未验证 提交 1899505d 编写于 作者: L LoneRanger 提交者: GitHub

fix the static op generation for einsum (#54723)

* fix the static op generation for group_norm

* fix bug

* fix bug

* Update op_compat.yaml
上级 cb4467f0
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include <string>
#include <vector>
#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/phi/core/ddim.h"
#include "paddle/phi/infermeta/unary.h"
namespace paddle {
namespace operators {
class EinsumOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
};
class EinsumOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Operands", "(TensorList), The input tensor of einsum op.")
.AsDuplicable();
AddOutput("Out", "(Tensor), The output tensor of einsum op.");
AddOutput(
"InnerCache",
"(Tensor), The cache of the forward transpose tensors: tA and tB.")
.AsDuplicable()
.AsExtra()
.AsIntermediate();
AddOutput("XShape", "(Tensor), The cache of the x_shape of: A and B.")
.AsDuplicable()
.AsExtra()
.AsIntermediate();
AddAttr<std::string>("equation",
"(string) A einsum equation. such as `ij,jk->ik`"
"There must have `->` and the number of operands in "
"equation must equals the `Operands` length.");
AddComment(R"DOC(
Einsum Operator.
This operator is used to perform einsum operation for given operands and equation.
)DOC");
}
};
class EinsumGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
auto x_name = "Operands";
auto x_grad_name = framework::GradVarName(x_name);
ctx->SetOutputsDim(x_grad_name, ctx->GetInputsDim("Operands"));
ctx->ShareAllLoD("Operands", x_grad_name);
}
protected:
phi::KernelKey GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto dtype = OperatorWithKernel::IndicateVarDataType(
ctx, framework::GradVarName("Out"));
return phi::KernelKey(dtype, ctx.GetPlace());
}
};
template <typename T>
class EinsumGradMaker : public framework::SingleGradOpMaker<T> {
public:
using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
void Apply(GradOpPtr<T> retv) const override {
retv->SetType("einsum_grad");
if (this->HasOutput("InnerCache")) {
retv->SetInput("InnerCache", this->Output("InnerCache"));
}
if (this->HasOutput("XShape")) {
// add if for compatibility.
retv->SetInput("Operands", this->Output("XShape")); // for memory save.
} else {
retv->SetInput("Operands", this->Input("Operands"));
}
retv->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
retv->SetAttrMap(this->Attrs());
retv->SetOutput(framework::GradVarName("Operands"),
this->InputGrad("Operands", false));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
DECLARE_INFER_SHAPE_FUNCTOR(einsum,
EinsumInferShapeFunctor,
PD_INFER_META(phi::EinsumRawInferMeta));
REGISTER_OPERATOR(einsum,
ops::EinsumOp,
ops::EinsumOpMaker,
EinsumInferShapeFunctor,
ops::EinsumGradMaker<paddle::framework::OpDesc>,
ops::EinsumGradMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(einsum_grad, ops::EinsumGradOp);
......@@ -844,6 +844,16 @@
attrs :
uplo : UPLO
- op : einsum
backward : einsum_grad
inputs :
x : Operands
outputs:
{out : Out, inner_cache: InnerCache, xshape : XShape}
drop_empty_grad: [x_grad]
extra:
outputs: [inner_cache, xshape]
- op : elementwise_pow
backward : elementwise_pow_grad
extra :
......
......@@ -74,6 +74,17 @@
data_type : x
optional : bias
- backward_op : einsum_grad
forward : einsum (Tensor[] x, str equation) -> Tensor(out), Tensor[](inner_cache), Tensor[](x_shape)
args : (Tensor[] x_shape, Tensor[] inner_cache, Tensor out_grad, str equation)
output : Tensor[](x_grad){x_shape.size()}
infer_meta :
func : UnchangedMultiInferMeta
param : [x_shape]
kernel :
func : einsum_grad
data_type : out_grad
- backward_op : embedding_grad
forward : embedding (Tensor x, Tensor weight, int64_t padding_idx=-1) -> Tensor(out)
args : (Tensor x, Tensor weight, Tensor out_grad, int64_t padding_idx=-1)
......
......@@ -156,6 +156,17 @@
optional : bias
backward : depthwise_conv2d_transpose_grad
- op : einsum
args : (Tensor[] x, str equation)
output : Tensor(out), Tensor[](inner_cache){x.size()}, Tensor[](xshape){x.size()}
infer_meta :
func : EinsumRawInferMeta
param : [x, equation]
kernel :
func : einsum
backward : einsum_grad
intermediate : inner_cache, xshape
- op : embedding
args : (Tensor x, Tensor weight, int64_t padding_idx=-1)
output : Tensor
......
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/phi/core/compat/op_utils.h"
namespace phi {
KernelSignature EinsumOpArgumentMapping(
const ArgumentMappingContext& ctx UNUSED) {
return KernelSignature(
"einsum", {"Operands"}, {"equation"}, {"Out", "InnerCache", "XShape"});
}
KernelSignature EinsumGradOpArgumentMapping(
const ArgumentMappingContext& ctx UNUSED) {
return KernelSignature("einsum_grad",
{"Operands", "InnerCache", "Out@GRAD"},
{"equation"},
{"Operands@GRAD"});
}
} // namespace phi
PD_REGISTER_ARG_MAPPING_FN(einsum, phi::EinsumOpArgumentMapping);
PD_REGISTER_ARG_MAPPING_FN(einsum_grad, phi::EinsumGradOpArgumentMapping);
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册