Skip to content

Commit

Permalink
Lower onnx.split to torch.aten (#2686)
Browse files Browse the repository at this point in the history
  • Loading branch information
@renxida
renxida authored Dec 28, 2023
1 parent 1b40b63 commit d560698
Show file tree
Hide file tree
Showing 2 changed files with 172 additions and 0 deletions.
142 changes: 142 additions & 0 deletions lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -794,6 +794,148 @@ void mlir::torch::onnx_c::populateDefaultDomainQtoZ(
return success();
});

// split with fixed-size parts
// Arguments:
// - input: the tensor to split
// Attributes:
// - axis: the axis along which to split the input
// - num_outputs: the number of outputs to produce
// Outputs:
// - outputs: the produced outputs. Variadic with num_outputs elements.
// Note: torch.aten gives a list of tensors, but ONNX gives a variadic list of
// tensors
// so we need to unpack the list
patterns.onOp(
"Split", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Value self;
int64_t axis;
int64_t num_outputs;
if (binder.tensorOperand(self))
return rewriter.notifyMatchFailure(
binder.op, "Not converting to AtenSplitTensorOp due to input "
"tensor mismatch");
if (binder.s64IntegerAttr(axis, "axis", 0))
return rewriter.notifyMatchFailure(binder.op,
"Failed to get axis attribute");
if (binder.s64IntegerAttr(num_outputs, "num_outputs", 0))
return rewriter.notifyMatchFailure(
binder.op, "Failed to get num_outputs attribute");

auto result0Ty =
binder.op->getResult(0).getType().cast<Torch::ValueTensorType>();
auto selfTy = self.getType().cast<Torch::ValueTensorType>();

int64_t dim = axis;
if (dim < 0)
dim += selfTy.getSizes().size();

// set intermediate shape to the shape of the first result
// if the results are of different shapes
// set the splitted axis to variable shape
llvm::SmallVector<int64_t> intermediateShape(result0Ty.getSizes());
for (auto result : binder.op->getResultTypes()) {
int64_t d = result.cast<Torch::ValueTensorType>().getSizes()[dim];
intermediateShape[dim] = d == intermediateShape[dim] ? d : -1;
}

Value dimValue = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), dim));

Value splitSize = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), num_outputs));

// TODO: Attempting to use the shape expected by the ONNX mlir as ground
// truth. For now just use dynamic shapes.
auto resultOuterType =
Torch::ListType::get(rewriter.getType<Torch::ValueTensorType>(
/*std::optional<llvm::ArrayRef<int64_t>>=*/intermediateShape,
result0Ty.getOptionalDtype()));
Torch::AtenSplitTensorOp new_op =
rewriter.create<Torch::AtenSplitTensorOp>(
binder.getLoc(), resultOuterType, self, splitSize, dimValue);

// the onnx op is variadic with multiple results, but AtenSplitWithSizes
// outputs a list so we need to unpack the list
rewriter.replaceOpWithNewOp<Torch::PrimListUnpackOp>(
binder.op, binder.op->getResults().getType(), new_op.getResult());

return success();
});

// split with variable parts
// Arguments:
// - input: the tensor to split
// - split: the sizes of the splits to be produced
// Attributes:
// - axis: the axis along which to split the input
// - num_outputs: the number of outputs to produce
// Outputs:
// - outputs: the produced outputs. Variadic with num_outputs elements.
// Note: torch.aten gives a list of tensors, but ONNX gives a variadic list of
// tensors
// so we need to unpack the list
patterns.onOp(
"Split", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Value self;
Value split;
int64_t axis;
int64_t num_outputs;
if (binder.tensorOperandAtIndex(self, 0) ||
binder.tensorOperandAtIndex(split, 1))
return rewriter.notifyMatchFailure(
binder.op, "Not converting to AtenSplitWithSizesOp due to input "
"tensor mismatch");
if (binder.s64IntegerAttr(axis, "axis", 0))
return rewriter.notifyMatchFailure(binder.op,
"Failed to get axis attribute");
if (binder.s64IntegerAttr(num_outputs, "num_outputs", 0))
return rewriter.notifyMatchFailure(
binder.op, "Failed to get num_outputs attribute");

auto result0Ty =
binder.op->getResult(0).getType().cast<Torch::ValueTensorType>();
auto selfTy =
cast<Torch::ValueTensorType>(binder.op->getOperand(0).getType());

int64_t dim = axis;
if (dim < 0)
dim += selfTy.getSizes().size();

llvm::SmallVector<int64_t> intermediateShape(result0Ty.getSizes());
for (auto result : binder.op->getResultTypes()) {
int64_t d = result.cast<Torch::ValueTensorType>().getSizes()[dim];
intermediateShape[dim] = d == intermediateShape[dim] ? d : -1;
}

Torch::PrimTolistOp splitToList = rewriter.create<Torch::PrimTolistOp>(
binder.getLoc(),
Torch::ListType::get(rewriter.getType<Torch::IntType>()), split);

Value dimValue = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), dim));

