bjacob

error: the actual and expected result matrices disagree at row 7, column 5.

actual value: 0
expected value: -24


left-hand side (rows 0..7 out of 0..7, columns 0..7 out of 0..7)
  96  -48  126   36 -107   83   -3  -97
 -52   41   45  127   62  120 -100   66
  43  -37  -23   69  -40  126  -77 -100

bjacob

Command line:

iree/tools/iree-translate -iree-mlir-to-vm-bytecode-module -mlir-print-op-on-diagnostic=false --iree-hal-target-backends=dylib-llvm-aot /tmp/a.mlir -o /tmp/a.vmfb -iree-llvm-embedded-linker-path=third_party/llvm-project/llvm/bin/lld

Output:

/tmp/a.mlir:10:8: 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

bjacob

// -----// IR Dump After mlir::iree_compiler::Shape::(anonymous namespace)::ExpandFunctionDynamicDimsPass //----- //
module  {
  func @f(%arg0: !hal.buffer_view) -> !hal.buffer_view attributes {iree.abi.stub} {
    %c5 = arith.constant 5 : index
    %c3 = arith.constant 3 : index
    %c0_i8 = arith.constant 0 : i8
    %c0 = arith.constant 0 : index
    %c1 = arith.constant 1 : index
    %0 = hal.tensor.cast %arg0 : !hal.buffer_view -> tensor<1x1xi8>
    %1 = linalg.pad_tensor %0 low[0, 0] high[%c5, %c3]  {

bjacob

diff --git a/iree/compiler/Dialect/Flow/Transforms/PadTensorToSubTensorInsert.cpp b/iree/compiler/Dialect/Flow/Transforms/PadTensorToSubTensorInsert.cpp
index 7f3ad376e..362d4df27 100644
--- a/iree/compiler/Dialect/Flow/Transforms/PadTensorToSubTensorInsert.cpp
+++ b/iree/compiler/Dialect/Flow/Transforms/PadTensorToSubTensorInsert.cpp
@@ -90,8 +90,10 @@ struct PadTensorOpConversion : public OpRewritePattern<linalg::PadTensorOp> {
     Value fill =
         rewriter.create<linalg::FillOp>(loc, yieldVal, initTensor).getResult(0);
     SmallVector<OpFoldResult> strides(rank, rewriter.getI64IntegerAttr(1));
-    rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
-        padTensorOp, source, fill, lowPad, sourceShape, strides);

bjacob

// -----// IR Dump After mlir::iree_compiler::IREE::ABI::WrapEntryPointsPass //----- //
#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map2 = affine_map<(d0, d1, d2, d3) -> (d1, d3, d0, d2)>
module  {
  func @matmul_123x456xi8_times_456x789xi8_into_123x789xi32(%arg0: !hal.buffer_view, %arg1: !hal.buffer_view, %arg2: !hal.buffer_view) -> !hal.buffer_view attributes {iree.abi.stub} {
    %0 = hal.tensor.cast %arg0 : !hal.buffer_view -> tensor<123x456xi8>
    %1 = hal.tensor.cast %arg1 : !hal.buffer_view -> tensor<456x789xi8>
    %2 = hal.tensor.cast %arg2 : !hal.buffer_view -> tensor<123x789xi32>
    %3 = call @_matmul_123x456xi8_times_456x789xi8_into_123x789xi32(%0, %1, %2) : (tensor<123x456xi8>, tensor<456x789xi8>, tensor<123x789xi32>) -> tensor<123x789xi32>

bjacob

benoitjacob@desk:~$ cc -DM0=8 -DK0=4 -DN0=8 -DMMT4D_VARIANT_GENERIC mmt4dkernels.c -o /tmp/mmt4dkernels && /tmp/mmt4dkernels
called kernel: mmt4d_kernel_generic_8x4x8
57 30 63 45 27 45 33 42 
60 27 60 51 24 39 33 60 
60 24 66 45 36 42 36 33 
54 30 54 51 24 60 18 48 
69 45 66 57 18 54 39 48 
33 18 42 39 18 30 12 36 
72 39 63 63 30 66 39 51 

bjacob

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

bjacob

commit 75f1a3862cf6d099e850fca6afe31a910cd8cbe5
Author: Benoit Jacob <[email protected]>
Date:   Wed Jun 9 16:35:14 2021 -0400

    wip
    
    forfkorfkr

diff --git a/mlir/include/mlir/Conversion/Passes.h b/mlir/include/mlir/Conversion/Passes.h
index f3f8687984fc..cad5b2713c8e 100644

bjacob

#map0 = affine_map<(d0, d1, d2, d3) -> (d0, d2, d1, d3)>
#map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map2 = affine_map<(d0, d1, d2, d3) -> (d1, d3, d0, d2)>
module  {
  func @matmul_8x8xi8_times_8x8xi8_into_8x8xi32(%arg0: tensor<8x8xi8>, %arg1: tensor<8x8xi8>, %arg2: tensor<8x8xi32>) -> tensor<8x8xi32> {
    %0 = linalg.tensor_expand_shape %arg0 [[0, 1], [2, 3]] : tensor<8x8xi8> into tensor<1x8x2x4xi8>
    %1 = linalg.tensor_expand_shape %arg1 [[0, 1], [2, 3]] : tensor<8x8xi8> into tensor<2x4x1x8xi8>
    %2 = linalg.tensor_expand_shape %arg2 [[0, 1], [2, 3]] : tensor<8x8xi32> into tensor<1x8x1x8xi32>
    %3 = linalg.init_tensor [1, 2, 8, 4] : tensor<1x2x8x4xi8>
    %4 = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%0 : tensor<1x8x2x4xi8>) outs(%3 : tensor<1x2x8x4xi8>) {

bjacob

--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_1x1xi8_times_1x1xi8_into_1x1xi32] ---
--- CALL[module.matmul_1xDYNxi8_times_DYNx1xi8_into_1x1xi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---
--- CALL[module.matmul_DYNxDYNxi8_times_DYNxDYNxi8_into_DYNxDYNxi32] ---