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] Add acc_type to TOSA conv ops in Torch to TOSA legalization #3958

Closed
wants to merge 1 commit into from
Closed

[TOSA] Add acc_type to TOSA conv ops in Torch to TOSA legalization #3958

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
6 changes: 6 additions & 0 deletions include/torch-mlir/Conversion/TorchToTosa/TosaLegalizeUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,6 +121,12 @@ void CreateReplaceOpAndInfer(PatternRewriter &rewriter, Operation *op,
LogicalResult getAvgPool2dAccType(PatternRewriter &rewriter, Value input,
TypeAttr &accType);

// Get accumulator type for TOSA convolution ops
LogicalResult getConvOpsAccType(PatternRewriter &rewriter,
RankedTensorType inputTy,
RankedTensorType weightTy,
RankedTensorType outputTy, TypeAttr &accType);

} // namespace tosa
} // namespace mlir

Expand Down
20 changes: 13 additions & 7 deletions lib/Conversion/TorchToTosa/TorchToTosa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2252,6 +2252,12 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
return rewriter.notifyMatchFailure(op,
"non-const dilation list unsupported");

TypeAttr accType;
if (failed(tosa::getConvOpsAccType(rewriter, inputTy, weightTy, outputTy,
accType)))
return rewriter.notifyMatchFailure(
op, "failed to get accumulator type for convolution ops");

// TOSA works in NHWC and takes OHWI (conv) / HWIM (depthwise conv) weights.
// Perform the necessary transformations.
std::optional<Value> nchwToNhwcTransposeConst =
Expand Down Expand Up @@ -2365,12 +2371,12 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
// full convolution
convOpResult =
rewriter
.create<tosa::Conv2DOp>(op->getLoc(),
getTypeConverter()->convertType(convOpTy),
transposedInput, transformedWeight, bias,
rewriter.getDenseI64ArrayAttr(padding),
rewriter.getDenseI64ArrayAttr(stride),
rewriter.getDenseI64ArrayAttr(dilation))
.create<tosa::Conv2DOp>(
op->getLoc(), getTypeConverter()->convertType(convOpTy),
transposedInput, transformedWeight, bias,
rewriter.getDenseI64ArrayAttr(padding),
rewriter.getDenseI64ArrayAttr(stride),
rewriter.getDenseI64ArrayAttr(dilation), accType)
.getResult();
} else if (weightShape[1] == 1) {
// depthwise convolution
Expand All @@ -2381,7 +2387,7 @@ LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
transposedInput, transformedWeight, bias,
rewriter.getDenseI64ArrayAttr(padding),
rewriter.getDenseI64ArrayAttr(stride),
rewriter.getDenseI64ArrayAttr(dilation))
rewriter.getDenseI64ArrayAttr(dilation), accType)
.getResult();
} else {
llvm_unreachable("Unhandled convolution type");
Expand Down
46 changes: 46 additions & 0 deletions lib/Conversion/TorchToTosa/TosaLegalizeUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,5 +436,51 @@ LogicalResult getAvgPool2dAccType(PatternRewriter &rewriter, Value input,
return success();
}

// Get accumulator type for TOSA convolution ops
LogicalResult getConvOpsAccType(PatternRewriter &rewriter,
RankedTensorType inputTy,
RankedTensorType weightTy,
RankedTensorType outputTy, TypeAttr &accType) {
auto inputElemTy = inputTy.getElementType();
auto weightElemTy = weightTy.getElementType();
auto outputElemTy = outputTy.getElementType();

auto quantTy = dyn_cast<quant::QuantizedType>(inputElemTy);
if (quantTy)
inputElemTy = quantTy.getStorageType();

