Skip to content

[TOSA] Fix conv2d.padding decomposition to emit 2D padding for aten.convolution #4380

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+52 −2
Open

[TOSA] Fix conv2d.padding decomposition to emit 2D padding for aten.convolution #4380

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions lib/Conversion/TorchToTosa/TorchToTosa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2380,8 +2380,12 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
// padding {height, width}. The PyTorch OFM computation uses 2*pad in each
// spatial direction, implying the same top=bottom=height and left=right=width
// values for TOSA.
SmallVector<int64_t> padding(
Copy link
Member

@sahas3 sahas3 Nov 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does linalg path need a similar fix? If so, it may be better to add a canonicalization pattern for torch.aten.convolution so that padding is extended to be 2 dimension for all TorchToXXX paths in one place.

Copy link
Contributor Author

@hariprasadravi hariprasadravi Nov 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

torch-to-linalg asserts due the same assumption here:

torch-mlir/lib/Conversion/TorchToLinalg/Utils.cpp

Line 77 in b834f94

assert(static_cast<int64_t>(unpaddedDims + padding.size()) ==

torch-to-stablehlo also asserts due to the same assumption here:

torch-mlir/lib/Conversion/TorchToStablehlo/Linear.cpp

Line 788 in b834f94

assert(padding.size() == dilation.size() &&

So it looks like this is similar to #3885. Shall I proceed with a cononicalization pattern similar to the proposed fix in #4250 ?

Thoughts @sjarus and @zjgarvey ?

{padding_2d[0], padding_2d[0], padding_2d[1], padding_2d[1]});

int64_t padH = padding_2d[0];
// When padding is 'Valid', Torch produces 1D padding with only one value.
int64_t padW = (padding_2d.size() > 1) ? padding_2d[1] : padding_2d[0];

SmallVector<int64_t> padding({padH, padH, padW, padW});

SmallVector<int64_t, 2> dilation;
if (!matchPattern(adaptor.getDilation(), m_TorchListOfConstantInts(dilation)))
Expand Down
46 changes: 46 additions & 0 deletions test/Conversion/TorchToTosa/basic.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3891,6 +3891,52 @@ func.func @torch.aten.convolution$full_dim_indivisible_by_stride_with_sliced_inp

