98 = vector.extract %96[0, 0] : vector<4x1xi32>
            %99 = vector.insert %98, %cst_1 [0] : i32 into vector<4xi32>
            %100 = vector.extract %96[1, 0] : vector<4x1xi32>
            %101 = vector.insert %100, %99 [1] : i32 into vector<4xi32>
            %102 = vector.extract %96[2, 0] : vector<4x1xi32>
            %103 = vector.insert %102, %101 [2] : i32 into vector<4xi32>
            %104 = vector.extract %96[3, 0] : vector<4x1xi32>
            %105 = vector.insert %104, %103 [3] : i32 into vector<4xi32>
            %106 = vector.extract %35[0, 3] : vector<1x4xi32>
            %107 = arith.muli %97, %105 : vector<4xi32>
            %108 = vector.reduction <add>, %107, %106 : vector<4xi32> into i32
            %109 = vector.insert %108, %cst_0 [0] : i32 into vector<1xi32>
            %110 = vector.insert %109, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %111 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %112 = arith.extsi %111 : vector<4x1xi8> to vector<4x1xi32>
            %113 = vector.extract %40[0] : vector<1x4xi32>
            %114 = vector.extract %112[0, 0] : vector<4x1xi32>
            %115 = vector.insert %114, %cst_1 [0] : i32 into vector<4xi32>
            %116 = vector.extract %112[1, 0] : vector<4x1xi32>
            %117 = vector.insert %116, %115 [1] : i32 into vector<4xi32>
            %118 = vector.extract %112[2, 0] : vector<4x1xi32>
            %119 = vector.insert %118, %117 [2] : i32 into vector<4xi32>
            %120 = vector.extract %112[3, 0] : vector<4x1xi32>
            %121 = vector.insert %120, %119 [3] : i32 into vector<4xi32>
            %122 = vector.extract %36[0, 0] : vector<1x4xi32>
            %123 = arith.muli %113, %121 : vector<4xi32>
            %124 = vector.reduction <add>, %123, %122 : vector<4xi32> into i32
            %125 = vector.insert %124, %cst_0 [0] : i32 into vector<1xi32>
            %126 = vector.insert %125, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %127 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %128 = arith.extsi %127 : vector<4x1xi8> to vector<4x1xi32>
            %129 = vector.extract %40[0] : vector<1x4xi32>
            %130 = vector.extract %128[0, 0] : vector<4x1xi32>
            %131 = vector.insert %130, %cst_1 [0] : i32 into vector<4xi32>
            %132 = vector.extract %128[1, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %131 [1] : i32 into vector<4xi32>
            %134 = vector.extract %128[2, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [2] : i32 into vector<4xi32>
            %136 = vector.extract %128[3, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [3] : i32 into vector<4xi32>
            %138 = vector.extract %36[0, 1] : vector<1x4xi32>
            %139 = arith.muli %129, %137 : vector<4xi32>
            %140 = vector.reduction <add>, %139, %138 : vector<4xi32> into i32
            %141 = vector.insert %140, %cst_0 [0] : i32 into vector<1xi32>
            %142 = vector.insert %141, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %143 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %144 = arith.extsi %143 : vector<4x1xi8> to vector<4x1xi32>
            %145 = vector.extract %40[0] : vector<1x4xi32>
            %146 = vector.extract %144[0, 0] : vector<4x1xi32>
            %147 = vector.insert %146, %cst_1 [0] : i32 into vector<4xi32>
            %148 = vector.extract %144[1, 0] : vector<4x1xi32>
            %149 = vector.insert %148, %147 [1] : i32 into vector<4xi32>
            %150 = vector.extract %144[2, 0] : vector<4x1xi32>
            %151 = vector.insert %150, %149 [2] : i32 into vector<4xi32>
            %152 = vector.extract %144[3, 0] : vector<4x1xi32>
            %153 = vector.insert %152, %151 [3] : i32 into vector<4xi32>
            %154 = vector.extract %36[0, 2] : vector<1x4xi32>
            %155 = arith.muli %145, %153 : vector<4xi32>
            %156 = vector.reduction <add>, %155, %154 : vector<4xi32> into i32
            %157 = vector.insert %156, %cst_0 [0] : i32 into vector<1xi32>
            %158 = vector.insert %157, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %159 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %160 = arith.extsi %159 : vector<4x1xi8> to vector<4x1xi32>
            %161 = vector.extract %40[0] : vector<1x4xi32>
            %162 = vector.extract %160[0, 0] : vector<4x1xi32>
            %163 = vector.insert %162, %cst_1 [0] : i32 into vector<4xi32>
            %164 = vector.extract %160[1, 0] : vector<4x1xi32>
            %165 = vector.insert %164, %163 [1] : i32 into vector<4xi32>
            %166 = vector.extract %160[2, 0] : vector<4x1xi32>
            %167 = vector.insert %166, %165 [2] : i32 into vector<4xi32>
            %168 = vector.extract %160[3, 0] : vector<4x1xi32>
            %169 = vector.insert %168, %167 [3] : i32 into vector<4xi32>
            %170 = vector.extract %36[0, 3] : vector<1x4xi32>
            %171 = arith.muli %161, %169 : vector<4xi32>
            %172 = vector.reduction <add>, %171, %170 : vector<4xi32> into i32
            %173 = vector.insert %172, %cst_0 [0] : i32 into vector<1xi32>
            %174 = vector.insert %173, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %175 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %176 = arith.extsi %175 : vector<4x1xi8> to vector<4x1xi32>
            %177 = vector.extract %41[0] : vector<1x4xi32>
            %178 = vector.extract %176[0, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %cst_1 [0] : i32 into vector<4xi32>
            %180 = vector.extract %176[1, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [1] : i32 into vector<4xi32>
            %182 = vector.extract %176[2, 0] : vector<4x1xi32>
            %183 = vector.insert %182, %181 [2] : i32 into vector<4xi32>
            %184 = vector.extract %176[3, 0] : vector<4x1xi32>
            %185 = vector.insert %184, %183 [3] : i32 into vector<4xi32>
            %186 = vector.extract %37[0, 0] : vector<1x4xi32>
            %187 = arith.muli %177, %185 : vector<4xi32>
            %188 = vector.reduction <add>, %187, %186 : vector<4xi32> into i32
            %189 = vector.insert %188, %cst_0 [0] : i32 into vector<1xi32>
            %190 = vector.insert %189, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %191 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %192 = arith.extsi %191 : vector<4x1xi8> to vector<4x1xi32>
            %193 = vector.extract %41[0] : vector<1x4xi32>
            %194 = vector.extract %192[0, 0] : vector<4x1xi32>
            %195 = vector.insert %194, %cst_1 [0] : i32 into vector<4xi32>
            %196 = vector.extract %192[1, 0] : vector<4x1xi32>
            %197 = vector.insert %196, %195 [1] : i32 into vector<4xi32>
            %198 = vector.extract %192[2, 0] : vector<4x1xi32>
            %199 = vector.insert %198, %197 [2] : i32 into vector<4xi32>
            %200 = vector.extract %192[3, 0] : vector<4x1xi32>
            %201 = vector.insert %200, %199 [3] : i32 into vector<4xi32>
            %202 = vector.extract %37[0, 1] : vector<1x4xi32>
            %203 = arith.muli %193, %201 : vector<4xi32>
            %204 = vector.reduction <add>, %203, %202 : vector<4xi32> into i32
            %205 = vector.insert %204, %cst_0 [0] : i32 into vector<1xi32>
            %206 = vector.insert %205, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %207 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %208 = arith.extsi %207 : vector<4x1xi8> to vector<4x1xi32>
            %209 = vector.extract %41[0] : vector<1x4xi32>
            %210 = vector.extract %208[0, 0] : vector<4x1xi32>
            %211 = vector.insert %210, %cst_1 [0] : i32 into vector<4xi32>
            %212 = vector.extract %208[1, 0] : vector<4x1xi32>
            %213 = vector.insert %212, %211 [1] : i32 into vector<4xi32>
            %214 = vector.extract %208[2, 0] : vector<4x1xi32>
            %215 = vector.insert %214, %213 [2] : i32 into vector<4xi32>
            %216 = vector.extract %208[3, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %215 [3] : i32 into vector<4xi32>
            %218 = vector.extract %37[0, 2] : vector<1x4xi32>
            %219 = arith.muli %209, %217 : vector<4xi32>
            %220 = vector.reduction <add>, %219, %218 : vector<4xi32> into i32
            %221 = vector.insert %220, %cst_0 [0] : i32 into vector<1xi32>
            %222 = vector.insert %221, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %223 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %224 = arith.extsi %223 : vector<4x1xi8> to vector<4x1xi32>
            %225 = vector.extract %41[0] : vector<1x4xi32>
            %226 = vector.extract %224[0, 0] : vector<4x1xi32>
            %227 = vector.insert %226, %cst_1 [0] : i32 into vector<4xi32>
            %228 = vector.extract %224[1, 0] : vector<4x1xi32>
            %229 = vector.insert %228, %227 [1] : i32 into vector<4xi32>
            %230 = vector.extract %224[2, 0] : vector<4x1xi32>
            %231 = vector.insert %230, %229 [2] : i32 into vector<4xi32>
            %232 = vector.extract %224[3, 0] : vector<4x1xi32>
            %233 = vector.insert %232, %231 [3] : i32 into vector<4xi32>
            %234 = vector.extract %37[0, 3] : vector<1x4xi32>
            %235 = arith.muli %225, %233 : vector<4xi32>
            %236 = vector.reduction <add>, %235, %234 : vector<4xi32> into i32
            %237 = vector.insert %236, %cst_0 [0] : i32 into vector<1xi32>
            %238 = vector.insert %237, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %239 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %240 = arith.extsi %239 : vector<4x1xi8> to vector<4x1xi32>
            %241 = vector.extract %42[0] : vector<1x4xi32>
            %242 = vector.extract %240[0, 0] : vector<4x1xi32>
            %243 = vector.insert %242, %cst_1 [0] : i32 into vector<4xi32>
            %244 = vector.extract %240[1, 0] : vector<4x1xi32>
            %245 = vector.insert %244, %243 [1] : i32 into vector<4xi32>
            %246 = vector.extract %240[2, 0] : vector<4x1xi32>
            %247 = vector.insert %246, %245 [2] : i32 into vector<4xi32>
            %248 = vector.extract %240[3, 0] : vector<4x1xi32>
            %249 = vector.insert %248, %247 [3] : i32 into vector<4xi32>
            %250 = vector.extract %38[0, 0] : vector<1x4xi32>
            %251 = arith.muli %241, %249 : vector<4xi32>
            %252 = vector.reduction <add>, %251, %250 : vector<4xi32> into i32
            %253 = vector.insert %252, %cst_0 [0] : i32 into vector<1xi32>
            %254 = vector.insert %253, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %255 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %256 = arith.extsi %255 : vector<4x1xi8> to vector<4x1xi32>
            %257 = vector.extract %42[0] : vector<1x4xi32>
            %258 = vector.extract %256[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %256[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %256[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %256[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %38[0, 1] : vector<1x4xi32>
            %267 = arith.muli %257, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.insert %269, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %271 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %272 = arith.extsi %271 : vector<4x1xi8> to vector<4x1xi32>
            %273 = vector.extract %42[0] : vector<1x4xi32>
            %274 = vector.extract %272[0, 0] : vector<4x1xi32>
            %275 = vector.insert %274, %cst_1 [0] : i32 into vector<4xi32>
            %276 = vector.extract %272[1, 0] : vector<4x1xi32>
            %277 = vector.insert %276, %275 [1] : i32 into vector<4xi32>
            %278 = vector.extract %272[2, 0] : vector<4x1xi32>
            %279 = vector.insert %278, %277 [2] : i32 into vector<4xi32>
            %280 = vector.extract %272[3, 0] : vector<4x1xi32>
            %281 = vector.insert %280, %279 [3] : i32 into vector<4xi32>
            %282 = vector.extract %38[0, 2] : vector<1x4xi32>
            %283 = arith.muli %273, %281 : vector<4xi32>
            %284 = vector.reduction <add>, %283, %282 : vector<4xi32> into i32
            %285 = vector.insert %284, %cst_0 [0] : i32 into vector<1xi32>
            %286 = vector.insert %285, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %287 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %288 = arith.extsi %287 : vector<4x1xi8> to vector<4x1xi32>
            %289 = vector.extract %42[0] : vector<1x4xi32>
            %290 = vector.extract %288[0, 0] : vector<4x1xi32>
            %291 = vector.insert %290, %cst_1 [0] : i32 into vector<4xi32>
            %292 = vector.extract %288[1, 0] : vector<4x1xi32>
            %293 = vector.insert %292, %291 [1] : i32 into vector<4xi32>
            %294 = vector.extract %288[2, 0] : vector<4x1xi32>
            %295 = vector.insert %294, %293 [2] : i32 into vector<4xi32>
            %296 = vector.extract %288[3, 0] : vector<4x1xi32>
            %297 = vector.insert %296, %295 [3] : i32 into vector<4xi32>
            %298 = vector.extract %38[0, 3] : vector<1x4xi32>
            %299 = arith.muli %289, %297 : vector<4xi32>
            %300 = vector.reduction <add>, %299, %298 : vector<4xi32> into i32
            %301 = vector.insert %300, %cst_0 [0] : i32 into vector<1xi32>
            %302 = vector.insert %301, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %303 = vector.insert_strided_slice %62, %cst_3 {offsets = [0, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %304 = vector.insert_strided_slice %78, %303 {offsets = [0, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %305 = vector.insert_strided_slice %94, %304 {offsets = [0, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %306 = vector.insert_strided_slice %110, %305 {offsets = [0, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %307 = vector.insert_strided_slice %126, %306 {offsets = [1, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %308 = vector.insert_strided_slice %142, %307 {offsets = [1, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %309 = vector.insert_strided_slice %158, %308 {offsets = [1, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %310 = vector.insert_strided_slice %174, %309 {offsets = [1, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %311 = vector.insert_strided_slice %190, %310 {offsets = [2, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %312 = vector.insert_strided_slice %206, %311 {offsets = [2, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %313 = vector.insert_strided_slice %222, %312 {offsets = [2, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %314 = vector.insert_strided_slice %238, %313 {offsets = [2, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %315 = vector.insert_strided_slice %254, %314 {offsets = [3, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %316 = vector.insert_strided_slice %270, %315 {offsets = [3, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %317 = vector.insert_strided_slice %286, %316 {offsets = [3, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %318 = vector.insert_strided_slice %302, %317 {offsets = [3, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %319 = vector.extract_strided_slice %318 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %320 = vector.transfer_write %319, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %321 = vector.extract_strided_slice %318 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %322 = vector.transfer_write %321, %320[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %323 = vector.extract_strided_slice %318 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %324 = vector.transfer_write %323, %322[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %325 = vector.extract_strided_slice %318 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %326 = vector.transfer_write %325, %324[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            scf.yield %326 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %20 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After lowering transpose ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<1x1xi32>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %cst_2 = arith.constant dense<0> : vector<4x4xi8>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %cst_3 = arith.constant dense<0> : vector<4x4xi32>
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.extract_strided_slice %cst_3 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %13 = vector.transfer_write %12, %extracted_slice[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %14 = vector.extract_strided_slice %cst_3 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %15 = vector.transfer_write %14, %13[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %16 = vector.extract_strided_slice %cst_3 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %17 = vector.transfer_write %16, %15[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %18 = vector.extract_strided_slice %cst_3 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
          %19 = vector.transfer_write %18, %17[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
          %20 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %19) -> (tensor<4x4xi32>) {
            %21 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %22 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
            %23 = vector.transfer_read %8[%22, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %24 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
            %25 = vector.transfer_read %8[%24, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %26 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
            %27 = vector.transfer_read %8[%26, %arg6], %c0_i8 {in_bounds = [true, true]} : tensor<8x1024xi8>, vector<1x4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true, true]} : tensor<1024x32xi8>, vector<1x4xi8>
            %35 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %36 = vector.transfer_read %arg7[%c1, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %37 = vector.transfer_read %arg7[%c2, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %38 = vector.transfer_read %arg7[%c3, %c0], %c0_i32 {in_bounds = [true, true]} : tensor<4x4xi32>, vector<1x4xi32>
            %39 = arith.extsi %21 : vector<1x4xi8> to vector<1x4xi32>
            %40 = arith.extsi %23 : vector<1x4xi8> to vector<1x4xi32>
            %41 = arith.extsi %25 : vector<1x4xi8> to vector<1x4xi32>
            %42 = arith.extsi %27 : vector<1x4xi8> to vector<1x4xi32>
            %43 = vector.insert_strided_slice %28, %cst_2 {offsets = [0, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %44 = vector.insert_strided_slice %30, %43 {offsets = [1, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %45 = vector.insert_strided_slice %32, %44 {offsets = [2, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %46 = vector.insert_strided_slice %34, %45 {offsets = [3, 0], strides = [1, 1]} : vector<1x4xi8> into vector<4x4xi8>
            %47 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %48 = arith.extsi %47 : vector<4x1xi8> to vector<4x1xi32>
            %49 = vector.extract %39[0] : vector<1x4xi32>
            %50 = vector.extract %48[0, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %cst_1 [0] : i32 into vector<4xi32>
            %52 = vector.extract %48[1, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [1] : i32 into vector<4xi32>
            %54 = vector.extract %48[2, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [2] : i32 into vector<4xi32>
            %56 = vector.extract %48[3, 0] : vector<4x1xi32>
            %57 = vector.insert %56, %55 [3] : i32 into vector<4xi32>
            %58 = vector.extract %35[0, 0] : vector<1x4xi32>
            %59 = arith.muli %49, %57 : vector<4xi32>
            %60 = vector.reduction <add>, %59, %58 : vector<4xi32> into i32
            %61 = vector.insert %60, %cst_0 [0] : i32 into vector<1xi32>
            %62 = vector.insert %61, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %63 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
            %65 = vector.extract %39[0] : vector<1x4xi32>
            %66 = vector.extract %64[0, 0] : vector<4x1xi32>
            %67 = vector.insert %66, %cst_1 [0] : i32 into vector<4xi32>
            %68 = vector.extract %64[1, 0] : vector<4x1xi32>
            %69 = vector.insert %68, %67 [1] : i32 into vector<4xi32>
            %70 = vector.extract %64[2, 0] : vector<4x1xi32>
            %71 = vector.insert %70, %69 [2] : i32 into vector<4xi32>
            %72 = vector.extract %64[3, 0] : vector<4x1xi32>
            %73 = vector.insert %72, %71 [3] : i32 into vector<4xi32>
            %74 = vector.extract %35[0, 1] : vector<1x4xi32>
            %75 = arith.muli %65, %73 : vector<4xi32>
            %76 = vector.reduction <add>, %75, %74 : vector<4xi32> into i32
            %77 = vector.insert %76, %cst_0 [0] : i32 into vector<1xi32>
            %78 = vector.insert %77, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %79 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %80 = arith.extsi %79 : vector<4x1xi8> to vector<4x1xi32>
            %81 = vector.extract %39[0] : vector<1x4xi32>
            %82 = vector.extract %80[0, 0] : vector<4x1xi32>
            %83 = vector.insert %82, %cst_1 [0] : i32 into vector<4xi32>
            %84 = vector.extract %80[1, 0] : vector<4x1xi32>
            %85 = vector.insert %84, %83 [1] : i32 into vector<4xi32>
            %86 = vector.extract %80[2, 0] : vector<4x1xi32>
            %87 = vector.insert %86, %85 [2] : i32 into vector<4xi32>
            %88 = vector.extract %80[3, 0] : vector<4x1xi32>
            %89 = vector.insert %88, %87 [3] : i32 into vector<4xi32>
            %90 = vector.extract %35[0, 2] : vector<1x4xi32>
            %91 = arith.muli %81, %89 : vector<4xi32>
            %92 = vector.reduction <add>, %91, %90 : vector<4xi32> into i32
            %93 = vector.insert %92, %cst_0 [0] : i32 into vector<1xi32>
            %94 = vector.insert %93, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %95 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %96 = arith.extsi %95 : vector<4x1xi8> to vector<4x1xi32>
            %97 = vector.extract %39[0] : vector<1x4xi32>
            %98 = vector.extract %96[0, 0] : vector<4x1xi32>
            %99 = vector.insert %98, %cst_1 [0] : i32 into vector<4xi32>
            %100 = vector.extract %96[1, 0] : vector<4x1xi32>
            %101 = vector.insert %100, %99 [1] : i32 into vector<4xi32>
            %102 = vector.extract %96[2, 0] : vector<4x1xi32>
            %103 = vector.insert %102, %101 [2] : i32 into vector<4xi32>
            %104 = vector.extract %96[3, 0] : vector<4x1xi32>
            %105 = vector.insert %104, %103 [3] : i32 into vector<4xi32>
            %106 = vector.extract %35[0, 3] : vector<1x4xi32>
            %107 = arith.muli %97, %105 : vector<4xi32>
            %108 = vector.reduction <add>, %107, %106 : vector<4xi32> into i32
            %109 = vector.insert %108, %cst_0 [0] : i32 into vector<1xi32>
            %110 = vector.insert %109, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %111 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %112 = arith.extsi %111 : vector<4x1xi8> to vector<4x1xi32>
            %113 = vector.extract %40[0] : vector<1x4xi32>
            %114 = vector.extract %112[0, 0] : vector<4x1xi32>
            %115 = vector.insert %114, %cst_1 [0] : i32 into vector<4xi32>
            %116 = vector.extract %112[1, 0] : vector<4x1xi32>
            %117 = vector.insert %116, %115 [1] : i32 into vector<4xi32>
            %118 = vector.extract %112[2, 0] : vector<4x1xi32>
            %119 = vector.insert %118, %117 [2] : i32 into vector<4xi32>
            %120 = vector.extract %112[3, 0] : vector<4x1xi32>
            %121 = vector.insert %120, %119 [3] : i32 into vector<4xi32>
            %122 = vector.extract %36[0, 0] : vector<1x4xi32>
            %123 = arith.muli %113, %121 : vector<4xi32>
            %124 = vector.reduction <add>, %123, %122 : vector<4xi32> into i32
            %125 = vector.insert %124, %cst_0 [0] : i32 into vector<1xi32>
            %126 = vector.insert %125, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %127 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %128 = arith.extsi %127 : vector<4x1xi8> to vector<4x1xi32>
            %129 = vector.extract %40[0] : vector<1x4xi32>
            %130 = vector.extract %128[0, 0] : vector<4x1xi32>
            %131 = vector.insert %130, %cst_1 [0] : i32 into vector<4xi32>
            %132 = vector.extract %128[1, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %131 [1] : i32 into vector<4xi32>
            %134 = vector.extract %128[2, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [2] : i32 into vector<4xi32>
            %136 = vector.extract %128[3, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [3] : i32 into vector<4xi32>
            %138 = vector.extract %36[0, 1] : vector<1x4xi32>
            %139 = arith.muli %129, %137 : vector<4xi32>
            %140 = vector.reduction <add>, %139, %138 : vector<4xi32> into i32
            %141 = vector.insert %140, %cst_0 [0] : i32 into vector<1xi32>
            %142 = vector.insert %141, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %143 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %144 = arith.extsi %143 : vector<4x1xi8> to vector<4x1xi32>
            %145 = vector.extract %40[0] : vector<1x4xi32>
            %146 = vector.extract %144[0, 0] : vector<4x1xi32>
            %147 = vector.insert %146, %cst_1 [0] : i32 into vector<4xi32>
            %148 = vector.extract %144[1, 0] : vector<4x1xi32>
            %149 = vector.insert %148, %147 [1] : i32 into vector<4xi32>
            %150 = vector.extract %144[2, 0] : vector<4x1xi32>
            %151 = vector.insert %150, %149 [2] : i32 into vector<4xi32>
            %152 = vector.extract %144[3, 0] : vector<4x1xi32>
            %153 = vector.insert %152, %151 [3] : i32 into vector<4xi32>
            %154 = vector.extract %36[0, 2] : vector<1x4xi32>
            %155 = arith.muli %145, %153 : vector<4xi32>
            %156 = vector.reduction <add>, %155, %154 : vector<4xi32> into i32
            %157 = vector.insert %156, %cst_0 [0] : i32 into vector<1xi32>
            %158 = vector.insert %157, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %159 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %160 = arith.extsi %159 : vector<4x1xi8> to vector<4x1xi32>
            %161 = vector.extract %40[0] : vector<1x4xi32>
            %162 = vector.extract %160[0, 0] : vector<4x1xi32>
            %163 = vector.insert %162, %cst_1 [0] : i32 into vector<4xi32>
            %164 = vector.extract %160[1, 0] : vector<4x1xi32>
            %165 = vector.insert %164, %163 [1] : i32 into vector<4xi32>
            %166 = vector.extract %160[2, 0] : vector<4x1xi32>
            %167 = vector.insert %166, %165 [2] : i32 into vector<4xi32>
            %168 = vector.extract %160[3, 0] : vector<4x1xi32>
            %169 = vector.insert %168, %167 [3] : i32 into vector<4xi32>
            %170 = vector.extract %36[0, 3] : vector<1x4xi32>
            %171 = arith.muli %161, %169 : vector<4xi32>
            %172 = vector.reduction <add>, %171, %170 : vector<4xi32> into i32
            %173 = vector.insert %172, %cst_0 [0] : i32 into vector<1xi32>
            %174 = vector.insert %173, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %175 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %176 = arith.extsi %175 : vector<4x1xi8> to vector<4x1xi32>
            %177 = vector.extract %41[0] : vector<1x4xi32>
            %178 = vector.extract %176[0, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %cst_1 [0] : i32 into vector<4xi32>
            %180 = vector.extract %176[1, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [1] : i32 into vector<4xi32>
            %182 = vector.extract %176[2, 0] : vector<4x1xi32>
            %183 = vector.insert %182, %181 [2] : i32 into vector<4xi32>
            %184 = vector.extract %176[3, 0] : vector<4x1xi32>
            %185 = vector.insert %184, %183 [3] : i32 into vector<4xi32>
            %186 = vector.extract %37[0, 0] : vector<1x4xi32>
            %187 = arith.muli %177, %185 : vector<4xi32>
            %188 = vector.reduction <add>, %187, %186 : vector<4xi32> into i32
            %189 = vector.insert %188, %cst_0 [0] : i32 into vector<1xi32>
            %190 = vector.insert %189, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %191 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %192 = arith.extsi %191 : vector<4x1xi8> to vector<4x1xi32>
            %193 = vector.extract %41[0] : vector<1x4xi32>
            %194 = vector.extract %192[0, 0] : vector<4x1xi32>
            %195 = vector.insert %194, %cst_1 [0] : i32 into vector<4xi32>
            %196 = vector.extract %192[1, 0] : vector<4x1xi32>
            %197 = vector.insert %196, %195 [1] : i32 into vector<4xi32>
            %198 = vector.extract %192[2, 0] : vector<4x1xi32>
            %199 = vector.insert %198, %197 [2] : i32 into vector<4xi32>
            %200 = vector.extract %192[3, 0] : vector<4x1xi32>
            %201 = vector.insert %200, %199 [3] : i32 into vector<4xi32>
            %202 = vector.extract %37[0, 1] : vector<1x4xi32>
            %203 = arith.muli %193, %201 : vector<4xi32>
            %204 = vector.reduction <add>, %203, %202 : vector<4xi32> into i32
            %205 = vector.insert %204, %cst_0 [0] : i32 into vector<1xi32>
            %206 = vector.insert %205, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %207 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %208 = arith.extsi %207 : vector<4x1xi8> to vector<4x1xi32>
            %209 = vector.extract %41[0] : vector<1x4xi32>
            %210 = vector.extract %208[0, 0] : vector<4x1xi32>
            %211 = vector.insert %210, %cst_1 [0] : i32 into vector<4xi32>
            %212 = vector.extract %208[1, 0] : vector<4x1xi32>
            %213 = vector.insert %212, %211 [1] : i32 into vector<4xi32>
            %214 = vector.extract %208[2, 0] : vector<4x1xi32>
            %215 = vector.insert %214, %213 [2] : i32 into vector<4xi32>
            %216 = vector.extract %208[3, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %215 [3] : i32 into vector<4xi32>
            %218 = vector.extract %37[0, 2] : vector<1x4xi32>
            %219 = arith.muli %209, %217 : vector<4xi32>
            %220 = vector.reduction <add>, %219, %218 : vector<4xi32> into i32
            %221 = vector.insert %220, %cst_0 [0] : i32 into vector<1xi32>
            %222 = vector.insert %221, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %223 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %224 = arith.extsi %223 : vector<4x1xi8> to vector<4x1xi32>
            %225 = vector.extract %41[0] : vector<1x4xi32>
            %226 = vector.extract %224[0, 0] : vector<4x1xi32>