// Get TOSA conv ops acc type based on input, weight, and output types
// according to the spec:
// https://www.mlplatform.org/tosa/tosa_spec.html#_conv2d
// https://www.mlplatform.org/tosa/tosa_spec.html#_depthwise_conv2d
// https://www.mlplatform.org/tosa/tosa_spec.html#_conv3d
//
// For undefined dtypes in TOSA like I64 and F64, acc_type will be set to the
// output type but does not offer any guarantee on the numerical precision
// since such cases will fail TOSA validation.
if ((inputElemTy.isF32() && weightElemTy.isF32() && outputElemTy.isF32()) ||
(inputElemTy.isF16() && weightElemTy.isF16() && outputElemTy.isF16()) ||
(inputElemTy.isBF16() && weightElemTy.isBF16() &&
outputElemTy.isBF16())) {
accType = mlir::TypeAttr::get(rewriter.getF32Type());
} else if (inputElemTy.isInteger(8) &&
(weightElemTy.isInteger(8) || weightElemTy.isInteger(4)) &&
outputElemTy.isInteger(32)) {
accType = mlir::TypeAttr::get(rewriter.getIntegerType(32));
} else if (inputElemTy.isInteger(16) && weightElemTy.isInteger(8) &&
outputElemTy.isInteger(48)) {
accType = mlir::TypeAttr::get(rewriter.getIntegerType(48));
} else if ((inputElemTy.isFloat8E4M3() && weightElemTy.isFloat8E4M3() &&
outputElemTy.isF16()) ||
(inputElemTy.isFloat8E5M2() && weightElemTy.isFloat8E5M2() &&
outputElemTy.isF16())) {
accType = mlir::TypeAttr::get(rewriter.getF16Type());
} else {
accType = mlir::TypeAttr::get(outputElemTy);
}

return success();
}

} // namespace tosa
} // namespace mlir
82 changes: 82 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 @@ -3057,3 +3057,85 @@ func.func @torch.aten.expm1$int(%arg0: !torch.vtensor<[3,4],si32>) -> !torch.vte
}

