Skip to content
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

[TOSA] Fold Up-Casts into MatMul when supported by tosa.matmul #198

Merged
merged 6 commits into from
Aug 13, 2024
Merged

[TOSA] Fold Up-Casts into MatMul when supported by tosa.matmul #198

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
0a94521
feat(torch.aten.mm): fold up-casts into matmul when supported in TOSA
cmcgirr-amd Jul 15, 2024
e637e14
fix: address PR comments
cmcgirr-amd Jul 15, 2024
511bf68
test(TorchToTosa): add more torch.aten.mm cases
cmcgirr-amd Jul 15, 2024
0ef1774
refactor(TorchToTosa): aten.mm if f16 use tosa.matmul(f16, f16) -> f16
cmcgirr-amd Jul 16, 2024
8e10629
refactor(TorchToTosa): add guard for aten.mm si16->si48
cmcgirr-amd Jul 16, 2024
f0eb1b2
refactor(TorchToTosa): remove AtenToDtype case as the op was already …
cmcgirr-amd Jul 16, 2024
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
60 changes: 56 additions & 4 deletions lib/Conversion/TorchToTosa/TorchToTosa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionDialect.h"
#include "torch-mlir/Dialect/TorchConversion/Transforms/BackendTypeConversion.h"
#include "llvm/ADT/TypeSwitch.h"
#include <numeric>
#include <optional>

Expand Down Expand Up @@ -1132,6 +1133,34 @@ Type getMatMulOutputType(Type inputElemTy, PatternRewriter &rewriter) {
return outputElemTy;
}

RankedTensorType getCastedInputTypeForMatmul(Value inputValue,
PatternRewriter &rewriter) {
// Check to see if the inputs to the matmul where casted from another type
auto preCastType =
TypeSwitch<Operation *, RankedTensorType>(inputValue.getDefiningOp())
.Case([](AtenToDtypeOp op) {
return cast<RankedTensorType>(op->getOperand(0).getType());
})
.Case([](tosa::CastOp op) {
return cast<RankedTensorType>(op->getOperand(0).getType());
})
.Default([](Operation * /*op*/) { return RankedTensorType(); });
if (!preCastType) {
return preCastType;
}
// Calculate the expected accumulator type based on the input type of the cast
cmcgirr-amd marked this conversation as resolved.
Show resolved Hide resolved
auto accumulatorType =
getMatMulOutputType(preCastType.getElementType(), rewriter);
// If the expected accumulatorType for the given input type to the cast
// matches the output type of the cast then we can fold the casting into the
TinaAMD marked this conversation as resolved.
Show resolved Hide resolved
// matmul. Because the casting is an up-cast and does not affect the numeric
// values due to rounding or saturation.
cmcgirr-amd marked this conversation as resolved.
Show resolved Hide resolved
return accumulatorType ==
cast<RankedTensorType>(inputValue.getType()).getElementType()
? preCastType
: RankedTensorType();
}

// Perform the basic n-dim matmul operation encompassing the handling of
// broadcasting and dynamic shape propagation.
// All PyTorch ops that leverage matrix multiplication will derive this and
Expand Down Expand Up @@ -1173,6 +1202,28 @@ class ConvertAtenMatmulBaseOp : public OpConversionPattern<AtenOpT> {
return rewriter.notifyMatchFailure(op,
"Matmul: input datatypes mismatched");

// Step: check if the inputs have been casted from a supported input type to
// an accumulator type and insert casts back to the original type if true
RankedTensorType lhsPreCastedType =
getCastedInputTypeForMatmul(lhs, rewriter);
RankedTensorType rhsPreCastedType =
getCastedInputTypeForMatmul(rhs, rewriter);
if (lhsPreCastedType && (lhsPreCastedType.getElementType() ==
rhsPreCastedType.getElementType())) {
cmcgirr-amd marked this conversation as resolved.
Show resolved Hide resolved
lhs = rewriter.create<tosa::CastOp>(
lhs.getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
lhsPreCastedType),
lhs);
rhs = rewriter.create<tosa::CastOp>(
rhs.getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
rhsPreCastedType),
rhs);
lhsElemTy = cast<RankedTensorType>(lhsPreCastedType).getElementType();
rhsElemTy = cast<RankedTensorType>(rhsPreCastedType).getElementType();
}

auto outputElemTy = getMatMulOutputType(lhsElemTy, rewriter);
if (!outputElemTy) {
return rewriter.notifyMatchFailure(
Expand Down Expand Up @@ -1565,12 +1616,13 @@ class ConvertAtenMatmulBaseOp : public OpConversionPattern<AtenOpT> {
matmulLhs, matmulRhs)
.getResult();

auto torchOpOutputType = lhsTy.getElementType();
auto castOutputTy = RankedTensorType::get(
makeShapeLLVMCompatible(matmulOutputShape), lhsElemTy);
makeShapeLLVMCompatible(matmulOutputShape), torchOpOutputType);
auto castResult = rewriter.createOrFold<tosa::CastOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(castOutputTy),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
castOutputTy),
mmOpResult);

