Skip to content

P
Paddle

BaiXuePrincess / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 6c358a7c 编写于 作者: 0 0x45f 提交者： GitHub
浏览文件
操作

[Phi]Move cross OP to phi (#39829) 

* move cross forward OP

* move cross grad op to phi

* move infershape

* refine infershape

* rename ctx

* set dtype and layout in InferMeta

* refine code
上级 6fc5d88a
显示空白变更内容
内联 并排
Showing with 491 addition and 282 deletion
paddle/fluid/operators/cross_op.cc
浏览文件 @ 6c358a7c
...@@ -12,67 +12,23 @@ ...@@ -12,67 +12,23 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/operators/cross_op.h"
#include <memory> #include <memory>
#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/core/infermeta_utils.h"
#include "paddle/phi/infermeta/binary.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
using framework::Tensor; using framework::Tensor;
using framework::DDim; using framework::DDim;
const int kDefaultDim = framework::DDim::kMaxRank;
class CrossOp : public framework::OperatorWithKernel { class CrossOp : public framework::OperatorWithKernel {
public: public:
using framework::OperatorWithKernel::OperatorWithKernel; using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(ctx->HasInput("X"), true,
platform::errors::InvalidArgument(
"Input(X) of CrossOp should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasInput("Y"), true,
platform::errors::InvalidArgument(
"Input(Index) of CrossOp should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasOutput("Out"), true,
platform::errors::InvalidArgument(
"Output(Out) of CrossOp should not be null."));
auto x_dim = ctx->GetInputDim("X");
auto y_dim = ctx->GetInputDim("Y");
auto dim = ctx->Attrs().Get<int>("dim");
bool dims_match = CheckDims(x_dim, y_dim);
PADDLE_ENFORCE_EQ(dims_match, true,
platform::errors::InvalidArgument(
"The 'shape' of Input(X) should be equal to "
"the 'shape' of Input(Y). But received "
"Input(X).dimensions = [%s], "
"Input(Y).dimensions = [%s]",
x_dim, y_dim));
if (dim != kDefaultDim) {
PADDLE_ENFORCE_EQ(
dim < x_dim.size() && dim >= (0 - x_dim.size()), true,
platform::errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
x_dim.size(), x_dim.size() - 1, dim));
if (dim < 0) {
dim += x_dim.size();
}
PADDLE_ENFORCE_EQ(x_dim[dim] == 3 && y_dim[dim] == 3, true,
platform::errors::InvalidArgument(
"Input(X/Y).dims()[dim] should be equal to 3."
"But received Input(X/Y).dims()[dim] = %d.",
x_dim[dim]));
}
ctx->SetOutputDim("Out", x_dim);
auto type = ctx->GetInputsVarType("X")[0];
if (type == framework::proto::VarType::LOD_TENSOR) {
ctx->ShareLoD("X", /*->*/ "Out");
}
}
protected: protected:
framework::OpKernelType GetExpectedKernelType( framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override { const framework::ExecutionContext& ctx) const override {
...@@ -153,17 +109,10 @@ class CrossGradMaker : public framework::SingleGradOpMaker<T> { ...@@ -153,17 +109,10 @@ class CrossGradMaker : public framework::SingleGradOpMaker<T> {
} // namespace paddle } // namespace paddle
namespace ops = paddle::operators; namespace ops = paddle::operators;
DELCARE_INFER_SHAPE_FUNCTOR(cross, CrossInferShapeFunctor,
PT_INFER_META(phi::CrossInferMeta));
REGISTER_OPERATOR(cross, ops::CrossOp, ops::CrossOpMaker, REGISTER_OPERATOR(cross, ops::CrossOp, ops::CrossOpMaker,
ops::CrossGradMaker<paddle::framework::OpDesc>, ops::CrossGradMaker<paddle::framework::OpDesc>,
ops::CrossGradMaker<paddle::imperative::OpBase>); ops::CrossGradMaker<paddle::imperative::OpBase>,
CrossInferShapeFunctor);
REGISTER_OPERATOR(cross_grad, ops::CrossGradOp); REGISTER_OPERATOR(cross_grad, ops::CrossGradOp);
REGISTER_OP_CPU_KERNEL(
cross, ops::CrossKernel<paddle::platform::CPUDeviceContext, float>,
ops::CrossKernel<paddle::platform::CPUDeviceContext, double>,
ops::CrossKernel<paddle::platform::CPUDeviceContext, int>,
ops::CrossKernel<paddle::platform::CPUDeviceContext, int64_t>);
REGISTER_OP_CPU_KERNEL(
cross_grad, ops::CrossGradKernel<paddle::platform::CPUDeviceContext, float>,
ops::CrossGradKernel<paddle::platform::CPUDeviceContext, double>,