// -----

// CHECK-LABEL: func.func @torch.aten.convolution$basic(
// CHECK-SAME: %[[VAL_0:.*]]: !torch.vtensor<[5,2,10,20],f32>) -> !torch.vtensor<[5,10,14,24],f32> {
// CHECK: %[[VAL_1:.*]] = torch_c.to_builtin_tensor %[[VAL_0]] : !torch.vtensor<[5,2,10,20],f32> -> tensor<5x2x10x20xf32>
// CHECK: %[[VAL_2:.*]] = torch.constant.bool false
// CHECK: %[[VAL_3:.*]] = torch.constant.int 3
// CHECK: %[[VAL_4:.*]] = "tosa.const"() <{value = dense_resource<torch_tensor_10_2_3_3_torch.float32> : tensor<10x2x3x3xf32>}> : () -> tensor<10x2x3x3xf32>
// CHECK: %[[VAL_5:.*]] = torch.constant.none
// CHECK: %[[VAL_6:.*]] = torch.constant.int 1
// CHECK: %[[VAL_7:.*]] = torch.prim.ListConstruct %[[VAL_6]], %[[VAL_6]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_8:.*]] = torch.prim.ListConstruct %[[VAL_3]], %[[VAL_3]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_9:.*]] = torch.prim.ListConstruct %[[VAL_6]], %[[VAL_6]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_10:.*]] = torch.prim.ListConstruct : () -> !torch.list<int>
// CHECK: %[[VAL_11:.*]] = "tosa.const"() <{value = dense<0.000000e+00> : tensor<10xf32>}> : () -> tensor<10xf32>
// CHECK: %[[VAL_12:.*]] = "tosa.const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi32>}> : () -> tensor<4xi32>
// CHECK: %[[VAL_13:.*]] = tosa.transpose %[[VAL_1]], %[[VAL_12]] : (tensor<5x2x10x20xf32>, tensor<4xi32>) -> tensor<5x10x20x2xf32>
// CHECK: %[[VAL_14:.*]] = tosa.transpose %[[VAL_4]], %[[VAL_12]] : (tensor<10x2x3x3xf32>, tensor<4xi32>) -> tensor<10x3x3x2xf32>
// CHECK: %[[VAL_15:.*]] = tosa.conv2d %[[VAL_13]], %[[VAL_14]], %[[VAL_11]] {acc_type = f32, dilation = array<i64: 1, 1>, pad = array<i64: 3, 3, 3, 3>, stride = array<i64: 1, 1>} : (tensor<5x10x20x2xf32>, tensor<10x3x3x2xf32>, tensor<10xf32>) -> tensor<5x14x24x10xf32>
// CHECK: %[[VAL_16:.*]] = "tosa.const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}> : () -> tensor<4xi32>
// CHECK: %[[VAL_17:.*]] = tosa.transpose %[[VAL_15]], %[[VAL_16]] : (tensor<5x14x24x10xf32>, tensor<4xi32>) -> tensor<5x10x14x24xf32>
// CHECK: %[[VAL_18:.*]] = tensor.cast %[[VAL_17]] : tensor<5x10x14x24xf32> to tensor<5x10x14x24xf32>
// CHECK: %[[VAL_19:.*]] = torch_c.from_builtin_tensor %[[VAL_18]] : tensor<5x10x14x24xf32> -> !torch.vtensor<[5,10,14,24],f32>
// CHECK: return %[[VAL_19]] : !torch.vtensor<[5,10,14,24],f32>
// CHECK: }
func.func @torch.aten.convolution$basic(%arg0: !torch.vtensor<[5,2,10,20],f32>) -> !torch.vtensor<[5,10,14,24],f32> {
%false = torch.constant.bool false
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense_resource<torch_tensor_10_2_3_3_torch.float32> : tensor<10x2x3x3xf32>) : !torch.vtensor<[10,2,3,3],f32>
%none = torch.constant.none
%int1 = torch.constant.int 1
%1 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
%2 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
%3 = torch.prim.ListConstruct %int1, %int1 : (!torch.int, !torch.int) -> !torch.list<int>
%4 = torch.prim.ListConstruct : () -> !torch.list<int>
%5 = torch.aten.convolution %arg0, %0, %none, %1, %2, %3, %false, %4, %int1 : !torch.vtensor<[5,2,10,20],f32>, !torch.vtensor<[10,2,3,3],f32>, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[5,10,14,24],f32>
return %5 : !torch.vtensor<[5,10,14,24],f32>
}

