README.md

benoitjacob@cloud:~/iree-build-linux$ iree/tools/iree-translate -iree-mlir-to-vm-bytecode-module -mlir-print-op-on-diagnostic=false --iree-hal-target-backends=dylib-llvm-aot a.mlir -o /usr/local/google/home/benoitjacob/iree-build-linux/iree/test/e2e/regression/e2e_matmul_mmt4d_f32_large_dylib-llvm-aot_dylib.vmfb -iree-llvm-embedded-linker-path="/usr/local/google/home/benoitjacob/iree-build-linux/third_party/llvm-project/llvm/bin/lld" -print-ir-after-all

// -----// IR Dump After mlir::iree_compiler::Shape::(anonymous namespace)::ExpandFunctionDynamicDimsPass //----- //
module  {
  func @matmul_123x456xf32_times_456x789xf32_into_123x789xf32(%arg0: !hal.buffer_view) -> !hal.buffer_view attributes {iree.abi.stub} {
    %c5 = arith.constant 5 : index
    %c0 = arith.constant 0 : index
    %cst = arith.constant 0.000000e+00 : f32
    %c2 = arith.constant 2 : index
    %0 = hal.tensor.cast %arg0 : !hal.buffer_view -> tensor<123x456xf32>
    %1 = linalg.pad_tensor %0 low[0, 0] high[%c5, %c0]  {
    ^bb0(%arg1: index, %arg2: index):  // no predecessors
      linalg.yield %cst : f32
    } : tensor<123x456xf32> to tensor<?x?xf32>
    %2 = linalg.tensor_expand_shape %1 [[0, 1], [2, 3]] : tensor<?x?xf32> into tensor<?x8x?x1xf32>
    %3 = tensor.dim %2, %c0 : tensor<?x8x?x1xf32>
    %4 = tensor.dim %2, %c2 : tensor<?x8x?x1xf32>
    %5 = hal.tensor.cast %2 : tensor<?x8x?x1xf32>{%3, %4} -> !hal.buffer_view
    return %5 : !hal.buffer_view
  }
}


/usr/local/google/home/benoitjacob/iree-build-linux/iree/test/e2e/regression/e2e_matmul_mmt4d_f32_large_dylib-llvm-aot_dylib_module.opt.mlir:10:10: error: 'linalg.tensor_expand_shape' op expected dimension 0 of collapsed type to be dynamic since one or more of the corresponding dimensions in the expanded type is dynamic
    %3 = linalg.tensor_expand_shape %0 [[0, 1], [2, 3]] : tensor<?x?xf32> into tensor<?x8x?x1xf32>
         ^
/usr/local/google/home/benoitjacob/iree-build-linux/iree/test/e2e/regression/e2e_matmul_mmt4d_f32_large_dylib-llvm-aot_dylib_module.opt.mlir:2:3: note: called from
  func @matmul_123x456xf32_times_456x789xf32_into_123x789xf32(%arg0: tensor<123x456xf32>) -> tensor<?x8x?x1xf32> {
  ^
// -----// IR Dump After PadTensorToSubTensorInsert Failed //----- //
"builtin.module"() ( {
  "builtin.func"() ( {
  ^bb0(%arg0: !hal.buffer_view):  // no predecessors
    %0 = "arith.constant"() {value = 0 : index} : () -> index
    %1 = "arith.constant"() {value = 0.000000e+00 : f32} : () -> f32
    %2 = "arith.constant"() {value = 2 : index} : () -> index
    %3 = "arith.constant"() {value = 123 : index} : () -> index
    %4 = "arith.constant"() {value = 456 : index} : () -> index
    %5 = "hal.tensor.cast"(%arg0) {operand_segment_sizes = dense<[1, 0, 0]> : vector<3xi32>} : (!hal.buffer_view) -> tensor<123x456xf32>
    %6 = "linalg.init_tensor"() {static_sizes = [128, 456]} : () -> tensor<128x456xf32>
    %7 = "linalg.fill"(%1, %6) ( {
    ^bb0(%arg1: f32, %arg2: f32):  // no predecessors
      "linalg.yield"(%arg1) : (f32) -> ()
    }) : (f32, tensor<128x456xf32>) -> tensor<128x456xf32>
    %8 = "tensor.insert_slice"(%5, %7, %3, %4) {operand_segment_sizes = dense<[1, 1, 0, 2, 0]> : vector<5xi32>, static_offsets = [0, 0], static_sizes = [-1, -1], static_strides = [1, 1]} : (tensor<123x456xf32>, tensor<128x456xf32>, index, index) -> tensor<128x456xf32>
    %9 = "linalg.tensor_expand_shape"(%8) {reassociation = [[0, 1], [2, 3]]} : (tensor<128x456xf32>) -> tensor<?x8x?x1xf32>
    %10 = "tensor.dim"(%9, %0) : (tensor<?x8x?x1xf32>, index) -> index
    %11 = "tensor.dim"(%9, %2) : (tensor<?x8x?x1xf32>, index) -> index
    %12 = "hal.tensor.cast"(%9, %10, %11) {operand_segment_sizes = dense<[1, 2, 0]> : vector<3xi32>} : (tensor<?x8x?x1xf32>, index, index) -> !hal.buffer_view
    "std.return"(%12) : (!hal.buffer_view) -> ()
  }) {iree.abi.stub, sym_name = "matmul_123x456xf32_times_456x789xf32_into_123x789xf32", type = (!hal.buffer_view) -> !hal.buffer_view} : () -> ()
}) : () -> ()

a.mlir

module  {
  func @matmul_123x456xf32_times_456x789xf32_into_123x789xf32(%arg0: tensor<123x456xf32>) -> tensor<?x8x?x1xf32> {
    %c5 = arith.constant 5 : index
    %c0 = arith.constant 0 : index
    %cst = arith.constant 0.000000e+00 : f32
    %0 = linalg.pad_tensor %arg0 low[0, 0] high[%c5, %c0]  {
    ^bb0(%arg3: index, %arg4: index):  // no predecessors
      linalg.yield %cst : f32
    } : tensor<123x456xf32> to tensor<?x?xf32>
    %3 = linalg.tensor_expand_shape %0 [[0, 1], [2, 3]] : tensor<?x?xf32> into tensor<?x8x?x1xf32>
    return %3 : tensor<?x8x?x1xf32>
  }
}