            %227 = vector.insert %226, %cst_1 [0] : i32 into vector<4xi32>
            %228 = vector.extract %224[1, 0] : vector<4x1xi32>
            %229 = vector.insert %228, %227 [1] : i32 into vector<4xi32>
            %230 = vector.extract %224[2, 0] : vector<4x1xi32>
            %231 = vector.insert %230, %229 [2] : i32 into vector<4xi32>
            %232 = vector.extract %224[3, 0] : vector<4x1xi32>
            %233 = vector.insert %232, %231 [3] : i32 into vector<4xi32>
            %234 = vector.extract %37[0, 3] : vector<1x4xi32>
            %235 = arith.muli %225, %233 : vector<4xi32>
            %236 = vector.reduction <add>, %235, %234 : vector<4xi32> into i32
            %237 = vector.insert %236, %cst_0 [0] : i32 into vector<1xi32>
            %238 = vector.insert %237, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %239 = vector.extract_strided_slice %46 {offsets = [0, 0], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %240 = arith.extsi %239 : vector<4x1xi8> to vector<4x1xi32>
            %241 = vector.extract %42[0] : vector<1x4xi32>
            %242 = vector.extract %240[0, 0] : vector<4x1xi32>
            %243 = vector.insert %242, %cst_1 [0] : i32 into vector<4xi32>
            %244 = vector.extract %240[1, 0] : vector<4x1xi32>
            %245 = vector.insert %244, %243 [1] : i32 into vector<4xi32>
            %246 = vector.extract %240[2, 0] : vector<4x1xi32>
            %247 = vector.insert %246, %245 [2] : i32 into vector<4xi32>
            %248 = vector.extract %240[3, 0] : vector<4x1xi32>
            %249 = vector.insert %248, %247 [3] : i32 into vector<4xi32>
            %250 = vector.extract %38[0, 0] : vector<1x4xi32>
            %251 = arith.muli %241, %249 : vector<4xi32>
            %252 = vector.reduction <add>, %251, %250 : vector<4xi32> into i32
            %253 = vector.insert %252, %cst_0 [0] : i32 into vector<1xi32>
            %254 = vector.insert %253, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %255 = vector.extract_strided_slice %46 {offsets = [0, 1], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %256 = arith.extsi %255 : vector<4x1xi8> to vector<4x1xi32>
            %257 = vector.extract %42[0] : vector<1x4xi32>
            %258 = vector.extract %256[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %256[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %256[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %256[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %38[0, 1] : vector<1x4xi32>
            %267 = arith.muli %257, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.insert %269, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %271 = vector.extract_strided_slice %46 {offsets = [0, 2], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %272 = arith.extsi %271 : vector<4x1xi8> to vector<4x1xi32>
            %273 = vector.extract %42[0] : vector<1x4xi32>
            %274 = vector.extract %272[0, 0] : vector<4x1xi32>
            %275 = vector.insert %274, %cst_1 [0] : i32 into vector<4xi32>
            %276 = vector.extract %272[1, 0] : vector<4x1xi32>
            %277 = vector.insert %276, %275 [1] : i32 into vector<4xi32>
            %278 = vector.extract %272[2, 0] : vector<4x1xi32>
            %279 = vector.insert %278, %277 [2] : i32 into vector<4xi32>
            %280 = vector.extract %272[3, 0] : vector<4x1xi32>
            %281 = vector.insert %280, %279 [3] : i32 into vector<4xi32>
            %282 = vector.extract %38[0, 2] : vector<1x4xi32>
            %283 = arith.muli %273, %281 : vector<4xi32>
            %284 = vector.reduction <add>, %283, %282 : vector<4xi32> into i32
            %285 = vector.insert %284, %cst_0 [0] : i32 into vector<1xi32>
            %286 = vector.insert %285, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %287 = vector.extract_strided_slice %46 {offsets = [0, 3], sizes = [4, 1], strides = [1, 1]} : vector<4x4xi8> to vector<4x1xi8>
            %288 = arith.extsi %287 : vector<4x1xi8> to vector<4x1xi32>
            %289 = vector.extract %42[0] : vector<1x4xi32>
            %290 = vector.extract %288[0, 0] : vector<4x1xi32>
            %291 = vector.insert %290, %cst_1 [0] : i32 into vector<4xi32>
            %292 = vector.extract %288[1, 0] : vector<4x1xi32>
            %293 = vector.insert %292, %291 [1] : i32 into vector<4xi32>
            %294 = vector.extract %288[2, 0] : vector<4x1xi32>
            %295 = vector.insert %294, %293 [2] : i32 into vector<4xi32>
            %296 = vector.extract %288[3, 0] : vector<4x1xi32>
            %297 = vector.insert %296, %295 [3] : i32 into vector<4xi32>
            %298 = vector.extract %38[0, 3] : vector<1x4xi32>
            %299 = arith.muli %289, %297 : vector<4xi32>
            %300 = vector.reduction <add>, %299, %298 : vector<4xi32> into i32
            %301 = vector.insert %300, %cst_0 [0] : i32 into vector<1xi32>
            %302 = vector.insert %301, %cst [0] : vector<1xi32> into vector<1x1xi32>
            %303 = vector.insert_strided_slice %62, %cst_3 {offsets = [0, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %304 = vector.insert_strided_slice %78, %303 {offsets = [0, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %305 = vector.insert_strided_slice %94, %304 {offsets = [0, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %306 = vector.insert_strided_slice %110, %305 {offsets = [0, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %307 = vector.insert_strided_slice %126, %306 {offsets = [1, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %308 = vector.insert_strided_slice %142, %307 {offsets = [1, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %309 = vector.insert_strided_slice %158, %308 {offsets = [1, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %310 = vector.insert_strided_slice %174, %309 {offsets = [1, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %311 = vector.insert_strided_slice %190, %310 {offsets = [2, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %312 = vector.insert_strided_slice %206, %311 {offsets = [2, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %313 = vector.insert_strided_slice %222, %312 {offsets = [2, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %314 = vector.insert_strided_slice %238, %313 {offsets = [2, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %315 = vector.insert_strided_slice %254, %314 {offsets = [3, 0], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %316 = vector.insert_strided_slice %270, %315 {offsets = [3, 1], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %317 = vector.insert_strided_slice %286, %316 {offsets = [3, 2], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %318 = vector.insert_strided_slice %302, %317 {offsets = [3, 3], strides = [1, 1]} : vector<1x1xi32> into vector<4x4xi32>
            %319 = vector.extract_strided_slice %318 {offsets = [0, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %320 = vector.transfer_write %319, %arg7[%c0, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %321 = vector.extract_strided_slice %318 {offsets = [1, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %322 = vector.transfer_write %321, %320[%c1, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %323 = vector.extract_strided_slice %318 {offsets = [2, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %324 = vector.transfer_write %323, %322[%c2, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            %325 = vector.extract_strided_slice %318 {offsets = [3, 0], sizes = [1, 4], strides = [1, 1]} : vector<4x4xi32> to vector<1x4xi32>
            %326 = vector.transfer_write %325, %324[%c3, %c0] {in_bounds = [true, true]} : vector<1x4xi32>, tensor<4x4xi32>
            scf.yield %326 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %20 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After trimming leading unit dims ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %c0_i32 = arith.constant 0 : i32
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg1 = %5 to %c1024 step %6 {
      %7 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %8 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %7) -> (tensor<8x32xi32>) {
        %11 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %12 = vector.transfer_write %cst_1, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %13 = vector.transfer_write %cst_1, %12[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %14 = vector.transfer_write %cst_1, %13[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %15 = vector.transfer_write %cst_1, %14[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %16 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %15) -> (tensor<4x4xi32>) {
            %17 = vector.transfer_read %8[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %18 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
            %19 = vector.transfer_read %8[%18, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %20 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
            %21 = vector.transfer_read %8[%20, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %22 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
            %23 = vector.transfer_read %8[%22, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = vector.transfer_read %arg7[%c0, %c0], %c0_i32 {in_bounds = [true]} : tensor<4x4xi32>, vector<4xi32>
            %32 = vector.transfer_read %arg7[%c1, %c0], %c0_i32 {in_bounds = [true]} : tensor<4x4xi32>, vector<4xi32>
            %33 = vector.transfer_read %arg7[%c2, %c0], %c0_i32 {in_bounds = [true]} : tensor<4x4xi32>, vector<4xi32>
            %34 = vector.transfer_read %arg7[%c3, %c0], %c0_i32 {in_bounds = [true]} : tensor<4x4xi32>, vector<4xi32>
            %35 = arith.extsi %17 : vector<4xi8> to vector<4xi32>
            %36 = arith.extsi %19 : vector<4xi8> to vector<4xi32>
            %37 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
            %38 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
            %39 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [1] : vector<1xi8> into vector<4x1xi8>
            %43 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %44 = vector.insert %43, %42 [2] : vector<1xi8> into vector<4x1xi8>
            %45 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %46 = vector.insert %45, %44 [3] : vector<1xi8> into vector<4x1xi8>
            %47 = arith.extsi %46 : vector<4x1xi8> to vector<4x1xi32>
            %48 = vector.extract %47[0, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %cst_1 [0] : i32 into vector<4xi32>
            %50 = vector.extract %47[1, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [1] : i32 into vector<4xi32>
            %52 = vector.extract %47[2, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [2] : i32 into vector<4xi32>
            %54 = vector.extract %47[3, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [3] : i32 into vector<4xi32>
            %56 = vector.extract %31[0] : vector<4xi32>
            %57 = arith.muli %35, %55 : vector<4xi32>
            %58 = vector.reduction <add>, %57, %56 : vector<4xi32> into i32
            %59 = vector.insert %58, %cst_0 [0] : i32 into vector<1xi32>
            %60 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [1] : vector<1xi8> into vector<4x1xi8>
            %64 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %65 = vector.insert %64, %63 [2] : vector<1xi8> into vector<4x1xi8>
            %66 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %67 = vector.insert %66, %65 [3] : vector<1xi8> into vector<4x1xi8>
            %68 = arith.extsi %67 : vector<4x1xi8> to vector<4x1xi32>
            %69 = vector.extract %68[0, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %cst_1 [0] : i32 into vector<4xi32>
            %71 = vector.extract %68[1, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [1] : i32 into vector<4xi32>
            %73 = vector.extract %68[2, 0] : vector<4x1xi32>
            %74 = vector.insert %73, %72 [2] : i32 into vector<4xi32>
            %75 = vector.extract %68[3, 0] : vector<4x1xi32>
            %76 = vector.insert %75, %74 [3] : i32 into vector<4xi32>
            %77 = vector.extract %31[1] : vector<4xi32>
            %78 = arith.muli %35, %76 : vector<4xi32>
            %79 = vector.reduction <add>, %78, %77 : vector<4xi32> into i32
            %80 = vector.insert %79, %cst_0 [0] : i32 into vector<1xi32>
            %81 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [1] : vector<1xi8> into vector<4x1xi8>
            %85 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %86 = vector.insert %85, %84 [2] : vector<1xi8> into vector<4x1xi8>
            %87 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %88 = vector.insert %87, %86 [3] : vector<1xi8> into vector<4x1xi8>
            %89 = arith.extsi %88 : vector<4x1xi8> to vector<4x1xi32>
            %90 = vector.extract %89[0, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %cst_1 [0] : i32 into vector<4xi32>
            %92 = vector.extract %89[1, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [1] : i32 into vector<4xi32>
            %94 = vector.extract %89[2, 0] : vector<4x1xi32>
            %95 = vector.insert %94, %93 [2] : i32 into vector<4xi32>
            %96 = vector.extract %89[3, 0] : vector<4x1xi32>
            %97 = vector.insert %96, %95 [3] : i32 into vector<4xi32>
            %98 = vector.extract %31[2] : vector<4xi32>
            %99 = arith.muli %35, %97 : vector<4xi32>
            %100 = vector.reduction <add>, %99, %98 : vector<4xi32> into i32
            %101 = vector.insert %100, %cst_0 [0] : i32 into vector<1xi32>
            %102 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [1] : vector<1xi8> into vector<4x1xi8>
            %106 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %107 = vector.insert %106, %105 [2] : vector<1xi8> into vector<4x1xi8>
            %108 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %109 = vector.insert %108, %107 [3] : vector<1xi8> into vector<4x1xi8>
            %110 = arith.extsi %109 : vector<4x1xi8> to vector<4x1xi32>
            %111 = vector.extract %110[0, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %cst_1 [0] : i32 into vector<4xi32>
            %113 = vector.extract %110[1, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [1] : i32 into vector<4xi32>
            %115 = vector.extract %110[2, 0] : vector<4x1xi32>
            %116 = vector.insert %115, %114 [2] : i32 into vector<4xi32>
            %117 = vector.extract %110[3, 0] : vector<4x1xi32>
            %118 = vector.insert %117, %116 [3] : i32 into vector<4xi32>
            %119 = vector.extract %31[3] : vector<4xi32>
            %120 = arith.muli %35, %118 : vector<4xi32>
            %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
            %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
            %123 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [1] : vector<1xi8> into vector<4x1xi8>
            %127 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %128 = vector.insert %127, %126 [2] : vector<1xi8> into vector<4x1xi8>
            %129 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %130 = vector.insert %129, %128 [3] : vector<1xi8> into vector<4x1xi8>
            %131 = arith.extsi %130 : vector<4x1xi8> to vector<4x1xi32>
            %132 = vector.extract %131[0, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %cst_1 [0] : i32 into vector<4xi32>
            %134 = vector.extract %131[1, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [1] : i32 into vector<4xi32>
            %136 = vector.extract %131[2, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [2] : i32 into vector<4xi32>
            %138 = vector.extract %131[3, 0] : vector<4x1xi32>
            %139 = vector.insert %138, %137 [3] : i32 into vector<4xi32>
            %140 = vector.extract %32[0] : vector<4xi32>
            %141 = arith.muli %36, %139 : vector<4xi32>
            %142 = vector.reduction <add>, %141, %140 : vector<4xi32> into i32
            %143 = vector.insert %142, %cst_0 [0] : i32 into vector<1xi32>
            %144 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [1] : vector<1xi8> into vector<4x1xi8>
            %148 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %149 = vector.insert %148, %147 [2] : vector<1xi8> into vector<4x1xi8>
            %150 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %151 = vector.insert %150, %149 [3] : vector<1xi8> into vector<4x1xi8>
            %152 = arith.extsi %151 : vector<4x1xi8> to vector<4x1xi32>
            %153 = vector.extract %152[0, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %cst_1 [0] : i32 into vector<4xi32>
            %155 = vector.extract %152[1, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [1] : i32 into vector<4xi32>
            %157 = vector.extract %152[2, 0] : vector<4x1xi32>
            %158 = vector.insert %157, %156 [2] : i32 into vector<4xi32>
            %159 = vector.extract %152[3, 0] : vector<4x1xi32>
            %160 = vector.insert %159, %158 [3] : i32 into vector<4xi32>
            %161 = vector.extract %32[1] : vector<4xi32>
            %162 = arith.muli %36, %160 : vector<4xi32>
            %163 = vector.reduction <add>, %162, %161 : vector<4xi32> into i32
            %164 = vector.insert %163, %cst_0 [0] : i32 into vector<1xi32>
            %165 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [1] : vector<1xi8> into vector<4x1xi8>
            %169 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %170 = vector.insert %169, %168 [2] : vector<1xi8> into vector<4x1xi8>
            %171 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %172 = vector.insert %171, %170 [3] : vector<1xi8> into vector<4x1xi8>
            %173 = arith.extsi %172 : vector<4x1xi8> to vector<4x1xi32>
            %174 = vector.extract %173[0, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %cst_1 [0] : i32 into vector<4xi32>
            %176 = vector.extract %173[1, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [1] : i32 into vector<4xi32>
            %178 = vector.extract %173[2, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %177 [2] : i32 into vector<4xi32>
            %180 = vector.extract %173[3, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [3] : i32 into vector<4xi32>
            %182 = vector.extract %32[2] : vector<4xi32>
            %183 = arith.muli %36, %181 : vector<4xi32>
            %184 = vector.reduction <add>, %183, %182 : vector<4xi32> into i32
            %185 = vector.insert %184, %cst_0 [0] : i32 into vector<1xi32>
            %186 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [1] : vector<1xi8> into vector<4x1xi8>
            %190 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %191 = vector.insert %190, %189 [2] : vector<1xi8> into vector<4x1xi8>
            %192 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %193 = vector.insert %192, %191 [3] : vector<1xi8> into vector<4x1xi8>
            %194 = arith.extsi %193 : vector<4x1xi8> to vector<4x1xi32>
            %195 = vector.extract %194[0, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %cst_1 [0] : i32 into vector<4xi32>
            %197 = vector.extract %194[1, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [1] : i32 into vector<4xi32>
            %199 = vector.extract %194[2, 0] : vector<4x1xi32>
            %200 = vector.insert %199, %198 [2] : i32 into vector<4xi32>
            %201 = vector.extract %194[3, 0] : vector<4x1xi32>
            %202 = vector.insert %201, %200 [3] : i32 into vector<4xi32>
            %203 = vector.extract %32[3] : vector<4xi32>
            %204 = arith.muli %36, %202 : vector<4xi32>
            %205 = vector.reduction <add>, %204, %203 : vector<4xi32> into i32
            %206 = vector.insert %205, %cst_0 [0] : i32 into vector<1xi32>
            %207 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [1] : vector<1xi8> into vector<4x1xi8>
            %211 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %212 = vector.insert %211, %210 [2] : vector<1xi8> into vector<4x1xi8>
            %213 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %214 = vector.insert %213, %212 [3] : vector<1xi8> into vector<4x1xi8>
            %215 = arith.extsi %214 : vector<4x1xi8> to vector<4x1xi32>
            %216 = vector.extract %215[0, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %cst_1 [0] : i32 into vector<4xi32>
            %218 = vector.extract %215[1, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [1] : i32 into vector<4xi32>
            %220 = vector.extract %215[2, 0] : vector<4x1xi32>
            %221 = vector.insert %220, %219 [2] : i32 into vector<4xi32>
            %222 = vector.extract %215[3, 0] : vector<4x1xi32>
            %223 = vector.insert %222, %221 [3] : i32 into vector<4xi32>
            %224 = vector.extract %33[0] : vector<4xi32>
            %225 = arith.muli %37, %223 : vector<4xi32>
            %226 = vector.reduction <add>, %225, %224 : vector<4xi32> into i32
            %227 = vector.insert %226, %cst_0 [0] : i32 into vector<1xi32>
            %228 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [1] : vector<1xi8> into vector<4x1xi8>
            %232 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %233 = vector.insert %232, %231 [2] : vector<1xi8> into vector<4x1xi8>
            %234 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %235 = vector.insert %234, %233 [3] : vector<1xi8> into vector<4x1xi8>
            %236 = arith.extsi %235 : vector<4x1xi8> to vector<4x1xi32>
            %237 = vector.extract %236[0, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %cst_1 [0] : i32 into vector<4xi32>
            %239 = vector.extract %236[1, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [1] : i32 into vector<4xi32>
            %241 = vector.extract %236[2, 0] : vector<4x1xi32>
            %242 = vector.insert %241, %240 [2] : i32 into vector<4xi32>
            %243 = vector.extract %236[3, 0] : vector<4x1xi32>
            %244 = vector.insert %243, %242 [3] : i32 into vector<4xi32>
            %245 = vector.extract %33[1] : vector<4xi32>
            %246 = arith.muli %37, %244 : vector<4xi32>
            %247 = vector.reduction <add>, %246, %245 : vector<4xi32> into i32
            %248 = vector.insert %247, %cst_0 [0] : i32 into vector<1xi32>
            %249 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [1] : vector<1xi8> into vector<4x1xi8>
            %253 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %254 = vector.insert %253, %252 [2] : vector<1xi8> into vector<4x1xi8>
            %255 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %256 = vector.insert %255, %254 [3] : vector<1xi8> into vector<4x1xi8>
            %257 = arith.extsi %256 : vector<4x1xi8> to vector<4x1xi32>
            %258 = vector.extract %257[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %257[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %257[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %257[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %33[2] : vector<4xi32>
            %267 = arith.muli %37, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [1] : vector<1xi8> into vector<4x1xi8>
            %274 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %275 = vector.insert %274, %273 [2] : vector<1xi8> into vector<4x1xi8>
            %276 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %277 = vector.insert %276, %275 [3] : vector<1xi8> into vector<4x1xi8>
            %278 = arith.extsi %277 : vector<4x1xi8> to vector<4x1xi32>
            %279 = vector.extract %278[0, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %cst_1 [0] : i32 into vector<4xi32>
            %281 = vector.extract %278[1, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [1] : i32 into vector<4xi32>
            %283 = vector.extract %278[2, 0] : vector<4x1xi32>
            %284 = vector.insert %283, %282 [2] : i32 into vector<4xi32>
            %285 = vector.extract %278[3, 0] : vector<4x1xi32>
            %286 = vector.insert %285, %284 [3] : i32 into vector<4xi32>
            %287 = vector.extract %33[3] : vector<4xi32>
            %288 = arith.muli %37, %286 : vector<4xi32>
            %289 = vector.reduction <add>, %288, %287 : vector<4xi32> into i32
            %290 = vector.insert %289, %cst_0 [0] : i32 into vector<1xi32>
            %291 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [1] : vector<1xi8> into vector<4x1xi8>
            %295 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %296 = vector.insert %295, %294 [2] : vector<1xi8> into vector<4x1xi8>
            %297 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %298 = vector.insert %297, %296 [3] : vector<1xi8> into vector<4x1xi8>
            %299 = arith.extsi %298 : vector<4x1xi8> to vector<4x1xi32>
            %300 = vector.extract %299[0, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %cst_1 [0] : i32 into vector<4xi32>
            %302 = vector.extract %299[1, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [1] : i32 into vector<4xi32>
            %304 = vector.extract %299[2, 0] : vector<4x1xi32>
            %305 = vector.insert %304, %303 [2] : i32 into vector<4xi32>
            %306 = vector.extract %299[3, 0] : vector<4x1xi32>
            %307 = vector.insert %306, %305 [3] : i32 into vector<4xi32>
            %308 = vector.extract %34[0] : vector<4xi32>
            %309 = arith.muli %38, %307 : vector<4xi32>
            %310 = vector.reduction <add>, %309, %308 : vector<4xi32> into i32
            %311 = vector.insert %310, %cst_0 [0] : i32 into vector<1xi32>
            %312 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [1] : vector<1xi8> into vector<4x1xi8>
            %316 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %317 = vector.insert %316, %315 [2] : vector<1xi8> into vector<4x1xi8>
            %318 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %319 = vector.insert %318, %317 [3] : vector<1xi8> into vector<4x1xi8>
            %320 = arith.extsi %319 : vector<4x1xi8> to vector<4x1xi32>
            %321 = vector.extract %320[0, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %cst_1 [0] : i32 into vector<4xi32>
            %323 = vector.extract %320[1, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [1] : i32 into vector<4xi32>
            %325 = vector.extract %320[2, 0] : vector<4x1xi32>
            %326 = vector.insert %325, %324 [2] : i32 into vector<4xi32>
            %327 = vector.extract %320[3, 0] : vector<4x1xi32>
            %328 = vector.insert %327, %326 [3] : i32 into vector<4xi32>
            %329 = vector.extract %34[1] : vector<4xi32>
            %330 = arith.muli %38, %328 : vector<4xi32>
            %331 = vector.reduction <add>, %330, %329 : vector<4xi32> into i32
            %332 = vector.insert %331, %cst_0 [0] : i32 into vector<1xi32>
            %333 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [1] : vector<1xi8> into vector<4x1xi8>
            %337 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %338 = vector.insert %337, %336 [2] : vector<1xi8> into vector<4x1xi8>
            %339 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %340 = vector.insert %339, %338 [3] : vector<1xi8> into vector<4x1xi8>
            %341 = arith.extsi %340 : vector<4x1xi8> to vector<4x1xi32>
            %342 = vector.extract %341[0, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %cst_1 [0] : i32 into vector<4xi32>
            %344 = vector.extract %341[1, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [1] : i32 into vector<4xi32>
            %346 = vector.extract %341[2, 0] : vector<4x1xi32>
            %347 = vector.insert %346, %345 [2] : i32 into vector<4xi32>
            %348 = vector.extract %341[3, 0] : vector<4x1xi32>
            %349 = vector.insert %348, %347 [3] : i32 into vector<4xi32>
            %350 = vector.extract %34[2] : vector<4xi32>
            %351 = arith.muli %38, %349 : vector<4xi32>
            %352 = vector.reduction <add>, %351, %350 : vector<4xi32> into i32
            %353 = vector.insert %352, %cst_0 [0] : i32 into vector<1xi32>
            %354 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [1] : vector<1xi8> into vector<4x1xi8>
            %358 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %359 = vector.insert %358, %357 [2] : vector<1xi8> into vector<4x1xi8>
            %360 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %361 = vector.insert %360, %359 [3] : vector<1xi8> into vector<4x1xi8>
            %362 = arith.extsi %361 : vector<4x1xi8> to vector<4x1xi32>
            %363 = vector.extract %362[0, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %cst_1 [0] : i32 into vector<4xi32>
            %365 = vector.extract %362[1, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [1] : i32 into vector<4xi32>
            %367 = vector.extract %362[2, 0] : vector<4x1xi32>
            %368 = vector.insert %367, %366 [2] : i32 into vector<4xi32>
            %369 = vector.extract %362[3, 0] : vector<4x1xi32>
            %370 = vector.insert %369, %368 [3] : i32 into vector<4xi32>
            %371 = vector.extract %34[3] : vector<4xi32>
            %372 = arith.muli %38, %370 : vector<4xi32>
            %373 = vector.reduction <add>, %372, %371 : vector<4xi32> into i32
            %374 = vector.insert %373, %cst_0 [0] : i32 into vector<1xi32>
            %375 = vector.insert_strided_slice %59, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %80, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %101, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %122, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %143, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %164, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %185, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %206, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %227, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %248, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %269, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %290, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %387 = vector.insert_strided_slice %311, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %388 = vector.insert_strided_slice %332, %387 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %389 = vector.insert_strided_slice %353, %388 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %390 = vector.insert_strided_slice %374, %389 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %391 = vector.transfer_write %378, %arg7[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %392 = vector.transfer_write %382, %391[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %393 = vector.transfer_write %386, %392[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %394 = vector.transfer_write %390, %393[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            scf.yield %394 : tensor<4x4xi32>
          }