ops::CrossGradKernel<paddle::platform::CPUDeviceContext, int>,
ops::CrossGradKernel<paddle::platform::CPUDeviceContext, int64_t>);
paddle/fluid/operators/cross_op.h 已删除 100644 → 0
浏览文件 @ 6fc5d88a
// Copyright (c) 2020 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.
#pragma once
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
using DDim = framework::DDim;
const int kDefaultDim = framework::DDim::kMaxRank;
inline bool CheckDims(const DDim& dims_x, const DDim& dims_y) {
if (dims_x.size() != dims_y.size()) {
return false;
}
for (int i = 0; i < dims_x.size(); i++) {
if (dims_x[i] != dims_y[i]) {
return false;
}
}
return true;
}
template <typename DeviceContext, typename T>
class CrossKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* input_x_var = context.InputVar("X");
auto* input_y_var = context.InputVar("Y");
auto* output_var = context.OutputVar("Out");
auto& input_x = input_x_var->Get<LoDTensor>();
auto& input_y = input_y_var->Get<LoDTensor>();
auto* output = output_var->GetMutable<LoDTensor>();
int dim = context.Attr<int>("dim");
auto input_x_dims = input_x.dims();
auto input_y_dims = input_y.dims();
bool dims_match = CheckDims(input_x_dims, input_y_dims);
PADDLE_ENFORCE_EQ(dims_match, true,
platform::errors::InvalidArgument(
"The 'shape' of Input(X) should be equal to "
"the 'shape' of Input(Y). But received "
"Input(X).dimensions = [%s], "
"Input(Y).dimensions = [%s]",
input_x_dims, input_x_dims));
if (dim != kDefaultDim) {
PADDLE_ENFORCE_EQ(
dim < input_x_dims.size() && dim >= (0 - input_x_dims.size()), true,
platform::errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
input_x_dims.size(), input_x_dims.size() - 1, dim));
if (dim < 0) {
dim += input_x_dims.size();
}
PADDLE_ENFORCE_EQ(
input_x_dims[dim] == 3, true,
platform::errors::InvalidArgument(
"Input(X/Y).dims[dim] must be equal to 3. But received: "
"Input(X/Y).dims[dim] = [%d].",
input_x_dims[dim]));
} else {
for (auto i = 0; i < input_x_dims.size(); i++) {
if (input_x_dims[i] == 3) {
dim = i;
break;
}
}
PADDLE_ENFORCE_EQ(dim == kDefaultDim, false,
platform::errors::InvalidArgument(
"There must be at least one dimension 'd' so that "
"Input(X/Y).dims()[d] is equal to 3. "
"But received: Input(X/Y).dims() == [%s].",
input_x_dims));
}
auto outer_loops = 1;
for (auto i = 0; i < dim; i++) {
outer_loops *= input_x_dims[i];
}
auto slice_size = 1;
for (auto i = dim + 1; i < input_x_dims.size(); i++) {
slice_size *= input_x_dims[i];
}
std::vector<T> input_x_vec, input_y_vec;
framework::TensorToVector(input_x, context.device_context(), &input_x_vec);
framework::TensorToVector(input_y, context.device_context(), &input_y_vec);
std::vector<T> out_vec(output->numel());
output->mutable_data<T>(context.GetPlace());
for (auto i = 0; i < outer_loops; i++) {
for (auto j = 0; j < 3; j++) {
auto dst_pos = (3 * i + j) * slice_size;
auto in_pos1 = (3 * i + ((j + 1) % 3)) * slice_size;
auto in_pos2 = (3 * i + ((j + 2) % 3)) * slice_size;
for (auto k = 0; k < slice_size; k++) {
out_vec[dst_pos + k] =
input_x_vec[in_pos1 + k] * input_y_vec[in_pos2 + k] -
input_x_vec[in_pos2 + k] * input_y_vec[in_pos1 + k];
}
}
}
framework::TensorFromVector(out_vec, context.device_context(), output);
output->Resize(input_x_dims);
}
};
template <typename DeviceContext, typename T>
class CrossGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* input_x_var = context.InputVar("X");
auto* input_y_var = context.InputVar("Y");
auto* input_out_grad_var = context.InputVar(framework::GradVarName("Out"));
auto* output_x_grad_var = context.OutputVar(framework::GradVarName("X"));
auto* output_y_grad_var = context.OutputVar(framework::GradVarName("Y"));
auto& input_x = input_x_var->Get<LoDTensor>();
auto& input_y = input_y_var->Get<LoDTensor>();
auto& input_out_grad = input_out_grad_var->Get<LoDTensor>();
auto* output_x_grad = output_x_grad_var->GetMutable<LoDTensor>();
auto* output_y_grad = output_y_grad_var->GetMutable<LoDTensor>();
int dim = context.Attr<int>("dim");
auto input_x_dims = input_x.dims();
if (dim != kDefaultDim) {