// Perform the reshape to output shape. This is always required unless max
Expand Down Expand Up @@ -1673,7 +1725,7 @@ class ConvertAtenMatmulBaseOp : public OpConversionPattern<AtenOpT> {

// Perform reshape
auto reshapedOpType = RankedTensorType::get(
makeShapeLLVMCompatible(reshapedOpShape), lhsElemTy);
makeShapeLLVMCompatible(reshapedOpShape), outputElemTy);
auto reshapedOp = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
Expand Down
40 changes: 40 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 @@ -85,6 +85,46 @@ func.func @torch.aten.mm_2d(%arg0 : !torch.vtensor<[2,6],f32>, %arg1 : !torch.vt

// -----

// CHECK: tosa.matmul{{.*}} : (tensor<1x4x8xbf16>, tensor<1x8x16xbf16>) -> tensor<1x4x16xf32>
func.func @torch.aten.mm_bf16(%arg0: !torch.vtensor<[4,8],bf16>, %arg1: !torch.vtensor<[8,16],bf16>) -> !torch.vtensor<[4,16],f32> {
%false = torch.constant.bool false
%none = torch.constant.none
%int6 = torch.constant.int 6
%0 = torch.aten.to.dtype %arg0, %int6, %false, %false, %none : !torch.vtensor<[4,8],bf16>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[4,8],f32>
%1 = torch.aten.to.dtype %arg1, %int6, %false, %false, %none : !torch.vtensor<[8,16],bf16>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[8,16],f32>
%2 = torch.aten.mm %0, %1 : !torch.vtensor<[4,8],f32>, !torch.vtensor<[8,16],f32> -> !torch.vtensor<[4,16],f32>
return %2 : !torch.vtensor<[4,16],f32>
}

// -----

// CHECK: tosa.matmul{{.*}} : (tensor<1x4x8xf16>, tensor<1x8x16xf16>) -> tensor<1x4x16xf32>
func.func @torch.aten.mm_f16(%arg0: !torch.vtensor<[4,8],f16>, %arg1: !torch.vtensor<[8,16],f16>) -> !torch.vtensor<[4,16],f32> {
%false = torch.constant.bool false
%none = torch.constant.none
%int6 = torch.constant.int 6
%0 = torch.aten.to.dtype %arg0, %int6, %false, %false, %none : !torch.vtensor<[4,8],f16>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[4,8],f32>
%1 = torch.aten.to.dtype %arg1, %int6, %false, %false, %none : !torch.vtensor<[8,16],f16>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[8,16],f32>
%2 = torch.aten.mm %0, %1 : !torch.vtensor<[4,8],f32>, !torch.vtensor<[8,16],f32> -> !torch.vtensor<[4,16],f32>
return %2 : !torch.vtensor<[4,16],f32>
}


// -----

// CHECK: tosa.matmul{{.*}} : (tensor<1x4x8xi8>, tensor<1x8x16xi8>) -> tensor<1x4x16xi32>
func.func @torch.aten.mm_i8(%arg0: !torch.vtensor<[4,8],si8>, %arg1: !torch.vtensor<[8,16],si8>) -> !torch.vtensor<[4,16],si32> {
%false = torch.constant.bool false
%none = torch.constant.none
%int3 = torch.constant.int 3
%0 = torch.aten.to.dtype %arg0, %int3, %false, %false, %none : !torch.vtensor<[4,8],si8>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[4,8],si32>
%1 = torch.aten.to.dtype %arg1, %int3, %false, %false, %none : !torch.vtensor<[8,16],si8>, !torch.int, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[8,16],si32>
%2 = torch.aten.mm %0, %1 : !torch.vtensor<[4,8],si32>, !torch.vtensor<[8,16],si32> -> !torch.vtensor<[4,16],si32>
return %2 : !torch.vtensor<[4,16],si32>
}

// -----

// CHECK: %[[VAL_2:.+]] = tosa.reshape %0 {new_shape = array<i64: 100, 6, 2>} : (tensor<10x10x6x2xf32>) -> tensor<100x6x2xf32>
// CHECK-NEXT: %[[VAL_3:.+]] = tosa.reshape %1 {new_shape = array<i64: 100, 2, 6>} : (tensor<10x10x2x6xf32>) -> tensor<100x2x6xf32>
// CHECK-NEXT: %[[VAL_4:.+]] = tosa.matmul %[[VAL_2]], %[[VAL_3]] : (tensor<100x6x2xf32>, tensor<100x2x6xf32>) -> tensor<100x6x6xf32>
Expand Down
Loading