          %inserted_slice = tensor.insert_slice %16 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %11 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After hoisting transfers ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
  %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
    scf.for %arg1 = %5 to %c1024 step %6 {
      %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
        %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
        %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
        %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
        %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %15 = vector.transfer_write %cst_1, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %16 = vector.transfer_write %cst_1, %15[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %17 = vector.transfer_write %cst_1, %16[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %18 = vector.transfer_write %cst_1, %17[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %19:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
            %24 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %25 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %26 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %27 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %35 = arith.extsi %24 : vector<4xi8> to vector<4xi32>
            %36 = arith.extsi %25 : vector<4xi8> to vector<4xi32>
            %37 = arith.extsi %26 : vector<4xi8> to vector<4xi32>
            %38 = arith.extsi %27 : vector<4xi8> to vector<4xi32>
            %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [1] : vector<1xi8> into vector<4x1xi8>
            %43 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %44 = vector.insert %43, %42 [2] : vector<1xi8> into vector<4x1xi8>
            %45 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %46 = vector.insert %45, %44 [3] : vector<1xi8> into vector<4x1xi8>
            %47 = arith.extsi %46 : vector<4x1xi8> to vector<4x1xi32>
            %48 = vector.extract %47[0, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %cst_1 [0] : i32 into vector<4xi32>
            %50 = vector.extract %47[1, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [1] : i32 into vector<4xi32>
            %52 = vector.extract %47[2, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [2] : i32 into vector<4xi32>
            %54 = vector.extract %47[3, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [3] : i32 into vector<4xi32>
            %56 = vector.extract %arg10[0] : vector<4xi32>
            %57 = arith.muli %35, %55 : vector<4xi32>
            %58 = vector.reduction <add>, %57, %56 : vector<4xi32> into i32
            %59 = vector.insert %58, %cst_0 [0] : i32 into vector<1xi32>
            %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [1] : vector<1xi8> into vector<4x1xi8>
            %64 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %65 = vector.insert %64, %63 [2] : vector<1xi8> into vector<4x1xi8>
            %66 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %67 = vector.insert %66, %65 [3] : vector<1xi8> into vector<4x1xi8>
            %68 = arith.extsi %67 : vector<4x1xi8> to vector<4x1xi32>
            %69 = vector.extract %68[0, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %cst_1 [0] : i32 into vector<4xi32>
            %71 = vector.extract %68[1, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [1] : i32 into vector<4xi32>
            %73 = vector.extract %68[2, 0] : vector<4x1xi32>
            %74 = vector.insert %73, %72 [2] : i32 into vector<4xi32>
            %75 = vector.extract %68[3, 0] : vector<4x1xi32>
            %76 = vector.insert %75, %74 [3] : i32 into vector<4xi32>
            %77 = vector.extract %arg10[1] : vector<4xi32>
            %78 = arith.muli %35, %76 : vector<4xi32>
            %79 = vector.reduction <add>, %78, %77 : vector<4xi32> into i32
            %80 = vector.insert %79, %cst_0 [0] : i32 into vector<1xi32>
            %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [1] : vector<1xi8> into vector<4x1xi8>
            %85 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %86 = vector.insert %85, %84 [2] : vector<1xi8> into vector<4x1xi8>
            %87 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %88 = vector.insert %87, %86 [3] : vector<1xi8> into vector<4x1xi8>
            %89 = arith.extsi %88 : vector<4x1xi8> to vector<4x1xi32>
            %90 = vector.extract %89[0, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %cst_1 [0] : i32 into vector<4xi32>
            %92 = vector.extract %89[1, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [1] : i32 into vector<4xi32>
            %94 = vector.extract %89[2, 0] : vector<4x1xi32>
            %95 = vector.insert %94, %93 [2] : i32 into vector<4xi32>
            %96 = vector.extract %89[3, 0] : vector<4x1xi32>
            %97 = vector.insert %96, %95 [3] : i32 into vector<4xi32>
            %98 = vector.extract %arg10[2] : vector<4xi32>
            %99 = arith.muli %35, %97 : vector<4xi32>
            %100 = vector.reduction <add>, %99, %98 : vector<4xi32> into i32
            %101 = vector.insert %100, %cst_0 [0] : i32 into vector<1xi32>
            %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [1] : vector<1xi8> into vector<4x1xi8>
            %106 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %107 = vector.insert %106, %105 [2] : vector<1xi8> into vector<4x1xi8>
            %108 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %109 = vector.insert %108, %107 [3] : vector<1xi8> into vector<4x1xi8>
            %110 = arith.extsi %109 : vector<4x1xi8> to vector<4x1xi32>
            %111 = vector.extract %110[0, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %cst_1 [0] : i32 into vector<4xi32>
            %113 = vector.extract %110[1, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [1] : i32 into vector<4xi32>
            %115 = vector.extract %110[2, 0] : vector<4x1xi32>
            %116 = vector.insert %115, %114 [2] : i32 into vector<4xi32>
            %117 = vector.extract %110[3, 0] : vector<4x1xi32>
            %118 = vector.insert %117, %116 [3] : i32 into vector<4xi32>
            %119 = vector.extract %arg10[3] : vector<4xi32>
            %120 = arith.muli %35, %118 : vector<4xi32>
            %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
            %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
            %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [1] : vector<1xi8> into vector<4x1xi8>
            %127 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %128 = vector.insert %127, %126 [2] : vector<1xi8> into vector<4x1xi8>
            %129 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %130 = vector.insert %129, %128 [3] : vector<1xi8> into vector<4x1xi8>
            %131 = arith.extsi %130 : vector<4x1xi8> to vector<4x1xi32>
            %132 = vector.extract %131[0, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %cst_1 [0] : i32 into vector<4xi32>
            %134 = vector.extract %131[1, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [1] : i32 into vector<4xi32>
            %136 = vector.extract %131[2, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [2] : i32 into vector<4xi32>
            %138 = vector.extract %131[3, 0] : vector<4x1xi32>
            %139 = vector.insert %138, %137 [3] : i32 into vector<4xi32>
            %140 = vector.extract %arg9[0] : vector<4xi32>
            %141 = arith.muli %36, %139 : vector<4xi32>
            %142 = vector.reduction <add>, %141, %140 : vector<4xi32> into i32
            %143 = vector.insert %142, %cst_0 [0] : i32 into vector<1xi32>
            %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [1] : vector<1xi8> into vector<4x1xi8>
            %148 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %149 = vector.insert %148, %147 [2] : vector<1xi8> into vector<4x1xi8>
            %150 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %151 = vector.insert %150, %149 [3] : vector<1xi8> into vector<4x1xi8>
            %152 = arith.extsi %151 : vector<4x1xi8> to vector<4x1xi32>
            %153 = vector.extract %152[0, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %cst_1 [0] : i32 into vector<4xi32>
            %155 = vector.extract %152[1, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [1] : i32 into vector<4xi32>
            %157 = vector.extract %152[2, 0] : vector<4x1xi32>
            %158 = vector.insert %157, %156 [2] : i32 into vector<4xi32>
            %159 = vector.extract %152[3, 0] : vector<4x1xi32>
            %160 = vector.insert %159, %158 [3] : i32 into vector<4xi32>
            %161 = vector.extract %arg9[1] : vector<4xi32>
            %162 = arith.muli %36, %160 : vector<4xi32>
            %163 = vector.reduction <add>, %162, %161 : vector<4xi32> into i32
            %164 = vector.insert %163, %cst_0 [0] : i32 into vector<1xi32>
            %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [1] : vector<1xi8> into vector<4x1xi8>
            %169 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %170 = vector.insert %169, %168 [2] : vector<1xi8> into vector<4x1xi8>
            %171 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %172 = vector.insert %171, %170 [3] : vector<1xi8> into vector<4x1xi8>
            %173 = arith.extsi %172 : vector<4x1xi8> to vector<4x1xi32>
            %174 = vector.extract %173[0, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %cst_1 [0] : i32 into vector<4xi32>
            %176 = vector.extract %173[1, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [1] : i32 into vector<4xi32>
            %178 = vector.extract %173[2, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %177 [2] : i32 into vector<4xi32>
            %180 = vector.extract %173[3, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [3] : i32 into vector<4xi32>
            %182 = vector.extract %arg9[2] : vector<4xi32>
            %183 = arith.muli %36, %181 : vector<4xi32>
            %184 = vector.reduction <add>, %183, %182 : vector<4xi32> into i32
            %185 = vector.insert %184, %cst_0 [0] : i32 into vector<1xi32>
            %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [1] : vector<1xi8> into vector<4x1xi8>
            %190 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %191 = vector.insert %190, %189 [2] : vector<1xi8> into vector<4x1xi8>
            %192 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %193 = vector.insert %192, %191 [3] : vector<1xi8> into vector<4x1xi8>
            %194 = arith.extsi %193 : vector<4x1xi8> to vector<4x1xi32>
            %195 = vector.extract %194[0, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %cst_1 [0] : i32 into vector<4xi32>
            %197 = vector.extract %194[1, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [1] : i32 into vector<4xi32>
            %199 = vector.extract %194[2, 0] : vector<4x1xi32>
            %200 = vector.insert %199, %198 [2] : i32 into vector<4xi32>
            %201 = vector.extract %194[3, 0] : vector<4x1xi32>
            %202 = vector.insert %201, %200 [3] : i32 into vector<4xi32>
            %203 = vector.extract %arg9[3] : vector<4xi32>
            %204 = arith.muli %36, %202 : vector<4xi32>
            %205 = vector.reduction <add>, %204, %203 : vector<4xi32> into i32
            %206 = vector.insert %205, %cst_0 [0] : i32 into vector<1xi32>
            %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [1] : vector<1xi8> into vector<4x1xi8>
            %211 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %212 = vector.insert %211, %210 [2] : vector<1xi8> into vector<4x1xi8>
            %213 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %214 = vector.insert %213, %212 [3] : vector<1xi8> into vector<4x1xi8>
            %215 = arith.extsi %214 : vector<4x1xi8> to vector<4x1xi32>
            %216 = vector.extract %215[0, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %cst_1 [0] : i32 into vector<4xi32>
            %218 = vector.extract %215[1, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [1] : i32 into vector<4xi32>
            %220 = vector.extract %215[2, 0] : vector<4x1xi32>
            %221 = vector.insert %220, %219 [2] : i32 into vector<4xi32>
            %222 = vector.extract %215[3, 0] : vector<4x1xi32>
            %223 = vector.insert %222, %221 [3] : i32 into vector<4xi32>
            %224 = vector.extract %arg8[0] : vector<4xi32>
            %225 = arith.muli %37, %223 : vector<4xi32>
            %226 = vector.reduction <add>, %225, %224 : vector<4xi32> into i32
            %227 = vector.insert %226, %cst_0 [0] : i32 into vector<1xi32>
            %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [1] : vector<1xi8> into vector<4x1xi8>
            %232 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %233 = vector.insert %232, %231 [2] : vector<1xi8> into vector<4x1xi8>
            %234 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %235 = vector.insert %234, %233 [3] : vector<1xi8> into vector<4x1xi8>
            %236 = arith.extsi %235 : vector<4x1xi8> to vector<4x1xi32>
            %237 = vector.extract %236[0, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %cst_1 [0] : i32 into vector<4xi32>
            %239 = vector.extract %236[1, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [1] : i32 into vector<4xi32>
            %241 = vector.extract %236[2, 0] : vector<4x1xi32>
            %242 = vector.insert %241, %240 [2] : i32 into vector<4xi32>
            %243 = vector.extract %236[3, 0] : vector<4x1xi32>
            %244 = vector.insert %243, %242 [3] : i32 into vector<4xi32>
            %245 = vector.extract %arg8[1] : vector<4xi32>
            %246 = arith.muli %37, %244 : vector<4xi32>
            %247 = vector.reduction <add>, %246, %245 : vector<4xi32> into i32
            %248 = vector.insert %247, %cst_0 [0] : i32 into vector<1xi32>
            %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [1] : vector<1xi8> into vector<4x1xi8>
            %253 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %254 = vector.insert %253, %252 [2] : vector<1xi8> into vector<4x1xi8>
            %255 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %256 = vector.insert %255, %254 [3] : vector<1xi8> into vector<4x1xi8>
            %257 = arith.extsi %256 : vector<4x1xi8> to vector<4x1xi32>
            %258 = vector.extract %257[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %257[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %257[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %257[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %arg8[2] : vector<4xi32>
            %267 = arith.muli %37, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [1] : vector<1xi8> into vector<4x1xi8>
            %274 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %275 = vector.insert %274, %273 [2] : vector<1xi8> into vector<4x1xi8>
            %276 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %277 = vector.insert %276, %275 [3] : vector<1xi8> into vector<4x1xi8>
            %278 = arith.extsi %277 : vector<4x1xi8> to vector<4x1xi32>
            %279 = vector.extract %278[0, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %cst_1 [0] : i32 into vector<4xi32>
            %281 = vector.extract %278[1, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [1] : i32 into vector<4xi32>
            %283 = vector.extract %278[2, 0] : vector<4x1xi32>
            %284 = vector.insert %283, %282 [2] : i32 into vector<4xi32>
            %285 = vector.extract %278[3, 0] : vector<4x1xi32>
            %286 = vector.insert %285, %284 [3] : i32 into vector<4xi32>
            %287 = vector.extract %arg8[3] : vector<4xi32>
            %288 = arith.muli %37, %286 : vector<4xi32>
            %289 = vector.reduction <add>, %288, %287 : vector<4xi32> into i32
            %290 = vector.insert %289, %cst_0 [0] : i32 into vector<1xi32>
            %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [1] : vector<1xi8> into vector<4x1xi8>
            %295 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %296 = vector.insert %295, %294 [2] : vector<1xi8> into vector<4x1xi8>
            %297 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %298 = vector.insert %297, %296 [3] : vector<1xi8> into vector<4x1xi8>
            %299 = arith.extsi %298 : vector<4x1xi8> to vector<4x1xi32>
            %300 = vector.extract %299[0, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %cst_1 [0] : i32 into vector<4xi32>
            %302 = vector.extract %299[1, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [1] : i32 into vector<4xi32>
            %304 = vector.extract %299[2, 0] : vector<4x1xi32>
            %305 = vector.insert %304, %303 [2] : i32 into vector<4xi32>
            %306 = vector.extract %299[3, 0] : vector<4x1xi32>
            %307 = vector.insert %306, %305 [3] : i32 into vector<4xi32>
            %308 = vector.extract %arg7[0] : vector<4xi32>
            %309 = arith.muli %38, %307 : vector<4xi32>
            %310 = vector.reduction <add>, %309, %308 : vector<4xi32> into i32
            %311 = vector.insert %310, %cst_0 [0] : i32 into vector<1xi32>
            %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [1] : vector<1xi8> into vector<4x1xi8>
            %316 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %317 = vector.insert %316, %315 [2] : vector<1xi8> into vector<4x1xi8>
            %318 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %319 = vector.insert %318, %317 [3] : vector<1xi8> into vector<4x1xi8>
            %320 = arith.extsi %319 : vector<4x1xi8> to vector<4x1xi32>
            %321 = vector.extract %320[0, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %cst_1 [0] : i32 into vector<4xi32>
            %323 = vector.extract %320[1, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [1] : i32 into vector<4xi32>
            %325 = vector.extract %320[2, 0] : vector<4x1xi32>
            %326 = vector.insert %325, %324 [2] : i32 into vector<4xi32>
            %327 = vector.extract %320[3, 0] : vector<4x1xi32>
            %328 = vector.insert %327, %326 [3] : i32 into vector<4xi32>
            %329 = vector.extract %arg7[1] : vector<4xi32>
            %330 = arith.muli %38, %328 : vector<4xi32>
            %331 = vector.reduction <add>, %330, %329 : vector<4xi32> into i32
            %332 = vector.insert %331, %cst_0 [0] : i32 into vector<1xi32>
            %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [1] : vector<1xi8> into vector<4x1xi8>
            %337 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %338 = vector.insert %337, %336 [2] : vector<1xi8> into vector<4x1xi8>
            %339 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %340 = vector.insert %339, %338 [3] : vector<1xi8> into vector<4x1xi8>
            %341 = arith.extsi %340 : vector<4x1xi8> to vector<4x1xi32>
            %342 = vector.extract %341[0, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %cst_1 [0] : i32 into vector<4xi32>
            %344 = vector.extract %341[1, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [1] : i32 into vector<4xi32>
            %346 = vector.extract %341[2, 0] : vector<4x1xi32>
            %347 = vector.insert %346, %345 [2] : i32 into vector<4xi32>
            %348 = vector.extract %341[3, 0] : vector<4x1xi32>
            %349 = vector.insert %348, %347 [3] : i32 into vector<4xi32>
            %350 = vector.extract %arg7[2] : vector<4xi32>
            %351 = arith.muli %38, %349 : vector<4xi32>
            %352 = vector.reduction <add>, %351, %350 : vector<4xi32> into i32
            %353 = vector.insert %352, %cst_0 [0] : i32 into vector<1xi32>
            %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [1] : vector<1xi8> into vector<4x1xi8>
            %358 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %359 = vector.insert %358, %357 [2] : vector<1xi8> into vector<4x1xi8>
            %360 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %361 = vector.insert %360, %359 [3] : vector<1xi8> into vector<4x1xi8>
            %362 = arith.extsi %361 : vector<4x1xi8> to vector<4x1xi32>
            %363 = vector.extract %362[0, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %cst_1 [0] : i32 into vector<4xi32>
            %365 = vector.extract %362[1, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [1] : i32 into vector<4xi32>
            %367 = vector.extract %362[2, 0] : vector<4x1xi32>
            %368 = vector.insert %367, %366 [2] : i32 into vector<4xi32>
            %369 = vector.extract %362[3, 0] : vector<4x1xi32>
            %370 = vector.insert %369, %368 [3] : i32 into vector<4xi32>
            %371 = vector.extract %arg7[3] : vector<4xi32>
            %372 = arith.muli %38, %370 : vector<4xi32>
            %373 = vector.reduction <add>, %372, %371 : vector<4xi32> into i32
            %374 = vector.insert %373, %cst_0 [0] : i32 into vector<1xi32>
            %375 = vector.insert_strided_slice %59, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %80, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %101, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %122, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %143, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %164, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %185, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %206, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %227, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %248, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %269, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %290, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %387 = vector.insert_strided_slice %311, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %388 = vector.insert_strided_slice %332, %387 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %389 = vector.insert_strided_slice %353, %388 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %390 = vector.insert_strided_slice %374, %389 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            scf.yield %390, %386, %382, %378 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
          }
          %20 = vector.transfer_write %19#3, %18[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %21 = vector.transfer_write %19#2, %20[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %22 = vector.transfer_write %19#1, %21[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %23 = vector.transfer_write %19#0, %22[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %inserted_slice = tensor.insert_slice %23 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %14 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After lowering transfer ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
  %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
    scf.for %arg1 = %5 to %c1024 step %6 {
      %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
        %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
        %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
        %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
        %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %15 = vector.transfer_write %cst_1, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %16 = vector.transfer_write %cst_1, %15[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %17 = vector.transfer_write %cst_1, %16[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %18 = vector.transfer_write %cst_1, %17[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %19:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
            %24 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %25 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %26 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %27 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %35 = arith.extsi %24 : vector<4xi8> to vector<4xi32>
            %36 = arith.extsi %25 : vector<4xi8> to vector<4xi32>
            %37 = arith.extsi %26 : vector<4xi8> to vector<4xi32>
            %38 = arith.extsi %27 : vector<4xi8> to vector<4xi32>
            %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [1] : vector<1xi8> into vector<4x1xi8>
            %43 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %44 = vector.insert %43, %42 [2] : vector<1xi8> into vector<4x1xi8>
            %45 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %46 = vector.insert %45, %44 [3] : vector<1xi8> into vector<4x1xi8>
            %47 = arith.extsi %46 : vector<4x1xi8> to vector<4x1xi32>
            %48 = vector.extract %47[0, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %cst_1 [0] : i32 into vector<4xi32>
            %50 = vector.extract %47[1, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [1] : i32 into vector<4xi32>
            %52 = vector.extract %47[2, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [2] : i32 into vector<4xi32>
            %54 = vector.extract %47[3, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [3] : i32 into vector<4xi32>
            %56 = vector.extract %arg10[0] : vector<4xi32>
            %57 = arith.muli %35, %55 : vector<4xi32>
            %58 = vector.reduction <add>, %57, %56 : vector<4xi32> into i32
            %59 = vector.insert %58, %cst_0 [0] : i32 into vector<1xi32>
            %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [1] : vector<1xi8> into vector<4x1xi8>
            %64 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %65 = vector.insert %64, %63 [2] : vector<1xi8> into vector<4x1xi8>
            %66 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %67 = vector.insert %66, %65 [3] : vector<1xi8> into vector<4x1xi8>
            %68 = arith.extsi %67 : vector<4x1xi8> to vector<4x1xi32>
            %69 = vector.extract %68[0, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %cst_1 [0] : i32 into vector<4xi32>
            %71 = vector.extract %68[1, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [1] : i32 into vector<4xi32>
            %73 = vector.extract %68[2, 0] : vector<4x1xi32>
            %74 = vector.insert %73, %72 [2] : i32 into vector<4xi32>
            %75 = vector.extract %68[3, 0] : vector<4x1xi32>
            %76 = vector.insert %75, %74 [3] : i32 into vector<4xi32>
            %77 = vector.extract %arg10[1] : vector<4xi32>
            %78 = arith.muli %35, %76 : vector<4xi32>
            %79 = vector.reduction <add>, %78, %77 : vector<4xi32> into i32
            %80 = vector.insert %79, %cst_0 [0] : i32 into vector<1xi32>
            %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [1] : vector<1xi8> into vector<4x1xi8>
            %85 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %86 = vector.insert %85, %84 [2] : vector<1xi8> into vector<4x1xi8>
            %87 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %88 = vector.insert %87, %86 [3] : vector<1xi8> into vector<4x1xi8>
            %89 = arith.extsi %88 : vector<4x1xi8> to vector<4x1xi32>
            %90 = vector.extract %89[0, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %cst_1 [0] : i32 into vector<4xi32>
            %92 = vector.extract %89[1, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [1] : i32 into vector<4xi32>
            %94 = vector.extract %89[2, 0] : vector<4x1xi32>
            %95 = vector.insert %94, %93 [2] : i32 into vector<4xi32>
            %96 = vector.extract %89[3, 0] : vector<4x1xi32>
            %97 = vector.insert %96, %95 [3] : i32 into vector<4xi32>
            %98 = vector.extract %arg10[2] : vector<4xi32>
            %99 = arith.muli %35, %97 : vector<4xi32>
            %100 = vector.reduction <add>, %99, %98 : vector<4xi32> into i32
            %101 = vector.insert %100, %cst_0 [0] : i32 into vector<1xi32>
            %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [1] : vector<1xi8> into vector<4x1xi8>
            %106 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %107 = vector.insert %106, %105 [2] : vector<1xi8> into vector<4x1xi8>
            %108 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %109 = vector.insert %108, %107 [3] : vector<1xi8> into vector<4x1xi8>
            %110 = arith.extsi %109 : vector<4x1xi8> to vector<4x1xi32>
            %111 = vector.extract %110[0, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %cst_1 [0] : i32 into vector<4xi32>
            %113 = vector.extract %110[1, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [1] : i32 into vector<4xi32>
            %115 = vector.extract %110[2, 0] : vector<4x1xi32>
            %116 = vector.insert %115, %114 [2] : i32 into vector<4xi32>
            %117 = vector.extract %110[3, 0] : vector<4x1xi32>
            %118 = vector.insert %117, %116 [3] : i32 into vector<4xi32>
            %119 = vector.extract %arg10[3] : vector<4xi32>
            %120 = arith.muli %35, %118 : vector<4xi32>
            %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
            %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
            %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [1] : vector<1xi8> into vector<4x1xi8>
            %127 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %128 = vector.insert %127, %126 [2] : vector<1xi8> into vector<4x1xi8>
            %129 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %130 = vector.insert %129, %128 [3] : vector<1xi8> into vector<4x1xi8>
            %131 = arith.extsi %130 : vector<4x1xi8> to vector<4x1xi32>
            %132 = vector.extract %131[0, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %cst_1 [0] : i32 into vector<4xi32>
            %134 = vector.extract %131[1, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [1] : i32 into vector<4xi32>
            %136 = vector.extract %131[2, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [2] : i32 into vector<4xi32>
            %138 = vector.extract %131[3, 0] : vector<4x1xi32>
            %139 = vector.insert %138, %137 [3] : i32 into vector<4xi32>
            %140 = vector.extract %arg9[0] : vector<4xi32>
            %141 = arith.muli %36, %139 : vector<4xi32>
            %142 = vector.reduction <add>, %141, %140 : vector<4xi32> into i32
            %143 = vector.insert %142, %cst_0 [0] : i32 into vector<1xi32>
            %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [1] : vector<1xi8> into vector<4x1xi8>
            %148 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %149 = vector.insert %148, %147 [2] : vector<1xi8> into vector<4x1xi8>
            %150 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %151 = vector.insert %150, %149 [3] : vector<1xi8> into vector<4x1xi8>
            %152 = arith.extsi %151 : vector<4x1xi8> to vector<4x1xi32>
            %153 = vector.extract %152[0, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %cst_1 [0] : i32 into vector<4xi32>
            %155 = vector.extract %152[1, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [1] : i32 into vector<4xi32>
            %157 = vector.extract %152[2, 0] : vector<4x1xi32>
            %158 = vector.insert %157, %156 [2] : i32 into vector<4xi32>
