Skip to content

[mlir][tosa] Fix several bugs in DepthwiseConv2DIsMul #129210

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
Mar 3, 2025
Merged

[mlir][tosa] Fix several bugs in DepthwiseConv2DIsMul #129210

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
35 changes: 19 additions & 16 deletions mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeDepthwise.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,25 @@ struct DepthwiseConv2DIsMul : public OpRewritePattern<tosa::DepthwiseConv2DOp> {
return failure();
}

Type inputETy = inputType.getElementType();
Type weightETy = weightType.getElementType();
if (!inputETy.isIntOrFloat() || !weightETy.isIntOrFloat())
return rewriter.notifyMatchFailure(op, "unsupported type");

// Get and verify zero points.
int64_t iZp;
int64_t wZp;

if (op.getInputZeroPoint(iZp).failed() ||
op.getWeightZeroPoint(wZp).failed())
return rewriter.notifyMatchFailure(
op, "bail out if zero points cannot statically be determined");

if (op.verifyInputZeroPoint(iZp).failed() ||
op.verifyWeightZeroPoint(wZp).failed())
return rewriter.notifyMatchFailure(
op, "zero point must be zero for non-int8 integer types");

// Reshape input to [N, H, W, C] -> [N, H, W, C, 1].
ArrayRef<int64_t> inputShape = inputType.getShape();
llvm::SmallVector<int64_t, 2> revisedInputShape{
Expand All @@ -62,8 +81,6 @@ struct DepthwiseConv2DIsMul : public OpRewritePattern<tosa::DepthwiseConv2DOp> {
revisedInputShapeValue)
.getResult();

Type inputETy = inputType.getElementType();
Type weightETy = weightType.getElementType();
Type resultETy = resultType.getElementType();

if (inputETy != resultETy) {
Expand All @@ -76,20 +93,6 @@ struct DepthwiseConv2DIsMul : public OpRewritePattern<tosa::DepthwiseConv2DOp> {
weight = rewriter.create<tosa::CastOp>(op.getLoc(), weightType, weight);
}

// Get and verify zero points.
int64_t iZp;
int64_t wZp;

if (op.getInputZeroPoint(iZp).failed() ||
op.getWeightZeroPoint(wZp).failed())
return rewriter.notifyMatchFailure(
op, "bail out if zero points cannot statically be determined");

if (op.verifyInputZeroPoint(iZp).failed() ||
op.verifyWeightZeroPoint(wZp).failed())
return rewriter.notifyMatchFailure(
op, "zero point must be zero for non-int8 integer types");

if (iZp != 0 || wZp != 0) {

auto applyZp = [&](Value val, int64_t zp) -> Value {
Expand Down
22 changes: 22 additions & 0 deletions mlir/test/Dialect/Tosa/tosa-decompose-depthwise.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,3 +76,25 @@ func.func @depthwise_conv2d_as_mul_padded(%arg0: tensor<4x10x10x2xf32>, %arg1: t
%0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %zp, %zp {acc_type = f32, pad = array<i64: 1, 1, 1, 1>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<4x10x10x2xf32>, tensor<1x1x2x3xf32>, tensor<6xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<4x12x12x6xf32>
return %0 : tensor<4x12x12x6xf32>
}

// -----

// Decompose only support integer or float types.

// CHECK-LABEL: @depthwise_conv2d_quant_type
func.func @depthwise_conv2d_quant_type(%arg0: tensor<4x10x10x2x!quant.uniform<i8:f32, 0.015684768557548523>>, %arg1: tensor<1x1x2x3x!quant.uniform<i8<-127:127>:f32, 0.015680249780416489>>, %arg2: tensor<6xi32>) -> tensor<4x10x10x6x!quant.uniform<i32:f32, 0.078431375324726104>> {
%0 = "tosa.const"() <{value = dense<7> : tensor<1xi8>}> : () -> tensor<1xi8>
%1 = "tosa.const"() <{value = dense<11> : tensor<1xi8>}> : () -> tensor<1xi8>
// CHECK: tosa.depthwise_conv2d
%2 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %0, %1 {acc_type = i32, dilation = array<i64: 1, 1>, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>} : (tensor<4x10x10x2x!quant.uniform<i8:f32, 0.015684768557548523>>, tensor<1x1x2x3x!quant.uniform<i8<-127:127>:f32, 0.015680249780416489>>, tensor<6xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<4x10x10x6x!quant.uniform<i32:f32, 0.078431375324726104>>
return %2 : tensor<4x10x10x6x!quant.uniform<i32:f32, 0.078431375324726104>>
}

// -----

// CHECK-LABEL: @depthwise_conv2d_no_const_zero_point
func.func @depthwise_conv2d_no_const_zero_point(%arg0: tensor<4x10x10x2xi8>, %arg1: tensor<1x1x2x3xi8>, %arg2: tensor<6xi32>, %arg3: tensor<1xi8>, %arg4: tensor<1xi8>) -> tensor<4x10x10x6xi32> {
// CHECK: tosa.depthwise_conv2d
%0 = tosa.depthwise_conv2d %arg0, %arg1, %arg2, %arg3, %arg4 {acc_type = i32, pad = array<i64: 0, 0, 0, 0>, stride = array<i64: 1, 1>, dilation = array<i64: 1, 1>} : (tensor<4x10x10x2xi8>, tensor<1x1x2x3xi8>, tensor<6xi32>, tensor<1xi8>, tensor<1xi8>) -> tensor<4x10x10x6xi32>
return %0 : tensor<4x10x10x6xi32>
}