// -----

// CHECK-LABEL: func.func @torch.aten.convolution$valid_padding(
// CHECK-SAME: %[[INPUT_TENSOR:.*]]: !torch.vtensor<[1,1,5,5],f32>) -> !torch.vtensor<[1,1,5,5],f32> {
// CHECK: %[[INPUT_BUILTIN:.*]] = torch_c.to_builtin_tensor %[[INPUT_TENSOR]] : !torch.vtensor<[1,1,5,5],f32> -> tensor<1x1x5x5xf32>
// CHECK: %[[WEIGHT_CONST:.*]] = "tosa.const"() <{values = dense<-7.486820e-03> : tensor<1x1x1x1xf32>}> : () -> tensor<1x1x1x1xf32>
// CHECK: %[[BIAS_CONST:.*]] = "tosa.const"() <{values = dense<0.536443591> : tensor<1xf32>}> : () -> tensor<1xf32>
// CHECK: %[[STRIDE_H:.*]] = torch.constant.int 1
// CHECK: %[[STRIDE_W:.*]] = torch.constant.int 1
// CHECK: %[[STRIDES_LIST:.*]] = torch.prim.ListConstruct %[[STRIDE_H]], %[[STRIDE_W]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[PADDING_VAL:.*]] = torch.constant.int 0
// CHECK: %[[PADDING_LIST:.*]] = torch.prim.ListConstruct %[[PADDING_VAL]] : (!torch.int) -> !torch.list<int>
// CHECK: %[[DILATION_H:.*]] = torch.constant.int 1
// CHECK: %[[DILATION_W:.*]] = torch.constant.int 1
// CHECK: %[[DILATIONS_LIST:.*]] = torch.prim.ListConstruct %[[DILATION_H]], %[[DILATION_W]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[GROUPS_VAL:.*]] = torch.constant.bool false
// CHECK: %[[OUTPUT_PADDING_VAL:.*]] = torch.constant.int 0
// CHECK: %[[OUTPUT_PADDING_LIST:.*]] = torch.prim.ListConstruct %[[OUTPUT_PADDING_VAL]] : (!torch.int) -> !torch.list<int>
// CHECK: %[[CONV_DIMENSIONS:.*]] = torch.constant.int 1
// CHECK: %[[WEIGHT_TRANSPOSED:.*]] = tosa.transpose %[[WEIGHT_CONST]] {perms = array<i32: 0, 2, 3, 1>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
// CHECK: %[[INPUT_TRANSPOSED:.*]] = tosa.transpose %[[INPUT_BUILTIN]] {perms = array<i32: 0, 2, 3, 1>} : (tensor<1x1x5x5xf32>) -> tensor<1x5x5x1xf32>
// CHECK: %[[ZERO_BIAS_OP:.*]] = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
// CHECK: %[[ZERO_BIAS_OP_2:.*]] = "tosa.const"() <{values = dense<0.000000e+00> : tensor<1xf32>}> : () -> tensor<1xf32>
// CHECK: %[[CONV_RESULT_TOSA:.*]] = tosa.conv2d %[[INPUT_TRANSPOSED]], %[[WEIGHT_TRANSPOSED]], %[[BIAS_CONST]], %[[ZERO_BIAS_OP]], %[[ZERO_BIAS_OP_2]] {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x5x5x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x5x5x1xf32>
// CHECK: %[[OUTPUT_TRANSPOSED:.*]] = tosa.transpose %[[CONV_RESULT_TOSA]] {perms = array<i32: 0, 3, 1, 2>} : (tensor<1x5x5x1xf32>) -> tensor<1x1x5x5xf32>
// CHECK: %[[OUTPUT_TENSOR:.*]] = torch_c.from_builtin_tensor %[[OUTPUT_TRANSPOSED]] : tensor<1x1x5x5xf32> -> !torch.vtensor<[1,1,5,5],f32>
// CHECK: return %[[OUTPUT_TENSOR]] : !torch.vtensor<[1,1,5,5],f32>
func.func @torch.aten.convolution$valid_padding(%arg0: !torch.vtensor<[1,1,5,5],f32>) -> !torch.vtensor<[1,1,5,5],f32> {
%0 = torch.vtensor.literal(dense<-7.486820e-03> : tensor<1x1x1x1xf32>) : !torch.vtensor<[1,1,1,1],f32>
%1 = torch.vtensor.literal(dense<0.536443591> : tensor<1xf32>) : !torch.vtensor<[1],f32>
%int1 = torch.constant.int 1
%int1_0 = torch.constant.int 1
%2 = torch.prim.ListConstruct %int1, %int1_0 : (!torch.int, !torch.int) -> !torch.list<int>
%int0 = torch.constant.int 0
%3 = torch.prim.ListConstruct %int0 : (!torch.int) -> !torch.list<int>
%int1_1 = torch.constant.int 1
%int1_2 = torch.constant.int 1
%4 = torch.prim.ListConstruct %int1_1, %int1_2 : (!torch.int, !torch.int) -> !torch.list<int>
%false = torch.constant.bool false
%int0_3 = torch.constant.int 0
%5 = torch.prim.ListConstruct %int0_3 : (!torch.int) -> !torch.list<int>
%int1_4 = torch.constant.int 1
%6 = torch.aten.convolution %arg0, %0, %1, %2, %3, %4, %false, %5, %int1_4 : !torch.vtensor<[1,1,5,5],f32>, !torch.vtensor<[1,1,1,1],f32>, !torch.vtensor<[1],f32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[1,1,5,5],f32>
return %6 : !torch.vtensor<[1,1,5,5],f32>
}

// -----

// CHECK-LABEL: func.func @torch.aten.max_pool2d$zero_pad_with_sliced_input(
// CHECK-SAME: %[[VAL_0:[0-9]+|[a-zA-Z$._-][a-zA-Z0-9$._-]*]]: !torch.vtensor<[1,1,56,56],f32>) -> !torch.vtensor<[1,1,27,27],f32> {
// CHECK: %[[VAL_1:.*]] = torch_c.to_builtin_tensor %[[VAL_0]] : !torch.vtensor<[1,1,56,56],f32> -> tensor<1x1x56x56xf32>
Expand Down
Loading