            %159 = vector.extract %152[3, 0] : vector<4x1xi32>
            %160 = vector.insert %159, %158 [3] : i32 into vector<4xi32>
            %161 = vector.extract %arg9[1] : vector<4xi32>
            %162 = arith.muli %36, %160 : vector<4xi32>
            %163 = vector.reduction <add>, %162, %161 : vector<4xi32> into i32
            %164 = vector.insert %163, %cst_0 [0] : i32 into vector<1xi32>
            %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [1] : vector<1xi8> into vector<4x1xi8>
            %169 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %170 = vector.insert %169, %168 [2] : vector<1xi8> into vector<4x1xi8>
            %171 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %172 = vector.insert %171, %170 [3] : vector<1xi8> into vector<4x1xi8>
            %173 = arith.extsi %172 : vector<4x1xi8> to vector<4x1xi32>
            %174 = vector.extract %173[0, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %cst_1 [0] : i32 into vector<4xi32>
            %176 = vector.extract %173[1, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [1] : i32 into vector<4xi32>
            %178 = vector.extract %173[2, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %177 [2] : i32 into vector<4xi32>
            %180 = vector.extract %173[3, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [3] : i32 into vector<4xi32>
            %182 = vector.extract %arg9[2] : vector<4xi32>
            %183 = arith.muli %36, %181 : vector<4xi32>
            %184 = vector.reduction <add>, %183, %182 : vector<4xi32> into i32
            %185 = vector.insert %184, %cst_0 [0] : i32 into vector<1xi32>
            %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [1] : vector<1xi8> into vector<4x1xi8>
            %190 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %191 = vector.insert %190, %189 [2] : vector<1xi8> into vector<4x1xi8>
            %192 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %193 = vector.insert %192, %191 [3] : vector<1xi8> into vector<4x1xi8>
            %194 = arith.extsi %193 : vector<4x1xi8> to vector<4x1xi32>
            %195 = vector.extract %194[0, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %cst_1 [0] : i32 into vector<4xi32>
            %197 = vector.extract %194[1, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [1] : i32 into vector<4xi32>
            %199 = vector.extract %194[2, 0] : vector<4x1xi32>
            %200 = vector.insert %199, %198 [2] : i32 into vector<4xi32>
            %201 = vector.extract %194[3, 0] : vector<4x1xi32>
            %202 = vector.insert %201, %200 [3] : i32 into vector<4xi32>
            %203 = vector.extract %arg9[3] : vector<4xi32>
            %204 = arith.muli %36, %202 : vector<4xi32>
            %205 = vector.reduction <add>, %204, %203 : vector<4xi32> into i32
            %206 = vector.insert %205, %cst_0 [0] : i32 into vector<1xi32>
            %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [1] : vector<1xi8> into vector<4x1xi8>
            %211 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %212 = vector.insert %211, %210 [2] : vector<1xi8> into vector<4x1xi8>
            %213 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %214 = vector.insert %213, %212 [3] : vector<1xi8> into vector<4x1xi8>
            %215 = arith.extsi %214 : vector<4x1xi8> to vector<4x1xi32>
            %216 = vector.extract %215[0, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %cst_1 [0] : i32 into vector<4xi32>
            %218 = vector.extract %215[1, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [1] : i32 into vector<4xi32>
            %220 = vector.extract %215[2, 0] : vector<4x1xi32>
            %221 = vector.insert %220, %219 [2] : i32 into vector<4xi32>
            %222 = vector.extract %215[3, 0] : vector<4x1xi32>
            %223 = vector.insert %222, %221 [3] : i32 into vector<4xi32>
            %224 = vector.extract %arg8[0] : vector<4xi32>
            %225 = arith.muli %37, %223 : vector<4xi32>
            %226 = vector.reduction <add>, %225, %224 : vector<4xi32> into i32
            %227 = vector.insert %226, %cst_0 [0] : i32 into vector<1xi32>
            %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [1] : vector<1xi8> into vector<4x1xi8>
            %232 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %233 = vector.insert %232, %231 [2] : vector<1xi8> into vector<4x1xi8>
            %234 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %235 = vector.insert %234, %233 [3] : vector<1xi8> into vector<4x1xi8>
            %236 = arith.extsi %235 : vector<4x1xi8> to vector<4x1xi32>
            %237 = vector.extract %236[0, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %cst_1 [0] : i32 into vector<4xi32>
            %239 = vector.extract %236[1, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [1] : i32 into vector<4xi32>
            %241 = vector.extract %236[2, 0] : vector<4x1xi32>
            %242 = vector.insert %241, %240 [2] : i32 into vector<4xi32>
            %243 = vector.extract %236[3, 0] : vector<4x1xi32>
            %244 = vector.insert %243, %242 [3] : i32 into vector<4xi32>
            %245 = vector.extract %arg8[1] : vector<4xi32>
            %246 = arith.muli %37, %244 : vector<4xi32>
            %247 = vector.reduction <add>, %246, %245 : vector<4xi32> into i32
            %248 = vector.insert %247, %cst_0 [0] : i32 into vector<1xi32>
            %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [1] : vector<1xi8> into vector<4x1xi8>
            %253 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %254 = vector.insert %253, %252 [2] : vector<1xi8> into vector<4x1xi8>
            %255 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %256 = vector.insert %255, %254 [3] : vector<1xi8> into vector<4x1xi8>
            %257 = arith.extsi %256 : vector<4x1xi8> to vector<4x1xi32>
            %258 = vector.extract %257[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %257[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %257[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %257[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %arg8[2] : vector<4xi32>
            %267 = arith.muli %37, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [1] : vector<1xi8> into vector<4x1xi8>
            %274 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %275 = vector.insert %274, %273 [2] : vector<1xi8> into vector<4x1xi8>
            %276 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %277 = vector.insert %276, %275 [3] : vector<1xi8> into vector<4x1xi8>
            %278 = arith.extsi %277 : vector<4x1xi8> to vector<4x1xi32>
            %279 = vector.extract %278[0, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %cst_1 [0] : i32 into vector<4xi32>
            %281 = vector.extract %278[1, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [1] : i32 into vector<4xi32>
            %283 = vector.extract %278[2, 0] : vector<4x1xi32>
            %284 = vector.insert %283, %282 [2] : i32 into vector<4xi32>
            %285 = vector.extract %278[3, 0] : vector<4x1xi32>
            %286 = vector.insert %285, %284 [3] : i32 into vector<4xi32>
            %287 = vector.extract %arg8[3] : vector<4xi32>
            %288 = arith.muli %37, %286 : vector<4xi32>
            %289 = vector.reduction <add>, %288, %287 : vector<4xi32> into i32
            %290 = vector.insert %289, %cst_0 [0] : i32 into vector<1xi32>
            %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [1] : vector<1xi8> into vector<4x1xi8>
            %295 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %296 = vector.insert %295, %294 [2] : vector<1xi8> into vector<4x1xi8>
            %297 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %298 = vector.insert %297, %296 [3] : vector<1xi8> into vector<4x1xi8>
            %299 = arith.extsi %298 : vector<4x1xi8> to vector<4x1xi32>
            %300 = vector.extract %299[0, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %cst_1 [0] : i32 into vector<4xi32>
            %302 = vector.extract %299[1, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [1] : i32 into vector<4xi32>
            %304 = vector.extract %299[2, 0] : vector<4x1xi32>
            %305 = vector.insert %304, %303 [2] : i32 into vector<4xi32>
            %306 = vector.extract %299[3, 0] : vector<4x1xi32>
            %307 = vector.insert %306, %305 [3] : i32 into vector<4xi32>
            %308 = vector.extract %arg7[0] : vector<4xi32>
            %309 = arith.muli %38, %307 : vector<4xi32>
            %310 = vector.reduction <add>, %309, %308 : vector<4xi32> into i32
            %311 = vector.insert %310, %cst_0 [0] : i32 into vector<1xi32>
            %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [1] : vector<1xi8> into vector<4x1xi8>
            %316 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %317 = vector.insert %316, %315 [2] : vector<1xi8> into vector<4x1xi8>
            %318 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %319 = vector.insert %318, %317 [3] : vector<1xi8> into vector<4x1xi8>
            %320 = arith.extsi %319 : vector<4x1xi8> to vector<4x1xi32>
            %321 = vector.extract %320[0, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %cst_1 [0] : i32 into vector<4xi32>
            %323 = vector.extract %320[1, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [1] : i32 into vector<4xi32>
            %325 = vector.extract %320[2, 0] : vector<4x1xi32>
            %326 = vector.insert %325, %324 [2] : i32 into vector<4xi32>
            %327 = vector.extract %320[3, 0] : vector<4x1xi32>
            %328 = vector.insert %327, %326 [3] : i32 into vector<4xi32>
            %329 = vector.extract %arg7[1] : vector<4xi32>
            %330 = arith.muli %38, %328 : vector<4xi32>
            %331 = vector.reduction <add>, %330, %329 : vector<4xi32> into i32
            %332 = vector.insert %331, %cst_0 [0] : i32 into vector<1xi32>
            %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [1] : vector<1xi8> into vector<4x1xi8>
            %337 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %338 = vector.insert %337, %336 [2] : vector<1xi8> into vector<4x1xi8>
            %339 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %340 = vector.insert %339, %338 [3] : vector<1xi8> into vector<4x1xi8>
            %341 = arith.extsi %340 : vector<4x1xi8> to vector<4x1xi32>
            %342 = vector.extract %341[0, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %cst_1 [0] : i32 into vector<4xi32>
            %344 = vector.extract %341[1, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [1] : i32 into vector<4xi32>
            %346 = vector.extract %341[2, 0] : vector<4x1xi32>
            %347 = vector.insert %346, %345 [2] : i32 into vector<4xi32>
            %348 = vector.extract %341[3, 0] : vector<4x1xi32>
            %349 = vector.insert %348, %347 [3] : i32 into vector<4xi32>
            %350 = vector.extract %arg7[2] : vector<4xi32>
            %351 = arith.muli %38, %349 : vector<4xi32>
            %352 = vector.reduction <add>, %351, %350 : vector<4xi32> into i32
            %353 = vector.insert %352, %cst_0 [0] : i32 into vector<1xi32>
            %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [1] : vector<1xi8> into vector<4x1xi8>
            %358 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %359 = vector.insert %358, %357 [2] : vector<1xi8> into vector<4x1xi8>
            %360 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %361 = vector.insert %360, %359 [3] : vector<1xi8> into vector<4x1xi8>
            %362 = arith.extsi %361 : vector<4x1xi8> to vector<4x1xi32>
            %363 = vector.extract %362[0, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %cst_1 [0] : i32 into vector<4xi32>
            %365 = vector.extract %362[1, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [1] : i32 into vector<4xi32>
            %367 = vector.extract %362[2, 0] : vector<4x1xi32>
            %368 = vector.insert %367, %366 [2] : i32 into vector<4xi32>
            %369 = vector.extract %362[3, 0] : vector<4x1xi32>
            %370 = vector.insert %369, %368 [3] : i32 into vector<4xi32>
            %371 = vector.extract %arg7[3] : vector<4xi32>
            %372 = arith.muli %38, %370 : vector<4xi32>
            %373 = vector.reduction <add>, %372, %371 : vector<4xi32> into i32
            %374 = vector.insert %373, %cst_0 [0] : i32 into vector<1xi32>
            %375 = vector.insert_strided_slice %59, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %80, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %101, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %122, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %143, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %164, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %185, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %206, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %227, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %248, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %269, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %290, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %387 = vector.insert_strided_slice %311, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %388 = vector.insert_strided_slice %332, %387 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %389 = vector.insert_strided_slice %353, %388 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %390 = vector.insert_strided_slice %374, %389 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            scf.yield %390, %386, %382, %378 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
          }
          %20 = vector.transfer_write %19#3, %18[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %21 = vector.transfer_write %19#2, %20[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %22 = vector.transfer_write %19#1, %21[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %23 = vector.transfer_write %19#0, %22[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %inserted_slice = tensor.insert_slice %23 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %14 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

--- After lowering various vector ops ---
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
  %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
    scf.for %arg1 = %5 to %c1024 step %6 {
      %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
        %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
        %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
        %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
        %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %15 = vector.transfer_write %cst_1, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %16 = vector.transfer_write %cst_1, %15[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %17 = vector.transfer_write %cst_1, %16[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %18 = vector.transfer_write %cst_1, %17[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %19:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
            %24 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %25 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %26 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %27 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %28 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %32 = vector.transfer_read %9[%31, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %33 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %34 = vector.transfer_read %9[%33, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %35 = arith.extsi %24 : vector<4xi8> to vector<4xi32>
            %36 = arith.extsi %25 : vector<4xi8> to vector<4xi32>
            %37 = arith.extsi %26 : vector<4xi8> to vector<4xi32>
            %38 = arith.extsi %27 : vector<4xi8> to vector<4xi32>
            %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [1] : vector<1xi8> into vector<4x1xi8>
            %43 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %44 = vector.insert %43, %42 [2] : vector<1xi8> into vector<4x1xi8>
            %45 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %46 = vector.insert %45, %44 [3] : vector<1xi8> into vector<4x1xi8>
            %47 = arith.extsi %46 : vector<4x1xi8> to vector<4x1xi32>
            %48 = vector.extract %47[0, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %cst_1 [0] : i32 into vector<4xi32>
            %50 = vector.extract %47[1, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [1] : i32 into vector<4xi32>
            %52 = vector.extract %47[2, 0] : vector<4x1xi32>
            %53 = vector.insert %52, %51 [2] : i32 into vector<4xi32>
            %54 = vector.extract %47[3, 0] : vector<4x1xi32>
            %55 = vector.insert %54, %53 [3] : i32 into vector<4xi32>
            %56 = vector.extract %arg10[0] : vector<4xi32>
            %57 = arith.muli %35, %55 : vector<4xi32>
            %58 = vector.reduction <add>, %57, %56 : vector<4xi32> into i32
            %59 = vector.insert %58, %cst_0 [0] : i32 into vector<1xi32>
            %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [1] : vector<1xi8> into vector<4x1xi8>
            %64 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %65 = vector.insert %64, %63 [2] : vector<1xi8> into vector<4x1xi8>
            %66 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %67 = vector.insert %66, %65 [3] : vector<1xi8> into vector<4x1xi8>
            %68 = arith.extsi %67 : vector<4x1xi8> to vector<4x1xi32>
            %69 = vector.extract %68[0, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %cst_1 [0] : i32 into vector<4xi32>
            %71 = vector.extract %68[1, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [1] : i32 into vector<4xi32>
            %73 = vector.extract %68[2, 0] : vector<4x1xi32>
            %74 = vector.insert %73, %72 [2] : i32 into vector<4xi32>
            %75 = vector.extract %68[3, 0] : vector<4x1xi32>
            %76 = vector.insert %75, %74 [3] : i32 into vector<4xi32>
            %77 = vector.extract %arg10[1] : vector<4xi32>
            %78 = arith.muli %35, %76 : vector<4xi32>
            %79 = vector.reduction <add>, %78, %77 : vector<4xi32> into i32
            %80 = vector.insert %79, %cst_0 [0] : i32 into vector<1xi32>
            %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [1] : vector<1xi8> into vector<4x1xi8>
            %85 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %86 = vector.insert %85, %84 [2] : vector<1xi8> into vector<4x1xi8>
            %87 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %88 = vector.insert %87, %86 [3] : vector<1xi8> into vector<4x1xi8>
            %89 = arith.extsi %88 : vector<4x1xi8> to vector<4x1xi32>
            %90 = vector.extract %89[0, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %cst_1 [0] : i32 into vector<4xi32>
            %92 = vector.extract %89[1, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [1] : i32 into vector<4xi32>
            %94 = vector.extract %89[2, 0] : vector<4x1xi32>
            %95 = vector.insert %94, %93 [2] : i32 into vector<4xi32>
            %96 = vector.extract %89[3, 0] : vector<4x1xi32>
            %97 = vector.insert %96, %95 [3] : i32 into vector<4xi32>
            %98 = vector.extract %arg10[2] : vector<4xi32>
            %99 = arith.muli %35, %97 : vector<4xi32>
            %100 = vector.reduction <add>, %99, %98 : vector<4xi32> into i32
            %101 = vector.insert %100, %cst_0 [0] : i32 into vector<1xi32>
            %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [1] : vector<1xi8> into vector<4x1xi8>
            %106 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %107 = vector.insert %106, %105 [2] : vector<1xi8> into vector<4x1xi8>
            %108 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %109 = vector.insert %108, %107 [3] : vector<1xi8> into vector<4x1xi8>
            %110 = arith.extsi %109 : vector<4x1xi8> to vector<4x1xi32>
            %111 = vector.extract %110[0, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %cst_1 [0] : i32 into vector<4xi32>
            %113 = vector.extract %110[1, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [1] : i32 into vector<4xi32>
            %115 = vector.extract %110[2, 0] : vector<4x1xi32>
            %116 = vector.insert %115, %114 [2] : i32 into vector<4xi32>
            %117 = vector.extract %110[3, 0] : vector<4x1xi32>
            %118 = vector.insert %117, %116 [3] : i32 into vector<4xi32>
            %119 = vector.extract %arg10[3] : vector<4xi32>
            %120 = arith.muli %35, %118 : vector<4xi32>
            %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
            %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
            %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [1] : vector<1xi8> into vector<4x1xi8>
            %127 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %128 = vector.insert %127, %126 [2] : vector<1xi8> into vector<4x1xi8>
            %129 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %130 = vector.insert %129, %128 [3] : vector<1xi8> into vector<4x1xi8>
            %131 = arith.extsi %130 : vector<4x1xi8> to vector<4x1xi32>
            %132 = vector.extract %131[0, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %cst_1 [0] : i32 into vector<4xi32>
            %134 = vector.extract %131[1, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [1] : i32 into vector<4xi32>
            %136 = vector.extract %131[2, 0] : vector<4x1xi32>
            %137 = vector.insert %136, %135 [2] : i32 into vector<4xi32>
            %138 = vector.extract %131[3, 0] : vector<4x1xi32>
            %139 = vector.insert %138, %137 [3] : i32 into vector<4xi32>
            %140 = vector.extract %arg9[0] : vector<4xi32>
            %141 = arith.muli %36, %139 : vector<4xi32>
            %142 = vector.reduction <add>, %141, %140 : vector<4xi32> into i32
            %143 = vector.insert %142, %cst_0 [0] : i32 into vector<1xi32>
            %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [1] : vector<1xi8> into vector<4x1xi8>
            %148 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %149 = vector.insert %148, %147 [2] : vector<1xi8> into vector<4x1xi8>
            %150 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %151 = vector.insert %150, %149 [3] : vector<1xi8> into vector<4x1xi8>
            %152 = arith.extsi %151 : vector<4x1xi8> to vector<4x1xi32>
            %153 = vector.extract %152[0, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %cst_1 [0] : i32 into vector<4xi32>
            %155 = vector.extract %152[1, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [1] : i32 into vector<4xi32>
            %157 = vector.extract %152[2, 0] : vector<4x1xi32>
            %158 = vector.insert %157, %156 [2] : i32 into vector<4xi32>
            %159 = vector.extract %152[3, 0] : vector<4x1xi32>
            %160 = vector.insert %159, %158 [3] : i32 into vector<4xi32>
            %161 = vector.extract %arg9[1] : vector<4xi32>
            %162 = arith.muli %36, %160 : vector<4xi32>
            %163 = vector.reduction <add>, %162, %161 : vector<4xi32> into i32
            %164 = vector.insert %163, %cst_0 [0] : i32 into vector<1xi32>
            %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [1] : vector<1xi8> into vector<4x1xi8>
            %169 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %170 = vector.insert %169, %168 [2] : vector<1xi8> into vector<4x1xi8>
            %171 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %172 = vector.insert %171, %170 [3] : vector<1xi8> into vector<4x1xi8>
            %173 = arith.extsi %172 : vector<4x1xi8> to vector<4x1xi32>
            %174 = vector.extract %173[0, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %cst_1 [0] : i32 into vector<4xi32>
            %176 = vector.extract %173[1, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [1] : i32 into vector<4xi32>
            %178 = vector.extract %173[2, 0] : vector<4x1xi32>
            %179 = vector.insert %178, %177 [2] : i32 into vector<4xi32>
            %180 = vector.extract %173[3, 0] : vector<4x1xi32>
            %181 = vector.insert %180, %179 [3] : i32 into vector<4xi32>
            %182 = vector.extract %arg9[2] : vector<4xi32>
            %183 = arith.muli %36, %181 : vector<4xi32>
            %184 = vector.reduction <add>, %183, %182 : vector<4xi32> into i32
            %185 = vector.insert %184, %cst_0 [0] : i32 into vector<1xi32>
            %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [1] : vector<1xi8> into vector<4x1xi8>
            %190 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %191 = vector.insert %190, %189 [2] : vector<1xi8> into vector<4x1xi8>
            %192 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %193 = vector.insert %192, %191 [3] : vector<1xi8> into vector<4x1xi8>
            %194 = arith.extsi %193 : vector<4x1xi8> to vector<4x1xi32>
            %195 = vector.extract %194[0, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %cst_1 [0] : i32 into vector<4xi32>
            %197 = vector.extract %194[1, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [1] : i32 into vector<4xi32>
            %199 = vector.extract %194[2, 0] : vector<4x1xi32>
            %200 = vector.insert %199, %198 [2] : i32 into vector<4xi32>
            %201 = vector.extract %194[3, 0] : vector<4x1xi32>
            %202 = vector.insert %201, %200 [3] : i32 into vector<4xi32>
            %203 = vector.extract %arg9[3] : vector<4xi32>
            %204 = arith.muli %36, %202 : vector<4xi32>
            %205 = vector.reduction <add>, %204, %203 : vector<4xi32> into i32
            %206 = vector.insert %205, %cst_0 [0] : i32 into vector<1xi32>
            %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [1] : vector<1xi8> into vector<4x1xi8>
            %211 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %212 = vector.insert %211, %210 [2] : vector<1xi8> into vector<4x1xi8>
            %213 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %214 = vector.insert %213, %212 [3] : vector<1xi8> into vector<4x1xi8>
            %215 = arith.extsi %214 : vector<4x1xi8> to vector<4x1xi32>
            %216 = vector.extract %215[0, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %cst_1 [0] : i32 into vector<4xi32>
            %218 = vector.extract %215[1, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [1] : i32 into vector<4xi32>
            %220 = vector.extract %215[2, 0] : vector<4x1xi32>
            %221 = vector.insert %220, %219 [2] : i32 into vector<4xi32>
            %222 = vector.extract %215[3, 0] : vector<4x1xi32>
            %223 = vector.insert %222, %221 [3] : i32 into vector<4xi32>
            %224 = vector.extract %arg8[0] : vector<4xi32>
            %225 = arith.muli %37, %223 : vector<4xi32>
            %226 = vector.reduction <add>, %225, %224 : vector<4xi32> into i32
            %227 = vector.insert %226, %cst_0 [0] : i32 into vector<1xi32>
            %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [1] : vector<1xi8> into vector<4x1xi8>
            %232 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %233 = vector.insert %232, %231 [2] : vector<1xi8> into vector<4x1xi8>
            %234 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %235 = vector.insert %234, %233 [3] : vector<1xi8> into vector<4x1xi8>
            %236 = arith.extsi %235 : vector<4x1xi8> to vector<4x1xi32>
            %237 = vector.extract %236[0, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %cst_1 [0] : i32 into vector<4xi32>
            %239 = vector.extract %236[1, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [1] : i32 into vector<4xi32>
            %241 = vector.extract %236[2, 0] : vector<4x1xi32>
            %242 = vector.insert %241, %240 [2] : i32 into vector<4xi32>
            %243 = vector.extract %236[3, 0] : vector<4x1xi32>
            %244 = vector.insert %243, %242 [3] : i32 into vector<4xi32>
            %245 = vector.extract %arg8[1] : vector<4xi32>
            %246 = arith.muli %37, %244 : vector<4xi32>
            %247 = vector.reduction <add>, %246, %245 : vector<4xi32> into i32
            %248 = vector.insert %247, %cst_0 [0] : i32 into vector<1xi32>
            %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [1] : vector<1xi8> into vector<4x1xi8>
            %253 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %254 = vector.insert %253, %252 [2] : vector<1xi8> into vector<4x1xi8>
            %255 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %256 = vector.insert %255, %254 [3] : vector<1xi8> into vector<4x1xi8>
            %257 = arith.extsi %256 : vector<4x1xi8> to vector<4x1xi32>
            %258 = vector.extract %257[0, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %cst_1 [0] : i32 into vector<4xi32>
            %260 = vector.extract %257[1, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [1] : i32 into vector<4xi32>
            %262 = vector.extract %257[2, 0] : vector<4x1xi32>
            %263 = vector.insert %262, %261 [2] : i32 into vector<4xi32>
            %264 = vector.extract %257[3, 0] : vector<4x1xi32>
            %265 = vector.insert %264, %263 [3] : i32 into vector<4xi32>
            %266 = vector.extract %arg8[2] : vector<4xi32>
            %267 = arith.muli %37, %265 : vector<4xi32>
            %268 = vector.reduction <add>, %267, %266 : vector<4xi32> into i32
            %269 = vector.insert %268, %cst_0 [0] : i32 into vector<1xi32>
            %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [1] : vector<1xi8> into vector<4x1xi8>
            %274 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %275 = vector.insert %274, %273 [2] : vector<1xi8> into vector<4x1xi8>
            %276 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %277 = vector.insert %276, %275 [3] : vector<1xi8> into vector<4x1xi8>
            %278 = arith.extsi %277 : vector<4x1xi8> to vector<4x1xi32>
            %279 = vector.extract %278[0, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %cst_1 [0] : i32 into vector<4xi32>
            %281 = vector.extract %278[1, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [1] : i32 into vector<4xi32>
            %283 = vector.extract %278[2, 0] : vector<4x1xi32>
            %284 = vector.insert %283, %282 [2] : i32 into vector<4xi32>
            %285 = vector.extract %278[3, 0] : vector<4x1xi32>
            %286 = vector.insert %285, %284 [3] : i32 into vector<4xi32>
            %287 = vector.extract %arg8[3] : vector<4xi32>
            %288 = arith.muli %37, %286 : vector<4xi32>
            %289 = vector.reduction <add>, %288, %287 : vector<4xi32> into i32
            %290 = vector.insert %289, %cst_0 [0] : i32 into vector<1xi32>
            %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [1] : vector<1xi8> into vector<4x1xi8>
            %295 = vector.extract_strided_slice %32 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %296 = vector.insert %295, %294 [2] : vector<1xi8> into vector<4x1xi8>