// -----

// CHECK-LABEL: func.func @torch.aten.convolution$depthwise(
// CHECK-SAME: %[[VAL_0:.*]]: !torch.vtensor<[5,4,10,20],f32>) -> !torch.vtensor<[5,4,5,10],f32> {
// CHECK: %[[VAL_1:.*]] = torch_c.to_builtin_tensor %[[VAL_0]] : !torch.vtensor<[5,4,10,20],f32> -> tensor<5x4x10x20xf32>
// CHECK: %[[VAL_2:.*]] = torch.constant.bool false
// CHECK: %[[VAL_3:.*]] = torch.constant.int 4
// CHECK: %[[VAL_4:.*]] = torch.constant.int 3
// CHECK: %[[VAL_5:.*]] = "tosa.const"() <{value = dense_resource<torch_tensor_4_1_3_3_torch.float32> : tensor<4x1x3x3xf32>}> : () -> tensor<4x1x3x3xf32>
// CHECK: %[[VAL_6:.*]] = torch.constant.none
// CHECK: %[[VAL_7:.*]] = torch.constant.int 2
// CHECK: %[[VAL_8:.*]] = torch.prim.ListConstruct %[[VAL_7]], %[[VAL_7]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_9:.*]] = torch.prim.ListConstruct %[[VAL_4]], %[[VAL_4]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_10:.*]] = torch.prim.ListConstruct %[[VAL_4]], %[[VAL_4]] : (!torch.int, !torch.int) -> !torch.list<int>
// CHECK: %[[VAL_11:.*]] = torch.prim.ListConstruct : () -> !torch.list<int>
// CHECK: %[[VAL_12:.*]] = "tosa.const"() <{value = dense<0.000000e+00> : tensor<4xf32>}> : () -> tensor<4xf32>
// CHECK: %[[VAL_13:.*]] = "tosa.const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi32>}> : () -> tensor<4xi32>
// CHECK: %[[VAL_14:.*]] = tosa.transpose %[[VAL_1]], %[[VAL_13]] : (tensor<5x4x10x20xf32>, tensor<4xi32>) -> tensor<5x10x20x4xf32>
// CHECK: %[[VAL_15:.*]] = "tosa.const"() <{value = dense<[2, 3, 0, 1]> : tensor<4xi32>}> : () -> tensor<4xi32>
// CHECK: %[[VAL_16:.*]] = tosa.transpose %[[VAL_5]], %[[VAL_15]] : (tensor<4x1x3x3xf32>, tensor<4xi32>) -> tensor<3x3x4x1xf32>
// CHECK: %[[VAL_17:.*]] = tosa.reshape %[[VAL_16]] {new_shape = array<i64: 3, 3, 4, 1>} : (tensor<3x3x4x1xf32>) -> tensor<3x3x4x1xf32>
// CHECK: %[[VAL_18:.*]] = tosa.depthwise_conv2d %[[VAL_14]], %[[VAL_17]], %[[VAL_12]] {acc_type = f32, dilation = array<i64: 3, 3>, pad = array<i64: 3, 3, 3, 3>, stride = array<i64: 2, 2>} : (tensor<5x10x20x4xf32>, tensor<3x3x4x1xf32>, tensor<4xf32>) -> tensor<5x5x10x4xf32>
// CHECK: %[[VAL_19:.*]] = "tosa.const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}> : () -> tensor<4xi32>
// CHECK: %[[VAL_20:.*]] = tosa.transpose %[[VAL_18]], %[[VAL_19]] : (tensor<5x5x10x4xf32>, tensor<4xi32>) -> tensor<5x4x5x10xf32>
// CHECK: %[[VAL_21:.*]] = tensor.cast %[[VAL_20]] : tensor<5x4x5x10xf32> to tensor<5x4x5x10xf32>
// CHECK: %[[VAL_22:.*]] = torch_c.from_builtin_tensor %[[VAL_21]] : tensor<5x4x5x10xf32> -> !torch.vtensor<[5,4,5,10],f32>
// CHECK: return %[[VAL_22]] : !torch.vtensor<[5,4,5,10],f32>
// CHECK: }
func.func @torch.aten.convolution$depthwise(%arg0: !torch.vtensor<[5,4,10,20],f32>) -> !torch.vtensor<[5,4,5,10],f32> {
%false = torch.constant.bool false
%int4 = torch.constant.int 4
%int3 = torch.constant.int 3
%0 = torch.vtensor.literal(dense_resource<torch_tensor_4_1_3_3_torch.float32> : tensor<4x1x3x3xf32>) : !torch.vtensor<[4,1,3,3],f32>
%none = torch.constant.none
%int2 = torch.constant.int 2
%1 = torch.prim.ListConstruct %int2, %int2 : (!torch.int, !torch.int) -> !torch.list<int>
%2 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
%3 = torch.prim.ListConstruct %int3, %int3 : (!torch.int, !torch.int) -> !torch.list<int>
%4 = torch.prim.ListConstruct : () -> !torch.list<int>
%5 = torch.aten.convolution %arg0, %0, %none, %1, %2, %3, %false, %4, %int4 : !torch.vtensor<[5,4,10,20],f32>, !torch.vtensor<[4,1,3,3],f32>, !torch.none, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.list<int>, !torch.int -> !torch.vtensor<[5,4,5,10],f32>
return %5 : !torch.vtensor<[5,4,5,10],f32>
}

// -----
Loading