maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[mlir][tosa] Do not fold transpose with quantized types 

For such cases, the type of the constant DenseElementsAttr is
different from the transpose op return type.

Reviewed By: rsuderman

Differential Revision: https://reviews.llvm.org/D110446
This commit is contained in:
Lei Zhang 2021-09-24 16:57:46 -04:00
parent d5f2013004
commit b45476c94c
2 changed files with 17 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
mlir/lib/Dialect/Tosa/IR/TosaOps.cpp
View File

@ -165,6 +165,12 @@ struct ConstantTransposeOptimization
LogicalResult matchAndRewrite(tosa::TransposeOp op,
PatternRewriter &rewriter) const override {
auto outputType = op.getType().cast<ShapedType>();
ArrayRef<int64_t> outputShape = outputType.getShape();
// TOSA supports quantized types.
if (!outputType.getElementType().isIntOrIndexOrFloat())
return failure();
DenseElementsAttr inputValues;
if (!matchPattern(op.input1(), m_Constant(&inputValues)))
return failure();
@ -184,9 +190,6 @@ struct ConstantTransposeOptimization
ArrayRef<int64_t> inputShape = inputType.getShape();
int64_t numElements = inputType.getNumElements();
auto outputType = op.getType().cast<ShapedType>();
ArrayRef<int64_t> outputShape = outputType.getShape();
SmallVector<Attribute, 4> outputValues;
outputValues.resize(numElements);

11
mlir/test/Dialect/Tosa/canonicalize.mlir
View File

@ -314,3 +314,14 @@ func @transpose_nofold_multi_users() -> (tensor<3x2xf32>, tensor<2x3xf32>) {
%1 = "tosa.transpose"(%input, %perms) : (tensor<2x3xf32>, tensor<2xi32>) -> tensor<3x2xf32>
return %1, %input : tensor<3x2xf32>, tensor<2x3xf32>
}
// -----
// CHECK-LABEL: @transpose_nofold_quantized_types
func @transpose_nofold_quantized_types() -> tensor<1x1x16x1x!quant.uniform<i8<-127:127>:f32:3, {1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,2.100000e+00,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01}>> {
%perms = "tosa.const"() {value = dense<[1, 2, 3, 0]> : tensor<4xi32>} : () -> tensor<4xi32>
%input = "tosa.const"() {value = dense<[[[[-127, 127, 127, -127, -127, -127, -127, -127, -127, 127, 127, 127, 127, 127, -127, 127]]]]> : tensor<1x1x1x16xi8>} : () -> tensor<1x1x1x16xi8>
// CHECK: tosa.transpose
%0 = "tosa.transpose"(%input, %perms) : (tensor<1x1x1x16xi8>, tensor<4xi32>) -> tensor<1x1x16x1x!quant.uniform<i8<-127:127>:f32:3, {1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,2.100000e+00,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01}>>
return %0: tensor<1x1x16x1x!quant.uniform<i8<-127:127>:f32:3, {1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,2.100000e+00,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01,1.000000e-01}>>
}