            %297 = vector.extract_strided_slice %34 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %298 = vector.insert %297, %296 [3] : vector<1xi8> into vector<4x1xi8>
            %299 = arith.extsi %298 : vector<4x1xi8> to vector<4x1xi32>
            %300 = vector.extract %299[0, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %cst_1 [0] : i32 into vector<4xi32>
            %302 = vector.extract %299[1, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [1] : i32 into vector<4xi32>
            %304 = vector.extract %299[2, 0] : vector<4x1xi32>
            %305 = vector.insert %304, %303 [2] : i32 into vector<4xi32>
            %306 = vector.extract %299[3, 0] : vector<4x1xi32>
            %307 = vector.insert %306, %305 [3] : i32 into vector<4xi32>
            %308 = vector.extract %arg7[0] : vector<4xi32>
            %309 = arith.muli %38, %307 : vector<4xi32>
            %310 = vector.reduction <add>, %309, %308 : vector<4xi32> into i32
            %311 = vector.insert %310, %cst_0 [0] : i32 into vector<1xi32>
            %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [1] : vector<1xi8> into vector<4x1xi8>
            %316 = vector.extract_strided_slice %32 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %317 = vector.insert %316, %315 [2] : vector<1xi8> into vector<4x1xi8>
            %318 = vector.extract_strided_slice %34 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %319 = vector.insert %318, %317 [3] : vector<1xi8> into vector<4x1xi8>
            %320 = arith.extsi %319 : vector<4x1xi8> to vector<4x1xi32>
            %321 = vector.extract %320[0, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %cst_1 [0] : i32 into vector<4xi32>
            %323 = vector.extract %320[1, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [1] : i32 into vector<4xi32>
            %325 = vector.extract %320[2, 0] : vector<4x1xi32>
            %326 = vector.insert %325, %324 [2] : i32 into vector<4xi32>
            %327 = vector.extract %320[3, 0] : vector<4x1xi32>
            %328 = vector.insert %327, %326 [3] : i32 into vector<4xi32>
            %329 = vector.extract %arg7[1] : vector<4xi32>
            %330 = arith.muli %38, %328 : vector<4xi32>
            %331 = vector.reduction <add>, %330, %329 : vector<4xi32> into i32
            %332 = vector.insert %331, %cst_0 [0] : i32 into vector<1xi32>
            %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [1] : vector<1xi8> into vector<4x1xi8>
            %337 = vector.extract_strided_slice %32 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %338 = vector.insert %337, %336 [2] : vector<1xi8> into vector<4x1xi8>
            %339 = vector.extract_strided_slice %34 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %340 = vector.insert %339, %338 [3] : vector<1xi8> into vector<4x1xi8>
            %341 = arith.extsi %340 : vector<4x1xi8> to vector<4x1xi32>
            %342 = vector.extract %341[0, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %cst_1 [0] : i32 into vector<4xi32>
            %344 = vector.extract %341[1, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [1] : i32 into vector<4xi32>
            %346 = vector.extract %341[2, 0] : vector<4x1xi32>
            %347 = vector.insert %346, %345 [2] : i32 into vector<4xi32>
            %348 = vector.extract %341[3, 0] : vector<4x1xi32>
            %349 = vector.insert %348, %347 [3] : i32 into vector<4xi32>
            %350 = vector.extract %arg7[2] : vector<4xi32>
            %351 = arith.muli %38, %349 : vector<4xi32>
            %352 = vector.reduction <add>, %351, %350 : vector<4xi32> into i32
            %353 = vector.insert %352, %cst_0 [0] : i32 into vector<1xi32>
            %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [1] : vector<1xi8> into vector<4x1xi8>
            %358 = vector.extract_strided_slice %32 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %359 = vector.insert %358, %357 [2] : vector<1xi8> into vector<4x1xi8>
            %360 = vector.extract_strided_slice %34 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %361 = vector.insert %360, %359 [3] : vector<1xi8> into vector<4x1xi8>
            %362 = arith.extsi %361 : vector<4x1xi8> to vector<4x1xi32>
            %363 = vector.extract %362[0, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %cst_1 [0] : i32 into vector<4xi32>
            %365 = vector.extract %362[1, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [1] : i32 into vector<4xi32>
            %367 = vector.extract %362[2, 0] : vector<4x1xi32>
            %368 = vector.insert %367, %366 [2] : i32 into vector<4xi32>
            %369 = vector.extract %362[3, 0] : vector<4x1xi32>
            %370 = vector.insert %369, %368 [3] : i32 into vector<4xi32>
            %371 = vector.extract %arg7[3] : vector<4xi32>
            %372 = arith.muli %38, %370 : vector<4xi32>
            %373 = vector.reduction <add>, %372, %371 : vector<4xi32> into i32
            %374 = vector.insert %373, %cst_0 [0] : i32 into vector<1xi32>
            %375 = vector.insert_strided_slice %59, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %80, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %101, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %122, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %143, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %164, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %185, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %206, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %227, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %248, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %269, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %290, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %387 = vector.insert_strided_slice %311, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %388 = vector.insert_strided_slice %332, %387 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %389 = vector.insert_strided_slice %353, %388 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %390 = vector.insert_strided_slice %374, %389 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            scf.yield %390, %386, %382, %378 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
          }
          %20 = vector.transfer_write %19#3, %18[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %21 = vector.transfer_write %19#2, %20[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %22 = vector.transfer_write %19#1, %21[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %23 = vector.transfer_write %19#0, %22[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %inserted_slice = tensor.insert_slice %23 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %14 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After SPIRVVectorize (iree-spirv-vectorize) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
  %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
    scf.for %arg1 = %5 to %c1024 step %6 {
      %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
        %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
        %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
        %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
        %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
            %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
            %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
            %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
            %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
            %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
            %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
            %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
            %44 = vector.extract %43[0, 0] : vector<4x1xi32>
            %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
            %46 = vector.extract %43[1, 0] : vector<4x1xi32>
            %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
            %48 = vector.extract %43[2, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
            %50 = vector.extract %43[3, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
            %52 = vector.extract %arg10[0] : vector<4xi32>
            %53 = arith.muli %31, %51 : vector<4xi32>
            %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
            %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
            %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
            %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
            %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
            %65 = vector.extract %64[0, 0] : vector<4x1xi32>
            %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
            %67 = vector.extract %64[1, 0] : vector<4x1xi32>
            %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
            %69 = vector.extract %64[2, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
            %71 = vector.extract %64[3, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
            %73 = vector.extract %arg10[1] : vector<4xi32>
            %74 = arith.muli %31, %72 : vector<4xi32>
            %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
            %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
            %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
            %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
            %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
            %86 = vector.extract %85[0, 0] : vector<4x1xi32>
            %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
            %88 = vector.extract %85[1, 0] : vector<4x1xi32>
            %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
            %90 = vector.extract %85[2, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
            %92 = vector.extract %85[3, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
            %94 = vector.extract %arg10[2] : vector<4xi32>
            %95 = arith.muli %31, %93 : vector<4xi32>
            %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
            %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
            %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
            %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
            %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
            %107 = vector.extract %106[0, 0] : vector<4x1xi32>
            %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
            %109 = vector.extract %106[1, 0] : vector<4x1xi32>
            %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
            %111 = vector.extract %106[2, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
            %113 = vector.extract %106[3, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
            %115 = vector.extract %arg10[3] : vector<4xi32>
            %116 = arith.muli %31, %114 : vector<4xi32>
            %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
            %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
            %119 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %120 = vector.insert %119, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %121 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %122 = vector.insert %121, %120 [1] : vector<1xi8> into vector<4x1xi8>
            %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %122 [2] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [3] : vector<1xi8> into vector<4x1xi8>
            %127 = arith.extsi %126 : vector<4x1xi8> to vector<4x1xi32>
            %128 = vector.extract %127[0, 0] : vector<4x1xi32>
            %129 = vector.insert %128, %cst_1 [0] : i32 into vector<4xi32>
            %130 = vector.extract %127[1, 0] : vector<4x1xi32>
            %131 = vector.insert %130, %129 [1] : i32 into vector<4xi32>
            %132 = vector.extract %127[2, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %131 [2] : i32 into vector<4xi32>
            %134 = vector.extract %127[3, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [3] : i32 into vector<4xi32>
            %136 = vector.extract %arg9[0] : vector<4xi32>
            %137 = arith.muli %32, %135 : vector<4xi32>
            %138 = vector.reduction <add>, %137, %136 : vector<4xi32> into i32
            %139 = vector.insert %138, %cst_0 [0] : i32 into vector<1xi32>
            %140 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %141 = vector.insert %140, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %142 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %143 = vector.insert %142, %141 [1] : vector<1xi8> into vector<4x1xi8>
            %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %143 [2] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [3] : vector<1xi8> into vector<4x1xi8>
            %148 = arith.extsi %147 : vector<4x1xi8> to vector<4x1xi32>
            %149 = vector.extract %148[0, 0] : vector<4x1xi32>
            %150 = vector.insert %149, %cst_1 [0] : i32 into vector<4xi32>
            %151 = vector.extract %148[1, 0] : vector<4x1xi32>
            %152 = vector.insert %151, %150 [1] : i32 into vector<4xi32>
            %153 = vector.extract %148[2, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %152 [2] : i32 into vector<4xi32>
            %155 = vector.extract %148[3, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [3] : i32 into vector<4xi32>
            %157 = vector.extract %arg9[1] : vector<4xi32>
            %158 = arith.muli %32, %156 : vector<4xi32>
            %159 = vector.reduction <add>, %158, %157 : vector<4xi32> into i32
            %160 = vector.insert %159, %cst_0 [0] : i32 into vector<1xi32>
            %161 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %162 = vector.insert %161, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %163 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %164 = vector.insert %163, %162 [1] : vector<1xi8> into vector<4x1xi8>
            %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %164 [2] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [3] : vector<1xi8> into vector<4x1xi8>
            %169 = arith.extsi %168 : vector<4x1xi8> to vector<4x1xi32>
            %170 = vector.extract %169[0, 0] : vector<4x1xi32>
            %171 = vector.insert %170, %cst_1 [0] : i32 into vector<4xi32>
            %172 = vector.extract %169[1, 0] : vector<4x1xi32>
            %173 = vector.insert %172, %171 [1] : i32 into vector<4xi32>
            %174 = vector.extract %169[2, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %173 [2] : i32 into vector<4xi32>
            %176 = vector.extract %169[3, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [3] : i32 into vector<4xi32>
            %178 = vector.extract %arg9[2] : vector<4xi32>
            %179 = arith.muli %32, %177 : vector<4xi32>
            %180 = vector.reduction <add>, %179, %178 : vector<4xi32> into i32
            %181 = vector.insert %180, %cst_0 [0] : i32 into vector<1xi32>
            %182 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %183 = vector.insert %182, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %184 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %185 = vector.insert %184, %183 [1] : vector<1xi8> into vector<4x1xi8>
            %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %185 [2] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [3] : vector<1xi8> into vector<4x1xi8>
            %190 = arith.extsi %189 : vector<4x1xi8> to vector<4x1xi32>
            %191 = vector.extract %190[0, 0] : vector<4x1xi32>
            %192 = vector.insert %191, %cst_1 [0] : i32 into vector<4xi32>
            %193 = vector.extract %190[1, 0] : vector<4x1xi32>
            %194 = vector.insert %193, %192 [1] : i32 into vector<4xi32>
            %195 = vector.extract %190[2, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %194 [2] : i32 into vector<4xi32>
            %197 = vector.extract %190[3, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [3] : i32 into vector<4xi32>
            %199 = vector.extract %arg9[3] : vector<4xi32>
            %200 = arith.muli %32, %198 : vector<4xi32>
            %201 = vector.reduction <add>, %200, %199 : vector<4xi32> into i32
            %202 = vector.insert %201, %cst_0 [0] : i32 into vector<1xi32>
            %203 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %204 = vector.insert %203, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %205 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %206 = vector.insert %205, %204 [1] : vector<1xi8> into vector<4x1xi8>
            %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %206 [2] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [3] : vector<1xi8> into vector<4x1xi8>
            %211 = arith.extsi %210 : vector<4x1xi8> to vector<4x1xi32>
            %212 = vector.extract %211[0, 0] : vector<4x1xi32>
            %213 = vector.insert %212, %cst_1 [0] : i32 into vector<4xi32>
            %214 = vector.extract %211[1, 0] : vector<4x1xi32>
            %215 = vector.insert %214, %213 [1] : i32 into vector<4xi32>
            %216 = vector.extract %211[2, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %215 [2] : i32 into vector<4xi32>
            %218 = vector.extract %211[3, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [3] : i32 into vector<4xi32>
            %220 = vector.extract %arg8[0] : vector<4xi32>
            %221 = arith.muli %33, %219 : vector<4xi32>
            %222 = vector.reduction <add>, %221, %220 : vector<4xi32> into i32
            %223 = vector.insert %222, %cst_0 [0] : i32 into vector<1xi32>
            %224 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %225 = vector.insert %224, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %226 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %227 = vector.insert %226, %225 [1] : vector<1xi8> into vector<4x1xi8>
            %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %227 [2] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [3] : vector<1xi8> into vector<4x1xi8>
            %232 = arith.extsi %231 : vector<4x1xi8> to vector<4x1xi32>
            %233 = vector.extract %232[0, 0] : vector<4x1xi32>
            %234 = vector.insert %233, %cst_1 [0] : i32 into vector<4xi32>
            %235 = vector.extract %232[1, 0] : vector<4x1xi32>
            %236 = vector.insert %235, %234 [1] : i32 into vector<4xi32>
            %237 = vector.extract %232[2, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %236 [2] : i32 into vector<4xi32>
            %239 = vector.extract %232[3, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [3] : i32 into vector<4xi32>
            %241 = vector.extract %arg8[1] : vector<4xi32>
            %242 = arith.muli %33, %240 : vector<4xi32>
            %243 = vector.reduction <add>, %242, %241 : vector<4xi32> into i32
            %244 = vector.insert %243, %cst_0 [0] : i32 into vector<1xi32>
            %245 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %246 = vector.insert %245, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %247 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %248 = vector.insert %247, %246 [1] : vector<1xi8> into vector<4x1xi8>
            %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %248 [2] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [3] : vector<1xi8> into vector<4x1xi8>
            %253 = arith.extsi %252 : vector<4x1xi8> to vector<4x1xi32>
            %254 = vector.extract %253[0, 0] : vector<4x1xi32>
            %255 = vector.insert %254, %cst_1 [0] : i32 into vector<4xi32>
            %256 = vector.extract %253[1, 0] : vector<4x1xi32>
            %257 = vector.insert %256, %255 [1] : i32 into vector<4xi32>
            %258 = vector.extract %253[2, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %257 [2] : i32 into vector<4xi32>
            %260 = vector.extract %253[3, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [3] : i32 into vector<4xi32>
            %262 = vector.extract %arg8[2] : vector<4xi32>
            %263 = arith.muli %33, %261 : vector<4xi32>
            %264 = vector.reduction <add>, %263, %262 : vector<4xi32> into i32
            %265 = vector.insert %264, %cst_0 [0] : i32 into vector<1xi32>
            %266 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %267 = vector.insert %266, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %268 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %269 = vector.insert %268, %267 [1] : vector<1xi8> into vector<4x1xi8>
            %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %269 [2] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [3] : vector<1xi8> into vector<4x1xi8>
            %274 = arith.extsi %273 : vector<4x1xi8> to vector<4x1xi32>
            %275 = vector.extract %274[0, 0] : vector<4x1xi32>
            %276 = vector.insert %275, %cst_1 [0] : i32 into vector<4xi32>
            %277 = vector.extract %274[1, 0] : vector<4x1xi32>
            %278 = vector.insert %277, %276 [1] : i32 into vector<4xi32>
            %279 = vector.extract %274[2, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %278 [2] : i32 into vector<4xi32>
            %281 = vector.extract %274[3, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [3] : i32 into vector<4xi32>
            %283 = vector.extract %arg8[3] : vector<4xi32>
            %284 = arith.muli %33, %282 : vector<4xi32>
            %285 = vector.reduction <add>, %284, %283 : vector<4xi32> into i32
            %286 = vector.insert %285, %cst_0 [0] : i32 into vector<1xi32>
            %287 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %288 = vector.insert %287, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %289 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %290 = vector.insert %289, %288 [1] : vector<1xi8> into vector<4x1xi8>
            %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %290 [2] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [3] : vector<1xi8> into vector<4x1xi8>
            %295 = arith.extsi %294 : vector<4x1xi8> to vector<4x1xi32>
            %296 = vector.extract %295[0, 0] : vector<4x1xi32>
            %297 = vector.insert %296, %cst_1 [0] : i32 into vector<4xi32>
            %298 = vector.extract %295[1, 0] : vector<4x1xi32>
            %299 = vector.insert %298, %297 [1] : i32 into vector<4xi32>
            %300 = vector.extract %295[2, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %299 [2] : i32 into vector<4xi32>
            %302 = vector.extract %295[3, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [3] : i32 into vector<4xi32>
            %304 = vector.extract %arg7[0] : vector<4xi32>
            %305 = arith.muli %34, %303 : vector<4xi32>
            %306 = vector.reduction <add>, %305, %304 : vector<4xi32> into i32
            %307 = vector.insert %306, %cst_0 [0] : i32 into vector<1xi32>
            %308 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %309 = vector.insert %308, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %310 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %311 = vector.insert %310, %309 [1] : vector<1xi8> into vector<4x1xi8>
            %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %311 [2] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [3] : vector<1xi8> into vector<4x1xi8>
            %316 = arith.extsi %315 : vector<4x1xi8> to vector<4x1xi32>
            %317 = vector.extract %316[0, 0] : vector<4x1xi32>
            %318 = vector.insert %317, %cst_1 [0] : i32 into vector<4xi32>
            %319 = vector.extract %316[1, 0] : vector<4x1xi32>
            %320 = vector.insert %319, %318 [1] : i32 into vector<4xi32>
            %321 = vector.extract %316[2, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %320 [2] : i32 into vector<4xi32>
            %323 = vector.extract %316[3, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [3] : i32 into vector<4xi32>
            %325 = vector.extract %arg7[1] : vector<4xi32>
            %326 = arith.muli %34, %324 : vector<4xi32>
            %327 = vector.reduction <add>, %326, %325 : vector<4xi32> into i32
            %328 = vector.insert %327, %cst_0 [0] : i32 into vector<1xi32>
            %329 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %330 = vector.insert %329, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %331 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %332 = vector.insert %331, %330 [1] : vector<1xi8> into vector<4x1xi8>
            %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %332 [2] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [3] : vector<1xi8> into vector<4x1xi8>
            %337 = arith.extsi %336 : vector<4x1xi8> to vector<4x1xi32>
            %338 = vector.extract %337[0, 0] : vector<4x1xi32>
            %339 = vector.insert %338, %cst_1 [0] : i32 into vector<4xi32>
            %340 = vector.extract %337[1, 0] : vector<4x1xi32>
            %341 = vector.insert %340, %339 [1] : i32 into vector<4xi32>
            %342 = vector.extract %337[2, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %341 [2] : i32 into vector<4xi32>
            %344 = vector.extract %337[3, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [3] : i32 into vector<4xi32>
            %346 = vector.extract %arg7[2] : vector<4xi32>
            %347 = arith.muli %34, %345 : vector<4xi32>
            %348 = vector.reduction <add>, %347, %346 : vector<4xi32> into i32
            %349 = vector.insert %348, %cst_0 [0] : i32 into vector<1xi32>
            %350 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %351 = vector.insert %350, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %352 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %353 = vector.insert %352, %351 [1] : vector<1xi8> into vector<4x1xi8>
            %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %353 [2] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [3] : vector<1xi8> into vector<4x1xi8>
            %358 = arith.extsi %357 : vector<4x1xi8> to vector<4x1xi32>
            %359 = vector.extract %358[0, 0] : vector<4x1xi32>
            %360 = vector.insert %359, %cst_1 [0] : i32 into vector<4xi32>
            %361 = vector.extract %358[1, 0] : vector<4x1xi32>
            %362 = vector.insert %361, %360 [1] : i32 into vector<4xi32>
            %363 = vector.extract %358[2, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %362 [2] : i32 into vector<4xi32>
            %365 = vector.extract %358[3, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [3] : i32 into vector<4xi32>
            %367 = vector.extract %arg7[3] : vector<4xi32>
            %368 = arith.muli %34, %366 : vector<4xi32>
            %369 = vector.reduction <add>, %368, %367 : vector<4xi32> into i32
            %370 = vector.insert %369, %cst_0 [0] : i32 into vector<1xi32>
            %371 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %372 = vector.insert_strided_slice %76, %371 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %373 = vector.insert_strided_slice %97, %372 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %374 = vector.insert_strided_slice %118, %373 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %375 = vector.insert_strided_slice %139, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %160, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %181, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %202, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %223, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %244, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %265, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %286, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %307, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %328, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %349, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %370, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            scf.yield %386, %382, %378, %374 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
          }
          %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %14 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After ForOpCanonicalization (iree-codegen-canonicalize-scf-for) //----- //
func.func @_main_dispatch_0_matmul_1024x1024x1024() {
  %cst = arith.constant dense<0> : vector<4x1xi8>
  %cst_0 = arith.constant dense<0> : vector<1xi32>
  %cst_1 = arith.constant dense<0> : vector<4xi32>
  %c1 = arith.constant 1 : index
  %c2 = arith.constant 2 : index
  %c3 = arith.constant 3 : index
  %c0_i8 = arith.constant 0 : i8
  %c32 = arith.constant 32 : index
  %c8 = arith.constant 8 : index
  %c4 = arith.constant 4 : index
  %c1024 = arith.constant 1024 : index
  %c0 = arith.constant 0 : index
  %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
  %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
  %workgroup_id_x = hal.interface.workgroup.id[0] : index
  %workgroup_count_x = hal.interface.workgroup.count[0] : index
  %workgroup_id_y = hal.interface.workgroup.id[1] : index
  %workgroup_count_y = hal.interface.workgroup.count[1] : index
  %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
  %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
  %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
  %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
  scf.for %arg0 = %3 to %c1024 step %4 {
    %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
    scf.for %arg1 = %5 to %c1024 step %6 {
      %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
      %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
      %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
        %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
        %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
        %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
        %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
          %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
          %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
            %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
            %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
            %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
            %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
            %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
            %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
            %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
            %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
            %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
            %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
            %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
            %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
            %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
