Skip to content

[mlir][tosa] Remove out_shape from transpose_conv2d #129133

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

Merged
merged 1 commit into from
Feb 28, 2025
Merged

[mlir][tosa] Remove out_shape from transpose_conv2d #129133

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
4 changes: 1 addition & 3 deletions mlir/include/mlir/Dialect/Tosa/IR/TosaOpBase.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -148,13 +148,11 @@ def Tosa_TransConvOpQuantInfoBuilder : OpBuilder<
"::mlir::Value":$weight, "mlir::Value":$bias,
"::mlir::DenseI64ArrayAttr":$outpad,
"::mlir::DenseI64ArrayAttr":$stride,
"::mlir::DenseI64ArrayAttr":$outputShape,
"::mlir::TypeAttr":$acc_type),
[{
buildTransConvOpWithQuantInfo($_builder, $_state, outputType,
input, weight, bias,
outpad, stride,
outputShape, acc_type);
outpad, stride, acc_type);
}]>;

// The tosa.matmul op is also intended to be generated where a fully_connected
Expand Down
1 change: 0 additions & 1 deletion mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -408,7 +408,6 @@ def Tosa_TransposeConv2DOp : Tosa_ConvOp<"transpose_conv2d"> {

Tosa_IntArrayAttr4:$out_pad,
Tosa_IntArrayAttr2:$stride,
Tosa_IntArrayAttr4:$out_shape,
TypeAttrOf<Tosa_AccType>:$acc_type,
DefaultValuedOptionalAttr<BoolAttr, "false">:$local_bound
);
Expand Down
14 changes: 6 additions & 8 deletions mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -569,15 +569,15 @@ static void buildConvOpWithQuantInfo(OpBuilder &builder, OperationState &result,

/// Handles tosa.transpose_conv2d which has outpad and output shape
/// attributes.
static void buildTransConvOpWithQuantInfo(
OpBuilder &builder, OperationState &result, Type outputType, Value input,
Value weight, Value bias, DenseI64ArrayAttr outpad,
DenseI64ArrayAttr stride, DenseI64ArrayAttr outputShape, TypeAttr accType) {
static void
buildTransConvOpWithQuantInfo(OpBuilder &builder, OperationState &result,
Type outputType, Value input, Value weight,
Value bias, DenseI64ArrayAttr outpad,
DenseI64ArrayAttr stride, TypeAttr accType) {
auto zps = createZPsAsConst(builder, input, weight);
result.addOperands({input, weight, bias, zps.first, zps.second});
result.addAttribute("out_pad", outpad);
result.addAttribute("stride", stride);
result.addAttribute("out_shape", outputShape);
result.addAttribute("acc_type", accType);
Type finalOutputType = outputType;
auto quantAttr = buildConvOpQuantizationAttr(builder, input, weight);
Expand Down Expand Up @@ -2327,9 +2327,7 @@ LogicalResult TransposeConv2DOp::inferReturnTypeComponents(
MLIRContext *context, ::std::optional<Location> location,
TransposeConv2DOp::Adaptor adaptor,
SmallVectorImpl<ShapedTypeComponents> &inferredReturnShapes) {
// outputShape is mutable.
llvm::SmallVector<int64_t> outputShape =
convertToMlirShape(adaptor.getOutShape());
llvm::SmallVector<int64_t> outputShape(4, ShapedType::kDynamic);

int64_t inputWidth = ShapedType::kDynamic;
int64_t inputHeight = ShapedType::kDynamic;
Expand Down
11 changes: 1 addition & 10 deletions mlir/test/Dialect/Tosa/invalid.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,7 +168,7 @@ func.func @test_depthwise_conv2d_acc_type(%arg0: tensor<1x4x4x4xi8>, %arg1: tens
func.func @test_transpose_conv2d(%arg0: tensor<1x32x32x8xi8>, %arg1: tensor<16x1x1x8xi8>, %arg2: tensor<16xi8>) -> tensor<1x32x32x16xi8> {
%zp = "tosa.const"() {value = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
// expected-error@+1 {{'tosa.transpose_conv2d' op accumulator type for i8 tensor is not i32}}
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = f16, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32, 16>, stride = array<i64: 1, 1>} : (tensor<1x32x32x8xi8>, tensor<16x1x1x8xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x32x32x16xi8>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = f16, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<1x32x32x8xi8>, tensor<16x1x1x8xi8>, tensor<16xi8>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x32x32x16xi8>
return %0 : tensor<1x32x32x16xi8>
}

Expand Down Expand Up @@ -741,15 +741,6 @@ func.func @test_table_io_shape_mismatch(%arg0: tensor<?x16xi16>, %arg1: tensor<6