// TODO: Attempting to use the shape expected by the ONNX mlir as ground
// truth. For now just use dynamic shapes.
auto resultOuterType =
Torch::ListType::get(rewriter.getType<Torch::ValueTensorType>(
/*std::optional<llvm::ArrayRef<int64_t>>=*/intermediateShape,
result0Ty.getOptionalDtype()));
Torch::AtenSplitWithSizesOp new_op =
rewriter.create<Torch::AtenSplitWithSizesOp>(
binder.getLoc(), resultOuterType, self,
splitToList.getResult(0), dimValue);

// the onnx op is variadic with multiple results, but AtenSplitWithSizes
// outputs a list so we need to unpack the list
rewriter.replaceOpWithNewOp<Torch::PrimListUnpackOp>(
binder.op, binder.op->getResults().getType(), new_op.getResult());

return success();
});

patterns.onOp("Tan", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Expand Down
30 changes: 30 additions & 0 deletions test/Conversion/TorchOnnxToTorch/simple_ops_q_to_z.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -795,6 +795,36 @@ func.func @test_sinh_example(%arg0: !torch.vtensor<[3],f32>) -> !torch.vtensor<[

// -----

// CHECK-LABEL: func.func @test_split_variable_parts_2d_opset18(
// CHECK-SAME: %[[VAL_INPUT:.*]]: !torch.vtensor<[2,6],f32>,
// CHECK-SAME: %[[VAL_SPLIT:.*]]: !torch.vtensor<[2],si64>
// CHECK: %[[VAL_SPLIT_LIST:.*]] = torch.prim.tolist(%[[VAL_SPLIT]]) : !torch.vtensor<[2],si64> -> !torch.list<int>
// CHECK: %[[VAL_AXIS:.*]] = torch.constant.int 1
// CHECK: %[[VAL_RESULT_LIST:.*]] = torch.aten.split_with_sizes %[[VAL_INPUT]], %[[VAL_SPLIT_LIST]], %[[VAL_AXIS]] : !torch.vtensor<[2,6],f32>, !torch.list<int>, !torch.int -> !torch.list<vtensor<[2,?],f32>>
// CHECK: %[[VAL_VARIADIC_RETURN_VALUE:.*]]:2 = torch.prim.ListUnpack %[[VAL_RESULT_LIST]] : !torch.list<vtensor<[2,?],f32>> -> !torch.vtensor<[2,2],f32>, !torch.vtensor<[2,4],f32>
// CHECK: return %[[VAL_VARIADIC_RETURN_VALUE]]#0, %[[VAL_VARIADIC_RETURN_VALUE]]#1 : !torch.vtensor<[2,2],f32>, !torch.vtensor<[2,4],f32>
func.func @test_split_variable_parts_2d_opset18(%arg0: !torch.vtensor<[2,6],f32>, %arg1: !torch.vtensor<[2],si64>) -> (!torch.vtensor<[2,2],f32>, !torch.vtensor<[2,4],f32>) attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 18 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
%0:2 = torch.operator "onnx.Split"(%arg0, %arg1) {torch.onnx.axis = 1 : si64} : (!torch.vtensor<[2,6],f32>, !torch.vtensor<[2],si64>) -> (!torch.vtensor<[2,2],f32>, !torch.vtensor<[2,4],f32>)
return %0#0, %0#1 : !torch.vtensor<[2,2],f32>, !torch.vtensor<[2,4],f32>
}

// -----

// CHECK-LABEL: func.func @test_split_2d_uneven_split_opset18(
// CHECK-SAME: %[[INPUT_TENSOR:.*]]: !torch.vtensor<[2,8],f32>) -> (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>) attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 18 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK: %[[AXIS:.*]] = torch.constant.int 1
// CHECK: %[[SPLIT_SIZE:.*]] = torch.constant.int 3
// CHECK: %[[SPLIT_RESULT:.*]] = torch.aten.split.Tensor %[[INPUT_TENSOR]], %[[SPLIT_SIZE]], %[[AXIS]] : !torch.vtensor<[2,8],f32>, !torch.int, !torch.int -> !torch.list<vtensor<[2,?],f32>>
// CHECK: %[[UNPACKED_TENSORS:.*]]:3 = torch.prim.ListUnpack %[[SPLIT_RESULT]] : !torch.list<vtensor<[2,?],f32>> -> !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>
// CHECK: return %[[UNPACKED_TENSORS]]#0, %[[UNPACKED_TENSORS]]#1, %[[UNPACKED_TENSORS]]#2 : !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>
// CHECK: }
func.func @test_split_2d_uneven_split_opset18(%arg0: !torch.vtensor<[2,8],f32>) -> (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>) attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 18 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
%0:3 = torch.operator "onnx.Split"(%arg0) {torch.onnx.axis = 1 : si64, torch.onnx.num_outputs = 3 : si64} : (!torch.vtensor<[2,8],f32>) -> (!torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>)
return %0#0, %0#1, %0#2 : !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,3],f32>, !torch.vtensor<[2,2],f32>
}

// -----

// CHECK-LABEL: func.func @test_tan
func.func @test_tan(%arg0: !torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 7 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
// CHECK: %[[TAN:.+]] = torch.aten.tan %arg0
Expand Down

0 comments on commit d560698

Please sign in to comment.