            %44 = vector.extract %43[0, 0] : vector<4x1xi32>
            %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
            %46 = vector.extract %43[1, 0] : vector<4x1xi32>
            %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
            %48 = vector.extract %43[2, 0] : vector<4x1xi32>
            %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
            %50 = vector.extract %43[3, 0] : vector<4x1xi32>
            %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
            %52 = vector.extract %arg10[0] : vector<4xi32>
            %53 = arith.muli %31, %51 : vector<4xi32>
            %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
            %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
            %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
            %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
            %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
            %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
            %65 = vector.extract %64[0, 0] : vector<4x1xi32>
            %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
            %67 = vector.extract %64[1, 0] : vector<4x1xi32>
            %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
            %69 = vector.extract %64[2, 0] : vector<4x1xi32>
            %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
            %71 = vector.extract %64[3, 0] : vector<4x1xi32>
            %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
            %73 = vector.extract %arg10[1] : vector<4xi32>
            %74 = arith.muli %31, %72 : vector<4xi32>
            %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
            %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
            %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
            %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
            %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
            %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
            %86 = vector.extract %85[0, 0] : vector<4x1xi32>
            %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
            %88 = vector.extract %85[1, 0] : vector<4x1xi32>
            %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
            %90 = vector.extract %85[2, 0] : vector<4x1xi32>
            %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
            %92 = vector.extract %85[3, 0] : vector<4x1xi32>
            %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
            %94 = vector.extract %arg10[2] : vector<4xi32>
            %95 = arith.muli %31, %93 : vector<4xi32>
            %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
            %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
            %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
            %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
            %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
            %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
            %107 = vector.extract %106[0, 0] : vector<4x1xi32>
            %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
            %109 = vector.extract %106[1, 0] : vector<4x1xi32>
            %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
            %111 = vector.extract %106[2, 0] : vector<4x1xi32>
            %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
            %113 = vector.extract %106[3, 0] : vector<4x1xi32>
            %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
            %115 = vector.extract %arg10[3] : vector<4xi32>
            %116 = arith.muli %31, %114 : vector<4xi32>
            %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
            %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
            %119 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %120 = vector.insert %119, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %121 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %122 = vector.insert %121, %120 [1] : vector<1xi8> into vector<4x1xi8>
            %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %124 = vector.insert %123, %122 [2] : vector<1xi8> into vector<4x1xi8>
            %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %126 = vector.insert %125, %124 [3] : vector<1xi8> into vector<4x1xi8>
            %127 = arith.extsi %126 : vector<4x1xi8> to vector<4x1xi32>
            %128 = vector.extract %127[0, 0] : vector<4x1xi32>
            %129 = vector.insert %128, %cst_1 [0] : i32 into vector<4xi32>
            %130 = vector.extract %127[1, 0] : vector<4x1xi32>
            %131 = vector.insert %130, %129 [1] : i32 into vector<4xi32>
            %132 = vector.extract %127[2, 0] : vector<4x1xi32>
            %133 = vector.insert %132, %131 [2] : i32 into vector<4xi32>
            %134 = vector.extract %127[3, 0] : vector<4x1xi32>
            %135 = vector.insert %134, %133 [3] : i32 into vector<4xi32>
            %136 = vector.extract %arg9[0] : vector<4xi32>
            %137 = arith.muli %32, %135 : vector<4xi32>
            %138 = vector.reduction <add>, %137, %136 : vector<4xi32> into i32
            %139 = vector.insert %138, %cst_0 [0] : i32 into vector<1xi32>
            %140 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %141 = vector.insert %140, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %142 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %143 = vector.insert %142, %141 [1] : vector<1xi8> into vector<4x1xi8>
            %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %145 = vector.insert %144, %143 [2] : vector<1xi8> into vector<4x1xi8>
            %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %147 = vector.insert %146, %145 [3] : vector<1xi8> into vector<4x1xi8>
            %148 = arith.extsi %147 : vector<4x1xi8> to vector<4x1xi32>
            %149 = vector.extract %148[0, 0] : vector<4x1xi32>
            %150 = vector.insert %149, %cst_1 [0] : i32 into vector<4xi32>
            %151 = vector.extract %148[1, 0] : vector<4x1xi32>
            %152 = vector.insert %151, %150 [1] : i32 into vector<4xi32>
            %153 = vector.extract %148[2, 0] : vector<4x1xi32>
            %154 = vector.insert %153, %152 [2] : i32 into vector<4xi32>
            %155 = vector.extract %148[3, 0] : vector<4x1xi32>
            %156 = vector.insert %155, %154 [3] : i32 into vector<4xi32>
            %157 = vector.extract %arg9[1] : vector<4xi32>
            %158 = arith.muli %32, %156 : vector<4xi32>
            %159 = vector.reduction <add>, %158, %157 : vector<4xi32> into i32
            %160 = vector.insert %159, %cst_0 [0] : i32 into vector<1xi32>
            %161 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %162 = vector.insert %161, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %163 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %164 = vector.insert %163, %162 [1] : vector<1xi8> into vector<4x1xi8>
            %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %166 = vector.insert %165, %164 [2] : vector<1xi8> into vector<4x1xi8>
            %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %168 = vector.insert %167, %166 [3] : vector<1xi8> into vector<4x1xi8>
            %169 = arith.extsi %168 : vector<4x1xi8> to vector<4x1xi32>
            %170 = vector.extract %169[0, 0] : vector<4x1xi32>
            %171 = vector.insert %170, %cst_1 [0] : i32 into vector<4xi32>
            %172 = vector.extract %169[1, 0] : vector<4x1xi32>
            %173 = vector.insert %172, %171 [1] : i32 into vector<4xi32>
            %174 = vector.extract %169[2, 0] : vector<4x1xi32>
            %175 = vector.insert %174, %173 [2] : i32 into vector<4xi32>
            %176 = vector.extract %169[3, 0] : vector<4x1xi32>
            %177 = vector.insert %176, %175 [3] : i32 into vector<4xi32>
            %178 = vector.extract %arg9[2] : vector<4xi32>
            %179 = arith.muli %32, %177 : vector<4xi32>
            %180 = vector.reduction <add>, %179, %178 : vector<4xi32> into i32
            %181 = vector.insert %180, %cst_0 [0] : i32 into vector<1xi32>
            %182 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %183 = vector.insert %182, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %184 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %185 = vector.insert %184, %183 [1] : vector<1xi8> into vector<4x1xi8>
            %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %187 = vector.insert %186, %185 [2] : vector<1xi8> into vector<4x1xi8>
            %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %189 = vector.insert %188, %187 [3] : vector<1xi8> into vector<4x1xi8>
            %190 = arith.extsi %189 : vector<4x1xi8> to vector<4x1xi32>
            %191 = vector.extract %190[0, 0] : vector<4x1xi32>
            %192 = vector.insert %191, %cst_1 [0] : i32 into vector<4xi32>
            %193 = vector.extract %190[1, 0] : vector<4x1xi32>
            %194 = vector.insert %193, %192 [1] : i32 into vector<4xi32>
            %195 = vector.extract %190[2, 0] : vector<4x1xi32>
            %196 = vector.insert %195, %194 [2] : i32 into vector<4xi32>
            %197 = vector.extract %190[3, 0] : vector<4x1xi32>
            %198 = vector.insert %197, %196 [3] : i32 into vector<4xi32>
            %199 = vector.extract %arg9[3] : vector<4xi32>
            %200 = arith.muli %32, %198 : vector<4xi32>
            %201 = vector.reduction <add>, %200, %199 : vector<4xi32> into i32
            %202 = vector.insert %201, %cst_0 [0] : i32 into vector<1xi32>
            %203 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %204 = vector.insert %203, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %205 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %206 = vector.insert %205, %204 [1] : vector<1xi8> into vector<4x1xi8>
            %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %208 = vector.insert %207, %206 [2] : vector<1xi8> into vector<4x1xi8>
            %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %210 = vector.insert %209, %208 [3] : vector<1xi8> into vector<4x1xi8>
            %211 = arith.extsi %210 : vector<4x1xi8> to vector<4x1xi32>
            %212 = vector.extract %211[0, 0] : vector<4x1xi32>
            %213 = vector.insert %212, %cst_1 [0] : i32 into vector<4xi32>
            %214 = vector.extract %211[1, 0] : vector<4x1xi32>
            %215 = vector.insert %214, %213 [1] : i32 into vector<4xi32>
            %216 = vector.extract %211[2, 0] : vector<4x1xi32>
            %217 = vector.insert %216, %215 [2] : i32 into vector<4xi32>
            %218 = vector.extract %211[3, 0] : vector<4x1xi32>
            %219 = vector.insert %218, %217 [3] : i32 into vector<4xi32>
            %220 = vector.extract %arg8[0] : vector<4xi32>
            %221 = arith.muli %33, %219 : vector<4xi32>
            %222 = vector.reduction <add>, %221, %220 : vector<4xi32> into i32
            %223 = vector.insert %222, %cst_0 [0] : i32 into vector<1xi32>
            %224 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %225 = vector.insert %224, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %226 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %227 = vector.insert %226, %225 [1] : vector<1xi8> into vector<4x1xi8>
            %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %229 = vector.insert %228, %227 [2] : vector<1xi8> into vector<4x1xi8>
            %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %231 = vector.insert %230, %229 [3] : vector<1xi8> into vector<4x1xi8>
            %232 = arith.extsi %231 : vector<4x1xi8> to vector<4x1xi32>
            %233 = vector.extract %232[0, 0] : vector<4x1xi32>
            %234 = vector.insert %233, %cst_1 [0] : i32 into vector<4xi32>
            %235 = vector.extract %232[1, 0] : vector<4x1xi32>
            %236 = vector.insert %235, %234 [1] : i32 into vector<4xi32>
            %237 = vector.extract %232[2, 0] : vector<4x1xi32>
            %238 = vector.insert %237, %236 [2] : i32 into vector<4xi32>
            %239 = vector.extract %232[3, 0] : vector<4x1xi32>
            %240 = vector.insert %239, %238 [3] : i32 into vector<4xi32>
            %241 = vector.extract %arg8[1] : vector<4xi32>
            %242 = arith.muli %33, %240 : vector<4xi32>
            %243 = vector.reduction <add>, %242, %241 : vector<4xi32> into i32
            %244 = vector.insert %243, %cst_0 [0] : i32 into vector<1xi32>
            %245 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %246 = vector.insert %245, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %247 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %248 = vector.insert %247, %246 [1] : vector<1xi8> into vector<4x1xi8>
            %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %250 = vector.insert %249, %248 [2] : vector<1xi8> into vector<4x1xi8>
            %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %252 = vector.insert %251, %250 [3] : vector<1xi8> into vector<4x1xi8>
            %253 = arith.extsi %252 : vector<4x1xi8> to vector<4x1xi32>
            %254 = vector.extract %253[0, 0] : vector<4x1xi32>
            %255 = vector.insert %254, %cst_1 [0] : i32 into vector<4xi32>
            %256 = vector.extract %253[1, 0] : vector<4x1xi32>
            %257 = vector.insert %256, %255 [1] : i32 into vector<4xi32>
            %258 = vector.extract %253[2, 0] : vector<4x1xi32>
            %259 = vector.insert %258, %257 [2] : i32 into vector<4xi32>
            %260 = vector.extract %253[3, 0] : vector<4x1xi32>
            %261 = vector.insert %260, %259 [3] : i32 into vector<4xi32>
            %262 = vector.extract %arg8[2] : vector<4xi32>
            %263 = arith.muli %33, %261 : vector<4xi32>
            %264 = vector.reduction <add>, %263, %262 : vector<4xi32> into i32
            %265 = vector.insert %264, %cst_0 [0] : i32 into vector<1xi32>
            %266 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %267 = vector.insert %266, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %268 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %269 = vector.insert %268, %267 [1] : vector<1xi8> into vector<4x1xi8>
            %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %271 = vector.insert %270, %269 [2] : vector<1xi8> into vector<4x1xi8>
            %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %273 = vector.insert %272, %271 [3] : vector<1xi8> into vector<4x1xi8>
            %274 = arith.extsi %273 : vector<4x1xi8> to vector<4x1xi32>
            %275 = vector.extract %274[0, 0] : vector<4x1xi32>
            %276 = vector.insert %275, %cst_1 [0] : i32 into vector<4xi32>
            %277 = vector.extract %274[1, 0] : vector<4x1xi32>
            %278 = vector.insert %277, %276 [1] : i32 into vector<4xi32>
            %279 = vector.extract %274[2, 0] : vector<4x1xi32>
            %280 = vector.insert %279, %278 [2] : i32 into vector<4xi32>
            %281 = vector.extract %274[3, 0] : vector<4x1xi32>
            %282 = vector.insert %281, %280 [3] : i32 into vector<4xi32>
            %283 = vector.extract %arg8[3] : vector<4xi32>
            %284 = arith.muli %33, %282 : vector<4xi32>
            %285 = vector.reduction <add>, %284, %283 : vector<4xi32> into i32
            %286 = vector.insert %285, %cst_0 [0] : i32 into vector<1xi32>
            %287 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %288 = vector.insert %287, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %289 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %290 = vector.insert %289, %288 [1] : vector<1xi8> into vector<4x1xi8>
            %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %292 = vector.insert %291, %290 [2] : vector<1xi8> into vector<4x1xi8>
            %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %294 = vector.insert %293, %292 [3] : vector<1xi8> into vector<4x1xi8>
            %295 = arith.extsi %294 : vector<4x1xi8> to vector<4x1xi32>
            %296 = vector.extract %295[0, 0] : vector<4x1xi32>
            %297 = vector.insert %296, %cst_1 [0] : i32 into vector<4xi32>
            %298 = vector.extract %295[1, 0] : vector<4x1xi32>
            %299 = vector.insert %298, %297 [1] : i32 into vector<4xi32>
            %300 = vector.extract %295[2, 0] : vector<4x1xi32>
            %301 = vector.insert %300, %299 [2] : i32 into vector<4xi32>
            %302 = vector.extract %295[3, 0] : vector<4x1xi32>
            %303 = vector.insert %302, %301 [3] : i32 into vector<4xi32>
            %304 = vector.extract %arg7[0] : vector<4xi32>
            %305 = arith.muli %34, %303 : vector<4xi32>
            %306 = vector.reduction <add>, %305, %304 : vector<4xi32> into i32
            %307 = vector.insert %306, %cst_0 [0] : i32 into vector<1xi32>
            %308 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %309 = vector.insert %308, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %310 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %311 = vector.insert %310, %309 [1] : vector<1xi8> into vector<4x1xi8>
            %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %313 = vector.insert %312, %311 [2] : vector<1xi8> into vector<4x1xi8>
            %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %315 = vector.insert %314, %313 [3] : vector<1xi8> into vector<4x1xi8>
            %316 = arith.extsi %315 : vector<4x1xi8> to vector<4x1xi32>
            %317 = vector.extract %316[0, 0] : vector<4x1xi32>
            %318 = vector.insert %317, %cst_1 [0] : i32 into vector<4xi32>
            %319 = vector.extract %316[1, 0] : vector<4x1xi32>
            %320 = vector.insert %319, %318 [1] : i32 into vector<4xi32>
            %321 = vector.extract %316[2, 0] : vector<4x1xi32>
            %322 = vector.insert %321, %320 [2] : i32 into vector<4xi32>
            %323 = vector.extract %316[3, 0] : vector<4x1xi32>
            %324 = vector.insert %323, %322 [3] : i32 into vector<4xi32>
            %325 = vector.extract %arg7[1] : vector<4xi32>
            %326 = arith.muli %34, %324 : vector<4xi32>
            %327 = vector.reduction <add>, %326, %325 : vector<4xi32> into i32
            %328 = vector.insert %327, %cst_0 [0] : i32 into vector<1xi32>
            %329 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %330 = vector.insert %329, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %331 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %332 = vector.insert %331, %330 [1] : vector<1xi8> into vector<4x1xi8>
            %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %334 = vector.insert %333, %332 [2] : vector<1xi8> into vector<4x1xi8>
            %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %336 = vector.insert %335, %334 [3] : vector<1xi8> into vector<4x1xi8>
            %337 = arith.extsi %336 : vector<4x1xi8> to vector<4x1xi32>
            %338 = vector.extract %337[0, 0] : vector<4x1xi32>
            %339 = vector.insert %338, %cst_1 [0] : i32 into vector<4xi32>
            %340 = vector.extract %337[1, 0] : vector<4x1xi32>
            %341 = vector.insert %340, %339 [1] : i32 into vector<4xi32>
            %342 = vector.extract %337[2, 0] : vector<4x1xi32>
            %343 = vector.insert %342, %341 [2] : i32 into vector<4xi32>
            %344 = vector.extract %337[3, 0] : vector<4x1xi32>
            %345 = vector.insert %344, %343 [3] : i32 into vector<4xi32>
            %346 = vector.extract %arg7[2] : vector<4xi32>
            %347 = arith.muli %34, %345 : vector<4xi32>
            %348 = vector.reduction <add>, %347, %346 : vector<4xi32> into i32
            %349 = vector.insert %348, %cst_0 [0] : i32 into vector<1xi32>
            %350 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %351 = vector.insert %350, %cst [0] : vector<1xi8> into vector<4x1xi8>
            %352 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %353 = vector.insert %352, %351 [1] : vector<1xi8> into vector<4x1xi8>
            %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %355 = vector.insert %354, %353 [2] : vector<1xi8> into vector<4x1xi8>
            %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
            %357 = vector.insert %356, %355 [3] : vector<1xi8> into vector<4x1xi8>
            %358 = arith.extsi %357 : vector<4x1xi8> to vector<4x1xi32>
            %359 = vector.extract %358[0, 0] : vector<4x1xi32>
            %360 = vector.insert %359, %cst_1 [0] : i32 into vector<4xi32>
            %361 = vector.extract %358[1, 0] : vector<4x1xi32>
            %362 = vector.insert %361, %360 [1] : i32 into vector<4xi32>
            %363 = vector.extract %358[2, 0] : vector<4x1xi32>
            %364 = vector.insert %363, %362 [2] : i32 into vector<4xi32>
            %365 = vector.extract %358[3, 0] : vector<4x1xi32>
            %366 = vector.insert %365, %364 [3] : i32 into vector<4xi32>
            %367 = vector.extract %arg7[3] : vector<4xi32>
            %368 = arith.muli %34, %366 : vector<4xi32>
            %369 = vector.reduction <add>, %368, %367 : vector<4xi32> into i32
            %370 = vector.insert %369, %cst_0 [0] : i32 into vector<1xi32>
            %371 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %372 = vector.insert_strided_slice %76, %371 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %373 = vector.insert_strided_slice %97, %372 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %374 = vector.insert_strided_slice %118, %373 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %375 = vector.insert_strided_slice %139, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %376 = vector.insert_strided_slice %160, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %377 = vector.insert_strided_slice %181, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %378 = vector.insert_strided_slice %202, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %379 = vector.insert_strided_slice %223, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %380 = vector.insert_strided_slice %244, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %381 = vector.insert_strided_slice %265, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %382 = vector.insert_strided_slice %286, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            %383 = vector.insert_strided_slice %307, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
            %384 = vector.insert_strided_slice %328, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
            %385 = vector.insert_strided_slice %349, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
            %386 = vector.insert_strided_slice %370, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
            scf.yield %386, %382, %378, %374 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
          }
          %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
          %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
          scf.yield %inserted_slice : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 0 : index}
        scf.yield %14 : tensor<8x32xi32>
      } {iree.spirv.distribute_dim = 1 : index}
      flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    }
  }
  return
}

// -----// IR Dump After Canonicalizer (canonicalize) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %cst = arith.constant dense<0> : vector<4x1xi8>
    %cst_0 = arith.constant dense<0> : vector<1xi32>
    %cst_1 = arith.constant dense<0> : vector<4xi32>
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      scf.for %arg1 = %5 to %c1024 step %6 {
        %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
          %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
          %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
          %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
          %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
              %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
              %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
              %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
              %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
              %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
              %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
              %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
              %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
              %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
              %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
              %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
              %44 = vector.extract %43[0, 0] : vector<4x1xi32>
              %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
              %46 = vector.extract %43[1, 0] : vector<4x1xi32>
              %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
              %48 = vector.extract %43[2, 0] : vector<4x1xi32>
              %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
              %50 = vector.extract %43[3, 0] : vector<4x1xi32>
              %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
              %52 = vector.extract %arg10[0] : vector<4xi32>
              %53 = arith.muli %31, %51 : vector<4xi32>
              %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
              %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
              %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
              %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
              %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
              %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
              %65 = vector.extract %64[0, 0] : vector<4x1xi32>
              %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
              %67 = vector.extract %64[1, 0] : vector<4x1xi32>
              %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
              %69 = vector.extract %64[2, 0] : vector<4x1xi32>
              %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
              %71 = vector.extract %64[3, 0] : vector<4x1xi32>
              %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
              %73 = vector.extract %arg10[1] : vector<4xi32>
              %74 = arith.muli %31, %72 : vector<4xi32>
              %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
              %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
              %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
              %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
              %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
              %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
              %86 = vector.extract %85[0, 0] : vector<4x1xi32>
              %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
              %88 = vector.extract %85[1, 0] : vector<4x1xi32>
              %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
              %90 = vector.extract %85[2, 0] : vector<4x1xi32>
              %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
              %92 = vector.extract %85[3, 0] : vector<4x1xi32>
              %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
              %94 = vector.extract %arg10[2] : vector<4xi32>
              %95 = arith.muli %31, %93 : vector<4xi32>
              %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
              %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
              %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
              %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
              %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
              %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
              %107 = vector.extract %106[0, 0] : vector<4x1xi32>
              %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
              %109 = vector.extract %106[1, 0] : vector<4x1xi32>
              %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
              %111 = vector.extract %106[2, 0] : vector<4x1xi32>
              %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
              %113 = vector.extract %106[3, 0] : vector<4x1xi32>
              %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
              %115 = vector.extract %arg10[3] : vector<4xi32>
              %116 = arith.muli %31, %114 : vector<4xi32>
              %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
              %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
              %119 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %120 = vector.insert %119, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %121 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %122 = vector.insert %121, %120 [1] : vector<1xi8> into vector<4x1xi8>
              %123 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %124 = vector.insert %123, %122 [2] : vector<1xi8> into vector<4x1xi8>
              %125 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %126 = vector.insert %125, %124 [3] : vector<1xi8> into vector<4x1xi8>
              %127 = arith.extsi %126 : vector<4x1xi8> to vector<4x1xi32>
              %128 = vector.extract %127[0, 0] : vector<4x1xi32>
              %129 = vector.insert %128, %cst_1 [0] : i32 into vector<4xi32>
              %130 = vector.extract %127[1, 0] : vector<4x1xi32>
              %131 = vector.insert %130, %129 [1] : i32 into vector<4xi32>
              %132 = vector.extract %127[2, 0] : vector<4x1xi32>
              %133 = vector.insert %132, %131 [2] : i32 into vector<4xi32>
              %134 = vector.extract %127[3, 0] : vector<4x1xi32>
              %135 = vector.insert %134, %133 [3] : i32 into vector<4xi32>
              %136 = vector.extract %arg9[0] : vector<4xi32>
              %137 = arith.muli %32, %135 : vector<4xi32>
              %138 = vector.reduction <add>, %137, %136 : vector<4xi32> into i32
              %139 = vector.insert %138, %cst_0 [0] : i32 into vector<1xi32>
              %140 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %141 = vector.insert %140, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %142 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %143 = vector.insert %142, %141 [1] : vector<1xi8> into vector<4x1xi8>
              %144 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %145 = vector.insert %144, %143 [2] : vector<1xi8> into vector<4x1xi8>
              %146 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %147 = vector.insert %146, %145 [3] : vector<1xi8> into vector<4x1xi8>
              %148 = arith.extsi %147 : vector<4x1xi8> to vector<4x1xi32>
              %149 = vector.extract %148[0, 0] : vector<4x1xi32>
              %150 = vector.insert %149, %cst_1 [0] : i32 into vector<4xi32>
              %151 = vector.extract %148[1, 0] : vector<4x1xi32>
              %152 = vector.insert %151, %150 [1] : i32 into vector<4xi32>
              %153 = vector.extract %148[2, 0] : vector<4x1xi32>