// -----

// CHECK-LABEL: test_transpose_conv2d_invalid_outshape
func.func @test_transpose_conv2d_invalid_outshape(%arg0: tensor<1x32x32x8xf32>, %arg1: tensor<16x1x1x8xf32>, %arg2: tensor<16xf32>) -> tensor<1x32x32x16xf32> {
// expected-error@+1 {{'tosa.transpose_conv2d' op attribute 'out_shape' failed to satisfy constraint: i64 dense array attribute with exactly 4 elements}}
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2 {out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: 1, 32, 32>, stride = array<i64: 1, 1>} : (tensor<1x32x32x8xf32>, tensor<16x1x1x8xf32>, tensor<16xf32>) -> tensor<1x32x32x16xf32>
return %0 : tensor<1x32x32x16xf32>
}

// -----

// CHECK-LABEL: test_mul_type_mismatch
func.func @test_mul_type_mismatch(%arg0: tensor<13x21x3xf32>, %arg1: tensor<13x1x3xf16>) -> tensor<13x21x3xf32> {
%shift = "tosa.const"() {value = dense<0> : tensor<1xi8>} : () -> tensor<1xi8>
Expand Down
22 changes: 9 additions & 13 deletions mlir/test/Dialect/Tosa/tosa-decompose-transpose-conv.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,19 +32,16 @@ func.func @transpose_conv2d_quantized(%arg0: tensor<2x16x14x3xi8>, %arg1: tensor

// CHECK-LABEL: @transpose_conv2d_quantized_padded
func.func @transpose_conv2d_quantized_padded(%arg0: tensor<2x16x14x3xi8>, %arg1: tensor<5x3x6x3xi8>, %arg2: tensor<5xi32>) -> (tensor<2x21x26x5xi32>) {
// CHECK-DAG: %[[INPUT_ZP:.+]] = "tosa.const"() <{value = dense<-22> : tensor<1xi8>}
// CHECK-DAG: %[[WEIGHT_ZP:.+]] = "tosa.const"() <{value = dense<42> : tensor<1xi8>}
// CHECK-DAG: %[[REV0:.+]] = tosa.reverse %2 {axis = 2 : i32}
// CHECK-DAG: %[[REV1:.+]] = tosa.reverse %arg1 {axis = 1 : i32}
// CHECK: tosa.conv2d %arg0, %3, %arg2, %[[INPUT_ZP]], %[[WEIGHT_ZP]]
// CHECK-SAME: dilation = array<i64: 1, 1>, pad = array<i64: 3, 4, 8, 9>,
// CHECK-SAME: stride = array<i64: 1, 1>}
%input_zp = "tosa.const"() {value = dense<-22> : tensor<1xi8>} : () -> tensor<1xi8>
%weight_zp = "tosa.const"() {value = dense<42> : tensor<1xi8>} : () -> tensor<1xi8>
// CHECK-DAG: %[[INPUT_ZP:.+]] = "tosa.const"() <{value = dense<-22> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK-DAG: %[[WEIGHT_ZP:.+]] = "tosa.const"() <{value = dense<42> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK-DAG: %[[REV0:.+]] = tosa.reverse %arg1 {axis = 1 : i32}
// CHECK-DAG: %[[REV1:.+]] = tosa.reverse %[[REV0]] {axis = 2 : i32}
// CHECK: tosa.conv2d %arg0, %[[REV1]], %arg2, %[[INPUT_ZP]], %[[WEIGHT_ZP]] {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 3, 4, 8, 9>, stride = array<i64: 1, 1>}
%input_zp = "tosa.const"() <{value = dense<-22> : tensor<1xi8>}> : () -> tensor<1xi8>
%weight_zp = "tosa.const"() <{value = dense<42> : tensor<1xi8>}> : () -> tensor<1xi8>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {
acc_type = i32,
out_pad = array<i64: 1, 2, 3, 4>,
out_shape = array<i64: -1, -1, -1, -1>,
stride = array<i64: 1, 1>} : (tensor<2x16x14x3xi8>, tensor<5x3x6x3xi8>, tensor<5xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<2x21x26x5xi32>
return %0 : tensor<2x21x26x5xi32>
}
Expand Down Expand Up @@ -160,12 +157,11 @@ func.func @transpose_conv2d_strided_overpad(%arg0 : tensor<1x16x1x1xi8>, %arg1 :
// CHECK: %[[PAD_RESULT:.+]] = tosa.pad %[[RESHAPE_RESULT_1]], %[[RESULT_PAD]]
// CHECK: %[[RESHAPE_ARG2:.+]] = tosa.reshape %arg2, %[[CONST10]]
// CHECK: %[[ADD:.+]] = tosa.add %[[PAD_RESULT]], %[[RESHAPE_ARG2]]
%input_zp = "tosa.const"() {value = dense<-103> : tensor<1xi8>} : () -> tensor<1xi8>
%weight_zp = "tosa.const"() {value = dense<93> : tensor<1xi8>} : () -> tensor<1xi8>
%input_zp = "tosa.const"() <{value = dense<-103> : tensor<1xi8>}> : () -> tensor<1xi8>
%weight_zp = "tosa.const"() <{value = dense<93> : tensor<1xi8>}> : () -> tensor<1xi8>
%2 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %input_zp, %weight_zp {
acc_type = i32,
out_pad = array<i64: 2, 0, 0, 1>,
out_shape = array<i64: 1, -1, -1, 1>,
stride = array<i64: 1, 2>} :
(tensor<1x16x1x1xi8>, tensor<1x2x1x1xi8>, tensor<1xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<1x19x2x1xi32>
"func.return" (%2) : (tensor<1x19x2x1xi32>) -> ()
Expand Down
16 changes: 8 additions & 8 deletions mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -907,7 +907,7 @@ func.func @depthwise_conv2d_strided(%arg0: tensor<1x13x14x1xf32>, %arg1: tensor<
// CHECK-LABEL: @transpose_conv2d_out_shape
func.func @transpose_conv2d_out_shape(%arg0: tensor<2x?x?x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x8x9x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, 8, 9, -1>, stride = array<i64: 1, 1>} : (tensor<2x?x?x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x?x?x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>
return
}