              %154 = vector.insert %153, %152 [2] : i32 into vector<4xi32>
              %155 = vector.extract %148[3, 0] : vector<4x1xi32>
              %156 = vector.insert %155, %154 [3] : i32 into vector<4xi32>
              %157 = vector.extract %arg9[1] : vector<4xi32>
              %158 = arith.muli %32, %156 : vector<4xi32>
              %159 = vector.reduction <add>, %158, %157 : vector<4xi32> into i32
              %160 = vector.insert %159, %cst_0 [0] : i32 into vector<1xi32>
              %161 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %162 = vector.insert %161, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %163 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %164 = vector.insert %163, %162 [1] : vector<1xi8> into vector<4x1xi8>
              %165 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %166 = vector.insert %165, %164 [2] : vector<1xi8> into vector<4x1xi8>
              %167 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %168 = vector.insert %167, %166 [3] : vector<1xi8> into vector<4x1xi8>
              %169 = arith.extsi %168 : vector<4x1xi8> to vector<4x1xi32>
              %170 = vector.extract %169[0, 0] : vector<4x1xi32>
              %171 = vector.insert %170, %cst_1 [0] : i32 into vector<4xi32>
              %172 = vector.extract %169[1, 0] : vector<4x1xi32>
              %173 = vector.insert %172, %171 [1] : i32 into vector<4xi32>
              %174 = vector.extract %169[2, 0] : vector<4x1xi32>
              %175 = vector.insert %174, %173 [2] : i32 into vector<4xi32>
              %176 = vector.extract %169[3, 0] : vector<4x1xi32>
              %177 = vector.insert %176, %175 [3] : i32 into vector<4xi32>
              %178 = vector.extract %arg9[2] : vector<4xi32>
              %179 = arith.muli %32, %177 : vector<4xi32>
              %180 = vector.reduction <add>, %179, %178 : vector<4xi32> into i32
              %181 = vector.insert %180, %cst_0 [0] : i32 into vector<1xi32>
              %182 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %183 = vector.insert %182, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %184 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %185 = vector.insert %184, %183 [1] : vector<1xi8> into vector<4x1xi8>
              %186 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %187 = vector.insert %186, %185 [2] : vector<1xi8> into vector<4x1xi8>
              %188 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %189 = vector.insert %188, %187 [3] : vector<1xi8> into vector<4x1xi8>
              %190 = arith.extsi %189 : vector<4x1xi8> to vector<4x1xi32>
              %191 = vector.extract %190[0, 0] : vector<4x1xi32>
              %192 = vector.insert %191, %cst_1 [0] : i32 into vector<4xi32>
              %193 = vector.extract %190[1, 0] : vector<4x1xi32>
              %194 = vector.insert %193, %192 [1] : i32 into vector<4xi32>
              %195 = vector.extract %190[2, 0] : vector<4x1xi32>
              %196 = vector.insert %195, %194 [2] : i32 into vector<4xi32>
              %197 = vector.extract %190[3, 0] : vector<4x1xi32>
              %198 = vector.insert %197, %196 [3] : i32 into vector<4xi32>
              %199 = vector.extract %arg9[3] : vector<4xi32>
              %200 = arith.muli %32, %198 : vector<4xi32>
              %201 = vector.reduction <add>, %200, %199 : vector<4xi32> into i32
              %202 = vector.insert %201, %cst_0 [0] : i32 into vector<1xi32>
              %203 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %204 = vector.insert %203, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %205 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %206 = vector.insert %205, %204 [1] : vector<1xi8> into vector<4x1xi8>
              %207 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %208 = vector.insert %207, %206 [2] : vector<1xi8> into vector<4x1xi8>
              %209 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %210 = vector.insert %209, %208 [3] : vector<1xi8> into vector<4x1xi8>
              %211 = arith.extsi %210 : vector<4x1xi8> to vector<4x1xi32>
              %212 = vector.extract %211[0, 0] : vector<4x1xi32>
              %213 = vector.insert %212, %cst_1 [0] : i32 into vector<4xi32>
              %214 = vector.extract %211[1, 0] : vector<4x1xi32>
              %215 = vector.insert %214, %213 [1] : i32 into vector<4xi32>
              %216 = vector.extract %211[2, 0] : vector<4x1xi32>
              %217 = vector.insert %216, %215 [2] : i32 into vector<4xi32>
              %218 = vector.extract %211[3, 0] : vector<4x1xi32>
              %219 = vector.insert %218, %217 [3] : i32 into vector<4xi32>
              %220 = vector.extract %arg8[0] : vector<4xi32>
              %221 = arith.muli %33, %219 : vector<4xi32>
              %222 = vector.reduction <add>, %221, %220 : vector<4xi32> into i32
              %223 = vector.insert %222, %cst_0 [0] : i32 into vector<1xi32>
              %224 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %225 = vector.insert %224, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %226 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %227 = vector.insert %226, %225 [1] : vector<1xi8> into vector<4x1xi8>
              %228 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %229 = vector.insert %228, %227 [2] : vector<1xi8> into vector<4x1xi8>
              %230 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %231 = vector.insert %230, %229 [3] : vector<1xi8> into vector<4x1xi8>
              %232 = arith.extsi %231 : vector<4x1xi8> to vector<4x1xi32>
              %233 = vector.extract %232[0, 0] : vector<4x1xi32>
              %234 = vector.insert %233, %cst_1 [0] : i32 into vector<4xi32>
              %235 = vector.extract %232[1, 0] : vector<4x1xi32>
              %236 = vector.insert %235, %234 [1] : i32 into vector<4xi32>
              %237 = vector.extract %232[2, 0] : vector<4x1xi32>
              %238 = vector.insert %237, %236 [2] : i32 into vector<4xi32>
              %239 = vector.extract %232[3, 0] : vector<4x1xi32>
              %240 = vector.insert %239, %238 [3] : i32 into vector<4xi32>
              %241 = vector.extract %arg8[1] : vector<4xi32>
              %242 = arith.muli %33, %240 : vector<4xi32>
              %243 = vector.reduction <add>, %242, %241 : vector<4xi32> into i32
              %244 = vector.insert %243, %cst_0 [0] : i32 into vector<1xi32>
              %245 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %246 = vector.insert %245, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %247 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %248 = vector.insert %247, %246 [1] : vector<1xi8> into vector<4x1xi8>
              %249 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %250 = vector.insert %249, %248 [2] : vector<1xi8> into vector<4x1xi8>
              %251 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %252 = vector.insert %251, %250 [3] : vector<1xi8> into vector<4x1xi8>
              %253 = arith.extsi %252 : vector<4x1xi8> to vector<4x1xi32>
              %254 = vector.extract %253[0, 0] : vector<4x1xi32>
              %255 = vector.insert %254, %cst_1 [0] : i32 into vector<4xi32>
              %256 = vector.extract %253[1, 0] : vector<4x1xi32>
              %257 = vector.insert %256, %255 [1] : i32 into vector<4xi32>
              %258 = vector.extract %253[2, 0] : vector<4x1xi32>
              %259 = vector.insert %258, %257 [2] : i32 into vector<4xi32>
              %260 = vector.extract %253[3, 0] : vector<4x1xi32>
              %261 = vector.insert %260, %259 [3] : i32 into vector<4xi32>
              %262 = vector.extract %arg8[2] : vector<4xi32>
              %263 = arith.muli %33, %261 : vector<4xi32>
              %264 = vector.reduction <add>, %263, %262 : vector<4xi32> into i32
              %265 = vector.insert %264, %cst_0 [0] : i32 into vector<1xi32>
              %266 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %267 = vector.insert %266, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %268 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %269 = vector.insert %268, %267 [1] : vector<1xi8> into vector<4x1xi8>
              %270 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %271 = vector.insert %270, %269 [2] : vector<1xi8> into vector<4x1xi8>
              %272 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %273 = vector.insert %272, %271 [3] : vector<1xi8> into vector<4x1xi8>
              %274 = arith.extsi %273 : vector<4x1xi8> to vector<4x1xi32>
              %275 = vector.extract %274[0, 0] : vector<4x1xi32>
              %276 = vector.insert %275, %cst_1 [0] : i32 into vector<4xi32>
              %277 = vector.extract %274[1, 0] : vector<4x1xi32>
              %278 = vector.insert %277, %276 [1] : i32 into vector<4xi32>
              %279 = vector.extract %274[2, 0] : vector<4x1xi32>
              %280 = vector.insert %279, %278 [2] : i32 into vector<4xi32>
              %281 = vector.extract %274[3, 0] : vector<4x1xi32>
              %282 = vector.insert %281, %280 [3] : i32 into vector<4xi32>
              %283 = vector.extract %arg8[3] : vector<4xi32>
              %284 = arith.muli %33, %282 : vector<4xi32>
              %285 = vector.reduction <add>, %284, %283 : vector<4xi32> into i32
              %286 = vector.insert %285, %cst_0 [0] : i32 into vector<1xi32>
              %287 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %288 = vector.insert %287, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %289 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %290 = vector.insert %289, %288 [1] : vector<1xi8> into vector<4x1xi8>
              %291 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %292 = vector.insert %291, %290 [2] : vector<1xi8> into vector<4x1xi8>
              %293 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %294 = vector.insert %293, %292 [3] : vector<1xi8> into vector<4x1xi8>
              %295 = arith.extsi %294 : vector<4x1xi8> to vector<4x1xi32>
              %296 = vector.extract %295[0, 0] : vector<4x1xi32>
              %297 = vector.insert %296, %cst_1 [0] : i32 into vector<4xi32>
              %298 = vector.extract %295[1, 0] : vector<4x1xi32>
              %299 = vector.insert %298, %297 [1] : i32 into vector<4xi32>
              %300 = vector.extract %295[2, 0] : vector<4x1xi32>
              %301 = vector.insert %300, %299 [2] : i32 into vector<4xi32>
              %302 = vector.extract %295[3, 0] : vector<4x1xi32>
              %303 = vector.insert %302, %301 [3] : i32 into vector<4xi32>
              %304 = vector.extract %arg7[0] : vector<4xi32>
              %305 = arith.muli %34, %303 : vector<4xi32>
              %306 = vector.reduction <add>, %305, %304 : vector<4xi32> into i32
              %307 = vector.insert %306, %cst_0 [0] : i32 into vector<1xi32>
              %308 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %309 = vector.insert %308, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %310 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %311 = vector.insert %310, %309 [1] : vector<1xi8> into vector<4x1xi8>
              %312 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %313 = vector.insert %312, %311 [2] : vector<1xi8> into vector<4x1xi8>
              %314 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %315 = vector.insert %314, %313 [3] : vector<1xi8> into vector<4x1xi8>
              %316 = arith.extsi %315 : vector<4x1xi8> to vector<4x1xi32>
              %317 = vector.extract %316[0, 0] : vector<4x1xi32>
              %318 = vector.insert %317, %cst_1 [0] : i32 into vector<4xi32>
              %319 = vector.extract %316[1, 0] : vector<4x1xi32>
              %320 = vector.insert %319, %318 [1] : i32 into vector<4xi32>
              %321 = vector.extract %316[2, 0] : vector<4x1xi32>
              %322 = vector.insert %321, %320 [2] : i32 into vector<4xi32>
              %323 = vector.extract %316[3, 0] : vector<4x1xi32>
              %324 = vector.insert %323, %322 [3] : i32 into vector<4xi32>
              %325 = vector.extract %arg7[1] : vector<4xi32>
              %326 = arith.muli %34, %324 : vector<4xi32>
              %327 = vector.reduction <add>, %326, %325 : vector<4xi32> into i32
              %328 = vector.insert %327, %cst_0 [0] : i32 into vector<1xi32>
              %329 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %330 = vector.insert %329, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %331 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %332 = vector.insert %331, %330 [1] : vector<1xi8> into vector<4x1xi8>
              %333 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %334 = vector.insert %333, %332 [2] : vector<1xi8> into vector<4x1xi8>
              %335 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %336 = vector.insert %335, %334 [3] : vector<1xi8> into vector<4x1xi8>
              %337 = arith.extsi %336 : vector<4x1xi8> to vector<4x1xi32>
              %338 = vector.extract %337[0, 0] : vector<4x1xi32>
              %339 = vector.insert %338, %cst_1 [0] : i32 into vector<4xi32>
              %340 = vector.extract %337[1, 0] : vector<4x1xi32>
              %341 = vector.insert %340, %339 [1] : i32 into vector<4xi32>
              %342 = vector.extract %337[2, 0] : vector<4x1xi32>
              %343 = vector.insert %342, %341 [2] : i32 into vector<4xi32>
              %344 = vector.extract %337[3, 0] : vector<4x1xi32>
              %345 = vector.insert %344, %343 [3] : i32 into vector<4xi32>
              %346 = vector.extract %arg7[2] : vector<4xi32>
              %347 = arith.muli %34, %345 : vector<4xi32>
              %348 = vector.reduction <add>, %347, %346 : vector<4xi32> into i32
              %349 = vector.insert %348, %cst_0 [0] : i32 into vector<1xi32>
              %350 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %351 = vector.insert %350, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %352 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %353 = vector.insert %352, %351 [1] : vector<1xi8> into vector<4x1xi8>
              %354 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %355 = vector.insert %354, %353 [2] : vector<1xi8> into vector<4x1xi8>
              %356 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %357 = vector.insert %356, %355 [3] : vector<1xi8> into vector<4x1xi8>
              %358 = arith.extsi %357 : vector<4x1xi8> to vector<4x1xi32>
              %359 = vector.extract %358[0, 0] : vector<4x1xi32>
              %360 = vector.insert %359, %cst_1 [0] : i32 into vector<4xi32>
              %361 = vector.extract %358[1, 0] : vector<4x1xi32>
              %362 = vector.insert %361, %360 [1] : i32 into vector<4xi32>
              %363 = vector.extract %358[2, 0] : vector<4x1xi32>
              %364 = vector.insert %363, %362 [2] : i32 into vector<4xi32>
              %365 = vector.extract %358[3, 0] : vector<4x1xi32>
              %366 = vector.insert %365, %364 [3] : i32 into vector<4xi32>
              %367 = vector.extract %arg7[3] : vector<4xi32>
              %368 = arith.muli %34, %366 : vector<4xi32>
              %369 = vector.reduction <add>, %368, %367 : vector<4xi32> into i32
              %370 = vector.insert %369, %cst_0 [0] : i32 into vector<1xi32>
              %371 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %372 = vector.insert_strided_slice %76, %371 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %373 = vector.insert_strided_slice %97, %372 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %374 = vector.insert_strided_slice %118, %373 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %375 = vector.insert_strided_slice %139, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %376 = vector.insert_strided_slice %160, %375 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %377 = vector.insert_strided_slice %181, %376 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %378 = vector.insert_strided_slice %202, %377 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %379 = vector.insert_strided_slice %223, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %380 = vector.insert_strided_slice %244, %379 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %381 = vector.insert_strided_slice %265, %380 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %382 = vector.insert_strided_slice %286, %381 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %383 = vector.insert_strided_slice %307, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %384 = vector.insert_strided_slice %328, %383 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %385 = vector.insert_strided_slice %349, %384 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %386 = vector.insert_strided_slice %370, %385 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              scf.yield %386, %382, %378, %374 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
            }
            %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %14 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After CSE (cse) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %cst = arith.constant dense<0> : vector<4x1xi8>
    %cst_0 = arith.constant dense<0> : vector<1xi32>
    %cst_1 = arith.constant dense<0> : vector<4xi32>
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      scf.for %arg1 = %5 to %c1024 step %6 {
        %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
          %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
          %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
          %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
          %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
              %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
              %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
              %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
              %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
              %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
              %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
              %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
              %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
              %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
              %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
              %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
              %44 = vector.extract %43[0, 0] : vector<4x1xi32>
              %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
              %46 = vector.extract %43[1, 0] : vector<4x1xi32>
              %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
              %48 = vector.extract %43[2, 0] : vector<4x1xi32>
              %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
              %50 = vector.extract %43[3, 0] : vector<4x1xi32>
              %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
              %52 = vector.extract %arg10[0] : vector<4xi32>
              %53 = arith.muli %31, %51 : vector<4xi32>
              %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
              %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
              %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
              %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
              %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
              %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
              %65 = vector.extract %64[0, 0] : vector<4x1xi32>
              %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
              %67 = vector.extract %64[1, 0] : vector<4x1xi32>
              %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
              %69 = vector.extract %64[2, 0] : vector<4x1xi32>
              %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
              %71 = vector.extract %64[3, 0] : vector<4x1xi32>
              %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
              %73 = vector.extract %arg10[1] : vector<4xi32>
              %74 = arith.muli %31, %72 : vector<4xi32>
              %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
              %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
              %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
              %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
              %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
              %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
              %86 = vector.extract %85[0, 0] : vector<4x1xi32>
              %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
              %88 = vector.extract %85[1, 0] : vector<4x1xi32>
              %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
              %90 = vector.extract %85[2, 0] : vector<4x1xi32>
              %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
              %92 = vector.extract %85[3, 0] : vector<4x1xi32>
              %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
              %94 = vector.extract %arg10[2] : vector<4xi32>
              %95 = arith.muli %31, %93 : vector<4xi32>
              %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
              %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
              %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
              %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
              %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
              %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
              %107 = vector.extract %106[0, 0] : vector<4x1xi32>
              %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
              %109 = vector.extract %106[1, 0] : vector<4x1xi32>
              %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
              %111 = vector.extract %106[2, 0] : vector<4x1xi32>
              %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
              %113 = vector.extract %106[3, 0] : vector<4x1xi32>
              %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
              %115 = vector.extract %arg10[3] : vector<4xi32>
              %116 = arith.muli %31, %114 : vector<4xi32>
              %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
              %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
              %119 = vector.extract %arg9[0] : vector<4xi32>
              %120 = arith.muli %32, %51 : vector<4xi32>
              %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
              %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
              %123 = vector.extract %arg9[1] : vector<4xi32>
              %124 = arith.muli %32, %72 : vector<4xi32>
              %125 = vector.reduction <add>, %124, %123 : vector<4xi32> into i32
              %126 = vector.insert %125, %cst_0 [0] : i32 into vector<1xi32>
              %127 = vector.extract %arg9[2] : vector<4xi32>
              %128 = arith.muli %32, %93 : vector<4xi32>
              %129 = vector.reduction <add>, %128, %127 : vector<4xi32> into i32
              %130 = vector.insert %129, %cst_0 [0] : i32 into vector<1xi32>
              %131 = vector.extract %arg9[3] : vector<4xi32>
              %132 = arith.muli %32, %114 : vector<4xi32>
              %133 = vector.reduction <add>, %132, %131 : vector<4xi32> into i32
              %134 = vector.insert %133, %cst_0 [0] : i32 into vector<1xi32>
              %135 = vector.extract %arg8[0] : vector<4xi32>
              %136 = arith.muli %33, %51 : vector<4xi32>
              %137 = vector.reduction <add>, %136, %135 : vector<4xi32> into i32
              %138 = vector.insert %137, %cst_0 [0] : i32 into vector<1xi32>
              %139 = vector.extract %arg8[1] : vector<4xi32>
              %140 = arith.muli %33, %72 : vector<4xi32>
              %141 = vector.reduction <add>, %140, %139 : vector<4xi32> into i32
              %142 = vector.insert %141, %cst_0 [0] : i32 into vector<1xi32>
              %143 = vector.extract %arg8[2] : vector<4xi32>
              %144 = arith.muli %33, %93 : vector<4xi32>
              %145 = vector.reduction <add>, %144, %143 : vector<4xi32> into i32
              %146 = vector.insert %145, %cst_0 [0] : i32 into vector<1xi32>
              %147 = vector.extract %arg8[3] : vector<4xi32>
              %148 = arith.muli %33, %114 : vector<4xi32>
              %149 = vector.reduction <add>, %148, %147 : vector<4xi32> into i32
              %150 = vector.insert %149, %cst_0 [0] : i32 into vector<1xi32>
              %151 = vector.extract %arg7[0] : vector<4xi32>
              %152 = arith.muli %34, %51 : vector<4xi32>
              %153 = vector.reduction <add>, %152, %151 : vector<4xi32> into i32
              %154 = vector.insert %153, %cst_0 [0] : i32 into vector<1xi32>
              %155 = vector.extract %arg7[1] : vector<4xi32>
              %156 = arith.muli %34, %72 : vector<4xi32>
              %157 = vector.reduction <add>, %156, %155 : vector<4xi32> into i32
              %158 = vector.insert %157, %cst_0 [0] : i32 into vector<1xi32>
              %159 = vector.extract %arg7[2] : vector<4xi32>
              %160 = arith.muli %34, %93 : vector<4xi32>
              %161 = vector.reduction <add>, %160, %159 : vector<4xi32> into i32
              %162 = vector.insert %161, %cst_0 [0] : i32 into vector<1xi32>
              %163 = vector.extract %arg7[3] : vector<4xi32>
              %164 = arith.muli %34, %114 : vector<4xi32>
              %165 = vector.reduction <add>, %164, %163 : vector<4xi32> into i32
              %166 = vector.insert %165, %cst_0 [0] : i32 into vector<1xi32>
              %167 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %168 = vector.insert_strided_slice %76, %167 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %169 = vector.insert_strided_slice %97, %168 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %170 = vector.insert_strided_slice %118, %169 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %171 = vector.insert_strided_slice %122, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %172 = vector.insert_strided_slice %126, %171 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %173 = vector.insert_strided_slice %130, %172 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %174 = vector.insert_strided_slice %134, %173 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %175 = vector.insert_strided_slice %138, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %176 = vector.insert_strided_slice %142, %175 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %177 = vector.insert_strided_slice %146, %176 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %178 = vector.insert_strided_slice %150, %177 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %179 = vector.insert_strided_slice %154, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %180 = vector.insert_strided_slice %158, %179 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %181 = vector.insert_strided_slice %162, %180 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %182 = vector.insert_strided_slice %166, %181 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              scf.yield %182, %178, %174, %170 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
            }
            %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %14 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After ResolveShapedTypeResultDims (resolve-shaped-type-result-dims) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %cst = arith.constant dense<0> : vector<4x1xi8>
    %cst_0 = arith.constant dense<0> : vector<1xi32>
    %cst_1 = arith.constant dense<0> : vector<4xi32>
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      scf.for %arg1 = %5 to %c1024 step %6 {
        %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
          %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
          %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
          %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
          %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
              %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
              %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
              %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
              %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
              %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
              %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
              %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
              %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
              %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
              %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
              %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
              %44 = vector.extract %43[0, 0] : vector<4x1xi32>
              %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
              %46 = vector.extract %43[1, 0] : vector<4x1xi32>
              %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
              %48 = vector.extract %43[2, 0] : vector<4x1xi32>
              %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
              %50 = vector.extract %43[3, 0] : vector<4x1xi32>
              %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
              %52 = vector.extract %arg10[0] : vector<4xi32>
              %53 = arith.muli %31, %51 : vector<4xi32>
              %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
              %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
              %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
              %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
              %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
              %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
              %65 = vector.extract %64[0, 0] : vector<4x1xi32>
              %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
              %67 = vector.extract %64[1, 0] : vector<4x1xi32>
              %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
              %69 = vector.extract %64[2, 0] : vector<4x1xi32>
              %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
              %71 = vector.extract %64[3, 0] : vector<4x1xi32>
              %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
              %73 = vector.extract %arg10[1] : vector<4xi32>
              %74 = arith.muli %31, %72 : vector<4xi32>
              %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
              %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
              %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
              %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
              %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
              %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
              %86 = vector.extract %85[0, 0] : vector<4x1xi32>
              %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
              %88 = vector.extract %85[1, 0] : vector<4x1xi32>
              %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
              %90 = vector.extract %85[2, 0] : vector<4x1xi32>
              %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
              %92 = vector.extract %85[3, 0] : vector<4x1xi32>
              %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
              %94 = vector.extract %arg10[2] : vector<4xi32>
              %95 = arith.muli %31, %93 : vector<4xi32>
              %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
              %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
              %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
              %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
              %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
              %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
              %107 = vector.extract %106[0, 0] : vector<4x1xi32>
              %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
              %109 = vector.extract %106[1, 0] : vector<4x1xi32>
              %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
              %111 = vector.extract %106[2, 0] : vector<4x1xi32>
              %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
              %113 = vector.extract %106[3, 0] : vector<4x1xi32>
              %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
              %115 = vector.extract %arg10[3] : vector<4xi32>
              %116 = arith.muli %31, %114 : vector<4xi32>
              %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
              %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
              %119 = vector.extract %arg9[0] : vector<4xi32>
              %120 = arith.muli %32, %51 : vector<4xi32>
              %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
              %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
              %123 = vector.extract %arg9[1] : vector<4xi32>
              %124 = arith.muli %32, %72 : vector<4xi32>
              %125 = vector.reduction <add>, %124, %123 : vector<4xi32> into i32
              %126 = vector.insert %125, %cst_0 [0] : i32 into vector<1xi32>
              %127 = vector.extract %arg9[2] : vector<4xi32>
              %128 = arith.muli %32, %93 : vector<4xi32>
              %129 = vector.reduction <add>, %128, %127 : vector<4xi32> into i32
              %130 = vector.insert %129, %cst_0 [0] : i32 into vector<1xi32>
              %131 = vector.extract %arg9[3] : vector<4xi32>
              %132 = arith.muli %32, %114 : vector<4xi32>
              %133 = vector.reduction <add>, %132, %131 : vector<4xi32> into i32
              %134 = vector.insert %133, %cst_0 [0] : i32 into vector<1xi32>
              %135 = vector.extract %arg8[0] : vector<4xi32>
              %136 = arith.muli %33, %51 : vector<4xi32>
              %137 = vector.reduction <add>, %136, %135 : vector<4xi32> into i32
              %138 = vector.insert %137, %cst_0 [0] : i32 into vector<1xi32>
              %139 = vector.extract %arg8[1] : vector<4xi32>
              %140 = arith.muli %33, %72 : vector<4xi32>
              %141 = vector.reduction <add>, %140, %139 : vector<4xi32> into i32
              %142 = vector.insert %141, %cst_0 [0] : i32 into vector<1xi32>
              %143 = vector.extract %arg8[2] : vector<4xi32>
              %144 = arith.muli %33, %93 : vector<4xi32>
              %145 = vector.reduction <add>, %144, %143 : vector<4xi32> into i32
              %146 = vector.insert %145, %cst_0 [0] : i32 into vector<1xi32>
              %147 = vector.extract %arg8[3] : vector<4xi32>
              %148 = arith.muli %33, %114 : vector<4xi32>
              %149 = vector.reduction <add>, %148, %147 : vector<4xi32> into i32
              %150 = vector.insert %149, %cst_0 [0] : i32 into vector<1xi32>
              %151 = vector.extract %arg7[0] : vector<4xi32>
              %152 = arith.muli %34, %51 : vector<4xi32>
              %153 = vector.reduction <add>, %152, %151 : vector<4xi32> into i32
              %154 = vector.insert %153, %cst_0 [0] : i32 into vector<1xi32>
              %155 = vector.extract %arg7[1] : vector<4xi32>
              %156 = arith.muli %34, %72 : vector<4xi32>
              %157 = vector.reduction <add>, %156, %155 : vector<4xi32> into i32
              %158 = vector.insert %157, %cst_0 [0] : i32 into vector<1xi32>
              %159 = vector.extract %arg7[2] : vector<4xi32>
              %160 = arith.muli %34, %93 : vector<4xi32>
              %161 = vector.reduction <add>, %160, %159 : vector<4xi32> into i32
              %162 = vector.insert %161, %cst_0 [0] : i32 into vector<1xi32>
              %163 = vector.extract %arg7[3] : vector<4xi32>
              %164 = arith.muli %34, %114 : vector<4xi32>
              %165 = vector.reduction <add>, %164, %163 : vector<4xi32> into i32
              %166 = vector.insert %165, %cst_0 [0] : i32 into vector<1xi32>
              %167 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %168 = vector.insert_strided_slice %76, %167 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %169 = vector.insert_strided_slice %97, %168 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %170 = vector.insert_strided_slice %118, %169 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %171 = vector.insert_strided_slice %122, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %172 = vector.insert_strided_slice %126, %171 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %173 = vector.insert_strided_slice %130, %172 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %174 = vector.insert_strided_slice %134, %173 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %175 = vector.insert_strided_slice %138, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %176 = vector.insert_strided_slice %142, %175 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %177 = vector.insert_strided_slice %146, %176 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %178 = vector.insert_strided_slice %150, %177 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %179 = vector.insert_strided_slice %154, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %180 = vector.insert_strided_slice %158, %179 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %181 = vector.insert_strided_slice %162, %180 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %182 = vector.insert_strided_slice %166, %181 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              scf.yield %182, %178, %174, %170 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
            }
            %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %14 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After EliminateEmptyTensors (iree-eliminate-empty-tensors) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %cst = arith.constant dense<0> : vector<4x1xi8>
    %cst_0 = arith.constant dense<0> : vector<1xi32>
    %cst_1 = arith.constant dense<0> : vector<4xi32>
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      scf.for %arg1 = %5 to %c1024 step %6 {
        %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
          %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
          %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
          %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
          %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
              %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
              %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
              %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
              %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
              %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
              %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
              %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
              %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
              %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
              %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
              %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
              %44 = vector.extract %43[0, 0] : vector<4x1xi32>
              %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
              %46 = vector.extract %43[1, 0] : vector<4x1xi32>
              %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
              %48 = vector.extract %43[2, 0] : vector<4x1xi32>
              %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
              %50 = vector.extract %43[3, 0] : vector<4x1xi32>
              %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
              %52 = vector.extract %arg10[0] : vector<4xi32>
              %53 = arith.muli %31, %51 : vector<4xi32>
              %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
              %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
              %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
              %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
              %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
              %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
              %65 = vector.extract %64[0, 0] : vector<4x1xi32>
              %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
              %67 = vector.extract %64[1, 0] : vector<4x1xi32>
              %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
              %69 = vector.extract %64[2, 0] : vector<4x1xi32>
              %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
              %71 = vector.extract %64[3, 0] : vector<4x1xi32>
              %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
              %73 = vector.extract %arg10[1] : vector<4xi32>
              %74 = arith.muli %31, %72 : vector<4xi32>
              %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
              %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
              %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
              %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
              %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
              %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
              %86 = vector.extract %85[0, 0] : vector<4x1xi32>
              %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
              %88 = vector.extract %85[1, 0] : vector<4x1xi32>
              %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
              %90 = vector.extract %85[2, 0] : vector<4x1xi32>
              %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
              %92 = vector.extract %85[3, 0] : vector<4x1xi32>
              %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
              %94 = vector.extract %arg10[2] : vector<4xi32>
              %95 = arith.muli %31, %93 : vector<4xi32>
              %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
              %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
              %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
              %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
              %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
              %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
              %107 = vector.extract %106[0, 0] : vector<4x1xi32>
              %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
              %109 = vector.extract %106[1, 0] : vector<4x1xi32>
              %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
              %111 = vector.extract %106[2, 0] : vector<4x1xi32>
              %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
              %113 = vector.extract %106[3, 0] : vector<4x1xi32>
              %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
              %115 = vector.extract %arg10[3] : vector<4xi32>
              %116 = arith.muli %31, %114 : vector<4xi32>
              %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
              %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
              %119 = vector.extract %arg9[0] : vector<4xi32>
              %120 = arith.muli %32, %51 : vector<4xi32>
              %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
              %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
              %123 = vector.extract %arg9[1] : vector<4xi32>
              %124 = arith.muli %32, %72 : vector<4xi32>
              %125 = vector.reduction <add>, %124, %123 : vector<4xi32> into i32
              %126 = vector.insert %125, %cst_0 [0] : i32 into vector<1xi32>
              %127 = vector.extract %arg9[2] : vector<4xi32>
              %128 = arith.muli %32, %93 : vector<4xi32>
              %129 = vector.reduction <add>, %128, %127 : vector<4xi32> into i32
              %130 = vector.insert %129, %cst_0 [0] : i32 into vector<1xi32>
              %131 = vector.extract %arg9[3] : vector<4xi32>
              %132 = arith.muli %32, %114 : vector<4xi32>
              %133 = vector.reduction <add>, %132, %131 : vector<4xi32> into i32
              %134 = vector.insert %133, %cst_0 [0] : i32 into vector<1xi32>
              %135 = vector.extract %arg8[0] : vector<4xi32>
              %136 = arith.muli %33, %51 : vector<4xi32>
              %137 = vector.reduction <add>, %136, %135 : vector<4xi32> into i32
              %138 = vector.insert %137, %cst_0 [0] : i32 into vector<1xi32>
              %139 = vector.extract %arg8[1] : vector<4xi32>
              %140 = arith.muli %33, %72 : vector<4xi32>
              %141 = vector.reduction <add>, %140, %139 : vector<4xi32> into i32
              %142 = vector.insert %141, %cst_0 [0] : i32 into vector<1xi32>
              %143 = vector.extract %arg8[2] : vector<4xi32>
              %144 = arith.muli %33, %93 : vector<4xi32>
              %145 = vector.reduction <add>, %144, %143 : vector<4xi32> into i32
              %146 = vector.insert %145, %cst_0 [0] : i32 into vector<1xi32>
              %147 = vector.extract %arg8[3] : vector<4xi32>
              %148 = arith.muli %33, %114 : vector<4xi32>
              %149 = vector.reduction <add>, %148, %147 : vector<4xi32> into i32
              %150 = vector.insert %149, %cst_0 [0] : i32 into vector<1xi32>
              %151 = vector.extract %arg7[0] : vector<4xi32>
              %152 = arith.muli %34, %51 : vector<4xi32>
              %153 = vector.reduction <add>, %152, %151 : vector<4xi32> into i32
              %154 = vector.insert %153, %cst_0 [0] : i32 into vector<1xi32>
              %155 = vector.extract %arg7[1] : vector<4xi32>
              %156 = arith.muli %34, %72 : vector<4xi32>
              %157 = vector.reduction <add>, %156, %155 : vector<4xi32> into i32
              %158 = vector.insert %157, %cst_0 [0] : i32 into vector<1xi32>
              %159 = vector.extract %arg7[2] : vector<4xi32>
              %160 = arith.muli %34, %93 : vector<4xi32>
              %161 = vector.reduction <add>, %160, %159 : vector<4xi32> into i32
              %162 = vector.insert %161, %cst_0 [0] : i32 into vector<1xi32>
              %163 = vector.extract %arg7[3] : vector<4xi32>
              %164 = arith.muli %34, %114 : vector<4xi32>
              %165 = vector.reduction <add>, %164, %163 : vector<4xi32> into i32
              %166 = vector.insert %165, %cst_0 [0] : i32 into vector<1xi32>
              %167 = vector.insert_strided_slice %55, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %168 = vector.insert_strided_slice %76, %167 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %169 = vector.insert_strided_slice %97, %168 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %170 = vector.insert_strided_slice %118, %169 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %171 = vector.insert_strided_slice %122, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %172 = vector.insert_strided_slice %126, %171 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %173 = vector.insert_strided_slice %130, %172 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %174 = vector.insert_strided_slice %134, %173 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %175 = vector.insert_strided_slice %138, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %176 = vector.insert_strided_slice %142, %175 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %177 = vector.insert_strided_slice %146, %176 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %178 = vector.insert_strided_slice %150, %177 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              %179 = vector.insert_strided_slice %154, %cst_1 {offsets = [0], strides = [1]} : vector<1xi32> into vector<4xi32>
              %180 = vector.insert_strided_slice %158, %179 {offsets = [1], strides = [1]} : vector<1xi32> into vector<4xi32>
              %181 = vector.insert_strided_slice %162, %180 {offsets = [2], strides = [1]} : vector<1xi32> into vector<4xi32>
              %182 = vector.insert_strided_slice %166, %181 {offsets = [3], strides = [1]} : vector<1xi32> into vector<4xi32>
              scf.yield %182, %178, %174, %170 : vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>
            }
            %16 = vector.transfer_write %15#3, %extracted_slice[%c0, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %17 = vector.transfer_write %15#2, %16[%c1, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %18 = vector.transfer_write %15#1, %17[%c2, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %19 = vector.transfer_write %15#0, %18[%c3, %c0] {in_bounds = [true]} : vector<4xi32>, tensor<4x4xi32>
            %inserted_slice = tensor.insert_slice %19 into %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<4x4xi32> into tensor<8x32xi32>
            scf.yield %inserted_slice : tensor<8x32xi32>
          } {iree.spirv.distribute_dim = 0 : index}
          scf.yield %14 : tensor<8x32xi32>
        } {iree.spirv.distribute_dim = 1 : index}
        flow.dispatch.tensor.store %10, %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : tensor<8x32xi32> -> !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
      }
    }
    return
  }
}

// -----// IR Dump After EmptyTensorToAllocTensor (empty-tensor-to-alloc-tensor) //----- //
module {
  func.func @_main_dispatch_0_matmul_1024x1024x1024() {
    %cst = arith.constant dense<0> : vector<4x1xi8>
    %cst_0 = arith.constant dense<0> : vector<1xi32>
    %cst_1 = arith.constant dense<0> : vector<4xi32>
    %c1 = arith.constant 1 : index
    %c2 = arith.constant 2 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c32 = arith.constant 32 : index
    %c8 = arith.constant 8 : index
    %c4 = arith.constant 4 : index
    %c1024 = arith.constant 1024 : index
    %c0 = arith.constant 0 : index
    %0 = hal.interface.binding.subspan set(0) binding(0) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %1 = hal.interface.binding.subspan set(0) binding(1) type(storage_buffer) alignment(64) offset(%c0) flags(ReadOnly) : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>>
    %2 = hal.interface.binding.subspan set(0) binding(2) type(storage_buffer) alignment(64) offset(%c0) : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>>
    %workgroup_id_x = hal.interface.workgroup.id[0] : index
    %workgroup_count_x = hal.interface.workgroup.count[0] : index
    %workgroup_id_y = hal.interface.workgroup.id[1] : index
    %workgroup_count_y = hal.interface.workgroup.count[1] : index
    %3 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_id_y]
    %4 = affine.apply affine_map<()[s0] -> (s0 * 8)>()[%workgroup_count_y]
    %5 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_id_x]
    %6 = affine.apply affine_map<()[s0] -> (s0 * 32)>()[%workgroup_count_x]
    scf.for %arg0 = %3 to %c1024 step %4 {
      %7 = flow.dispatch.tensor.load %0, offsets = [%arg0, 0], sizes = [8, 1024], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<8x1024xi8>
      scf.for %arg1 = %5 to %c1024 step %6 {
        %8 = flow.dispatch.tensor.load %2, offsets = [%arg0, %arg1], sizes = [8, 32], strides = [1, 1] : !flow.dispatch.tensor<writeonly:tensor<1024x1024xi32>> -> tensor<8x32xi32>
        %9 = flow.dispatch.tensor.load %1, offsets = [0, %arg1], sizes = [1024, 32], strides = [1, 1] : !flow.dispatch.tensor<readonly:tensor<1024x1024xi8>> -> tensor<1024x32xi8>
        %10 = scf.for %arg2 = %c0 to %c8 step %c4 iter_args(%arg3 = %8) -> (tensor<8x32xi32>) {
          %11 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg2)
          %12 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg2)
          %13 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg2)
          %14 = scf.for %arg4 = %c0 to %c32 step %c4 iter_args(%arg5 = %arg3) -> (tensor<8x32xi32>) {
            %extracted_slice = tensor.extract_slice %arg5[%arg2, %arg4] [4, 4] [1, 1] : tensor<8x32xi32> to tensor<4x4xi32>
            %15:4 = scf.for %arg6 = %c0 to %c1024 step %c4 iter_args(%arg7 = %cst_1, %arg8 = %cst_1, %arg9 = %cst_1, %arg10 = %cst_1) -> (vector<4xi32>, vector<4xi32>, vector<4xi32>, vector<4xi32>) {
              %20 = vector.transfer_read %7[%arg2, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %21 = vector.transfer_read %7[%11, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %22 = vector.transfer_read %7[%12, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %23 = vector.transfer_read %7[%13, %arg6], %c0_i8 {in_bounds = [true]} : tensor<8x1024xi8>, vector<4xi8>
              %24 = vector.transfer_read %9[%arg6, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %25 = affine.apply affine_map<(d0) -> (d0 + 1)>(%arg6)
              %26 = vector.transfer_read %9[%25, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %27 = affine.apply affine_map<(d0) -> (d0 + 2)>(%arg6)
              %28 = vector.transfer_read %9[%27, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %29 = affine.apply affine_map<(d0) -> (d0 + 3)>(%arg6)
              %30 = vector.transfer_read %9[%29, %arg4], %c0_i8 {in_bounds = [true]} : tensor<1024x32xi8>, vector<4xi8>
              %31 = arith.extsi %20 : vector<4xi8> to vector<4xi32>
              %32 = arith.extsi %21 : vector<4xi8> to vector<4xi32>
              %33 = arith.extsi %22 : vector<4xi8> to vector<4xi32>
              %34 = arith.extsi %23 : vector<4xi8> to vector<4xi32>
              %35 = vector.extract_strided_slice %24 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %36 = vector.insert %35, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %37 = vector.extract_strided_slice %26 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %38 = vector.insert %37, %36 [1] : vector<1xi8> into vector<4x1xi8>
              %39 = vector.extract_strided_slice %28 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %40 = vector.insert %39, %38 [2] : vector<1xi8> into vector<4x1xi8>
              %41 = vector.extract_strided_slice %30 {offsets = [0], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %42 = vector.insert %41, %40 [3] : vector<1xi8> into vector<4x1xi8>
              %43 = arith.extsi %42 : vector<4x1xi8> to vector<4x1xi32>
              %44 = vector.extract %43[0, 0] : vector<4x1xi32>
              %45 = vector.insert %44, %cst_1 [0] : i32 into vector<4xi32>
              %46 = vector.extract %43[1, 0] : vector<4x1xi32>
              %47 = vector.insert %46, %45 [1] : i32 into vector<4xi32>
              %48 = vector.extract %43[2, 0] : vector<4x1xi32>
              %49 = vector.insert %48, %47 [2] : i32 into vector<4xi32>
              %50 = vector.extract %43[3, 0] : vector<4x1xi32>
              %51 = vector.insert %50, %49 [3] : i32 into vector<4xi32>
              %52 = vector.extract %arg10[0] : vector<4xi32>
              %53 = arith.muli %31, %51 : vector<4xi32>
              %54 = vector.reduction <add>, %53, %52 : vector<4xi32> into i32
              %55 = vector.insert %54, %cst_0 [0] : i32 into vector<1xi32>
              %56 = vector.extract_strided_slice %24 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %57 = vector.insert %56, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %58 = vector.extract_strided_slice %26 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %59 = vector.insert %58, %57 [1] : vector<1xi8> into vector<4x1xi8>
              %60 = vector.extract_strided_slice %28 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %61 = vector.insert %60, %59 [2] : vector<1xi8> into vector<4x1xi8>
              %62 = vector.extract_strided_slice %30 {offsets = [1], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %63 = vector.insert %62, %61 [3] : vector<1xi8> into vector<4x1xi8>
              %64 = arith.extsi %63 : vector<4x1xi8> to vector<4x1xi32>
              %65 = vector.extract %64[0, 0] : vector<4x1xi32>
              %66 = vector.insert %65, %cst_1 [0] : i32 into vector<4xi32>
              %67 = vector.extract %64[1, 0] : vector<4x1xi32>
              %68 = vector.insert %67, %66 [1] : i32 into vector<4xi32>
              %69 = vector.extract %64[2, 0] : vector<4x1xi32>
              %70 = vector.insert %69, %68 [2] : i32 into vector<4xi32>
              %71 = vector.extract %64[3, 0] : vector<4x1xi32>
              %72 = vector.insert %71, %70 [3] : i32 into vector<4xi32>
              %73 = vector.extract %arg10[1] : vector<4xi32>
              %74 = arith.muli %31, %72 : vector<4xi32>
              %75 = vector.reduction <add>, %74, %73 : vector<4xi32> into i32
              %76 = vector.insert %75, %cst_0 [0] : i32 into vector<1xi32>
              %77 = vector.extract_strided_slice %24 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %78 = vector.insert %77, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %79 = vector.extract_strided_slice %26 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %80 = vector.insert %79, %78 [1] : vector<1xi8> into vector<4x1xi8>
              %81 = vector.extract_strided_slice %28 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %82 = vector.insert %81, %80 [2] : vector<1xi8> into vector<4x1xi8>
              %83 = vector.extract_strided_slice %30 {offsets = [2], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %84 = vector.insert %83, %82 [3] : vector<1xi8> into vector<4x1xi8>
              %85 = arith.extsi %84 : vector<4x1xi8> to vector<4x1xi32>
              %86 = vector.extract %85[0, 0] : vector<4x1xi32>
              %87 = vector.insert %86, %cst_1 [0] : i32 into vector<4xi32>
              %88 = vector.extract %85[1, 0] : vector<4x1xi32>
              %89 = vector.insert %88, %87 [1] : i32 into vector<4xi32>
              %90 = vector.extract %85[2, 0] : vector<4x1xi32>
              %91 = vector.insert %90, %89 [2] : i32 into vector<4xi32>
              %92 = vector.extract %85[3, 0] : vector<4x1xi32>
              %93 = vector.insert %92, %91 [3] : i32 into vector<4xi32>
              %94 = vector.extract %arg10[2] : vector<4xi32>
              %95 = arith.muli %31, %93 : vector<4xi32>
              %96 = vector.reduction <add>, %95, %94 : vector<4xi32> into i32
              %97 = vector.insert %96, %cst_0 [0] : i32 into vector<1xi32>
              %98 = vector.extract_strided_slice %24 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %99 = vector.insert %98, %cst [0] : vector<1xi8> into vector<4x1xi8>
              %100 = vector.extract_strided_slice %26 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %101 = vector.insert %100, %99 [1] : vector<1xi8> into vector<4x1xi8>
              %102 = vector.extract_strided_slice %28 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %103 = vector.insert %102, %101 [2] : vector<1xi8> into vector<4x1xi8>
              %104 = vector.extract_strided_slice %30 {offsets = [3], sizes = [1], strides = [1]} : vector<4xi8> to vector<1xi8>
              %105 = vector.insert %104, %103 [3] : vector<1xi8> into vector<4x1xi8>
              %106 = arith.extsi %105 : vector<4x1xi8> to vector<4x1xi32>
              %107 = vector.extract %106[0, 0] : vector<4x1xi32>
              %108 = vector.insert %107, %cst_1 [0] : i32 into vector<4xi32>
              %109 = vector.extract %106[1, 0] : vector<4x1xi32>
              %110 = vector.insert %109, %108 [1] : i32 into vector<4xi32>
              %111 = vector.extract %106[2, 0] : vector<4x1xi32>
              %112 = vector.insert %111, %110 [2] : i32 into vector<4xi32>
              %113 = vector.extract %106[3, 0] : vector<4x1xi32>
              %114 = vector.insert %113, %112 [3] : i32 into vector<4xi32>
              %115 = vector.extract %arg10[3] : vector<4xi32>
              %116 = arith.muli %31, %114 : vector<4xi32>
              %117 = vector.reduction <add>, %116, %115 : vector<4xi32> into i32
              %118 = vector.insert %117, %cst_0 [0] : i32 into vector<1xi32>
              %119 = vector.extract %arg9[0] : vector<4xi32>
              %120 = arith.muli %32, %51 : vector<4xi32>
              %121 = vector.reduction <add>, %120, %119 : vector<4xi32> into i32
              %122 = vector.insert %121, %cst_0 [0] : i32 into vector<1xi32>
              %123 = vector.extract %arg9[1] : vector<4xi32>
              %124 = arith.muli %32, %72 : vector<4xi32>
              %125 = vector.reduction <add>, %124, %123 : vector<4xi32> into i32
              %126 = vector.insert %125, %cst_0 [0] : i32 into vector<1xi32>
              %127 = vector.extract %arg9[2] : vector<4xi32>
              %128 = arith.muli %32, %93 : vector<4xi32>
              %129 = vector.reduction <add>, %128, %127 : vector<4xi32> into i32
              %130 = vector.insert %129, %cst_0 [0] : i32 into vector<1xi32>
              %131 = vector.extract %arg9[3] : vector<4xi32>
              %132 = arith.muli %32, %114 : vector<4xi32>
              %133 = vector.reduction <add>, %132, %131 : vector<4xi32> into i32
              %134 = vector.insert %133, %cst_0 [0] : i32 into vector<1xi32>
              %135 = vector.extract %arg8[0] : vector<4xi32>
              %136 = arith.muli %33, %51 : vector<4xi32>
              %137 = vector.reduction <add>, %136, %135 : vector<4xi32> into i32
              %138 = vector.insert %137, %cst_0 [0] : i32 into vector<1xi32>
              %139 = vector.extract %arg8[1] : vector<4xi32>
              %140 = arith.muli %33, %72 : vector<4xi32>
              %141 = vector.reduction <add>, %140, %139 : vector<4xi32> into i32
              %142 = vector.insert %141, %cst_0 [0] : i32 into vector<1xi32>
              %143 = vector.extract %arg8[2] : vector<4xi32>
              %144 = arith.muli %33, %93 : vector<4xi32>
              %145 = vector.reduction <add>, %144, %143 : vector<4xi32> into i32
              %146 = vector.insert %145, %cst_0 [0] : i32 into vector<1xi32>
              %147 = vector.extract %arg8[3] : vector<4xi32>
              %148 =