Expand All @@ -916,7 +916,7 @@ func.func @transpose_conv2d_out_shape(%arg0: tensor<2x?x?x3xf32>, %arg1: tensor<
// CHECK-LABEL: @transpose_conv2d_static
func.func @transpose_conv2d_static(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x18x19x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 1, 1>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
return
}

Expand All @@ -925,7 +925,7 @@ func.func @transpose_conv2d_static(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5
// CHECK-LABEL: @transpose_conv2d_static_strided
func.func @transpose_conv2d_static_strided(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x33x45x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 2, 3>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 2, 3>} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
return
}

Expand All @@ -934,7 +934,7 @@ func.func @transpose_conv2d_static_strided(%arg0: tensor<2x16x14x3xf32>, %arg1:
// CHECK-LABEL: @transpose_conv2d_dynamic_input
func.func @transpose_conv2d_dynamic_input(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<?x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<?x?x?x?xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<?x?x?x5xf32>
return
}

Expand All @@ -943,7 +943,7 @@ func.func @transpose_conv2d_dynamic_input(%arg0: tensor<?x?x?x?xf32>, %arg1: ten
// CHECK-LABEL: @transpose_conv2d_dynamic_weights
func.func @transpose_conv2d_dynamic_weights(%arg0: tensor<2x6x4x3xf32>, %arg1: tensor<?x?x?x?xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<?x?x?x?xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<?x?x?x?xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x?x?x5xf32>
return
}

Expand All @@ -952,7 +952,7 @@ func.func @transpose_conv2d_dynamic_weights(%arg0: tensor<2x6x4x3xf32>, %arg1: t
// CHECK-LABEL: @transpose_conv2d_dynamic_bias
func.func @transpose_conv2d_dynamic_bias(%arg0: tensor<2x6x4x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<?xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x8x9x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<5x3x6x3xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<2x6x4x3xf32>, tensor<5x3x6x3xf32>, tensor<?xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x8x9x5xf32>
return
}

Expand All @@ -961,14 +961,14 @@ func.func @transpose_conv2d_dynamic_bias(%arg0: tensor<2x6x4x3xf32>, %arg1: tens
// CHECK-LABEL: @transpose_conv2d_padded
func.func @transpose_conv2d_padded(%arg0: tensor<2x9x11x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<2x10x13x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 1, 0, 3, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 1, 1>} : (tensor<2x9x11x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x10x13x5xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 1, 0, 3, 0>, stride = array<i64: 1, 1>} : (tensor<2x9x11x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<2x10x13x5xf32>
return
}

// CHECK-LABEL: @transpose_conv2d_strided
func.func @transpose_conv2d_strided(%arg0: tensor<1x5x7x1xf32>, %arg1: tensor<1x1x1x1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>, %arg4: tensor<1xf32>) {
// CHECK: -> tensor<1x13x13x1xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, out_shape = array<i64: -1, -1, -1, -1>, stride = array<i64: 3, 2>} : (tensor<1x5x7x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x13x13x1xf32>
%0 = tosa.transpose_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = f32, out_pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 3, 2>} : (tensor<1x5x7x1xf32>, tensor<1x1x1x1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1x13x13x1xf32>
return
}

Expand Down