PADDLE_ENFORCE_EQ(
dim < input_x_dims.size() && dim >= (0 - input_x_dims.size()), true,
platform::errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
input_x_dims.size(), input_x_dims.size() - 1, dim));
if (dim < 0) {
dim += input_x_dims.size();
}
PADDLE_ENFORCE_EQ(
input_x_dims[dim] == 3, true,
platform::errors::InvalidArgument(
"Input(X/Y).dims[dim] must be equal to 3. But received: "
"Input(X/Y).dims[dim] = [%d].",
input_x_dims[dim]));
} else {
for (auto i = 0; i < input_x_dims.size(); i++) {
if (input_x_dims[i] == 3) {
dim = i;
break;
}
}
PADDLE_ENFORCE_EQ(dim == kDefaultDim, false,
platform::errors::InvalidArgument(
"There must be at least one dimension 'd' "
"so that Input(X/Y).dims()[d] is equal to 3. "
"But received: Input(X/Y).dims() == [%s].",
input_x_dims));
}
auto outer_loops = 1;
for (auto i = 0; i < dim; i++) {
outer_loops *= input_x_dims[i];
}
auto slice_size = 1;
for (auto i = dim + 1; i < input_x_dims.size(); i++) {
slice_size *= input_x_dims[i];
}
std::vector<T> input_x_vec, input_y_vec, input_dout_vec;
framework::TensorToVector(input_x, context.device_context(), &input_x_vec);
framework::TensorToVector(input_y, context.device_context(), &input_y_vec);
framework::TensorToVector(input_out_grad, context.device_context(),
&input_dout_vec);
std::vector<T> out_dx_vec(output_x_grad->numel());
std::vector<T> out_dy_vec(output_y_grad->numel());
output_x_grad->mutable_data<T>(context.GetPlace());
output_y_grad->mutable_data<T>(context.GetPlace());
for (auto i = 0; i < outer_loops; i++) {
for (auto j = 0; j < 3; j++) {
auto dst_pos = (3 * i + j) * slice_size;
auto in_pos1 = (3 * i + ((j + 1) % 3)) * slice_size;
auto in_pos2 = (3 * i + ((j + 2) % 3)) * slice_size;
for (auto k = 0; k < slice_size; k++) {
out_dx_vec[dst_pos + k] =
input_dout_vec[in_pos2 + k] * input_y_vec[in_pos1 + k] -
input_dout_vec[in_pos1 + k] * input_y_vec[in_pos2 + k];
out_dy_vec[dst_pos + k] =
input_dout_vec[in_pos1 + k] * input_x_vec[in_pos2 + k] -
input_dout_vec[in_pos2 + k] * input_x_vec[in_pos1 + k];
}
}
}
framework::TensorFromVector(out_dx_vec, context.device_context(),
output_x_grad);
framework::TensorFromVector(out_dy_vec, context.device_context(),
output_y_grad);
output_x_grad->Resize(input_x_dims);
output_y_grad->Resize(input_x_dims);
}
};
} // namespace operators
} // namespace paddle
paddle/phi/infermeta/binary.cc
浏览文件 @ 6c358a7c
...@@ -225,6 +225,51 @@ void HuberLossInferMeta(const MetaTensor& input, ...@@ -225,6 +225,51 @@ void HuberLossInferMeta(const MetaTensor& input,
out->share_lod(input); out->share_lod(input);
} }
void CrossInferMeta(const MetaTensor& x,
const MetaTensor& y,
int axis,
MetaTensor* out) {
auto x_dim = x.dims();
auto y_dim = y.dims();
auto dim = axis;
bool dims_match = phi::funcs::CheckDims(x_dim, y_dim);
PADDLE_ENFORCE_EQ(
dims_match,
true,
phi::errors::InvalidArgument("The 'shape' of Input(X) should be equal to "
"the 'shape' of Input(Y). But received "
"Input(X).dimensions = [%s], "
"Input(Y).dimensions = [%s]",
x_dim,
y_dim));
if (dim != DDim::kMaxRank) {
PADDLE_ENFORCE_EQ(
dim < x_dim.size() && dim >= (0 - x_dim.size()),
true,
phi::errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
x_dim.size(),
x_dim.size() - 1,
dim));
if (dim < 0) {
dim += x_dim.size();
}
PADDLE_ENFORCE_EQ(x_dim[dim] == 3 && y_dim[dim] == 3,
true,
phi::errors::InvalidArgument(
"Input(X/Y).dims()[dim] should be equal to 3."
"But received Input(X/Y).dims()[dim] = %d.",
x_dim[dim]));
}
out->set_dims(x_dim);
out->set_dtype(x.dtype());
out->set_layout(x.layout());
out->share_lod(x);
}
void Atan2InferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out) { void Atan2InferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out) {
auto in_dims = x.dims(); auto in_dims = x.dims();
out->set_dims(in_dims); out->set_dims(in_dims);
......
paddle/phi/infermeta/binary.h
浏览文件 @ 6c358a7c
...@@ -53,6 +53,11 @@ void HuberLossInferMeta(const MetaTensor& input_meta, ...@@ -53,6 +53,11 @@ void HuberLossInferMeta(const MetaTensor& input_meta,
MetaTensor* residual, MetaTensor* residual,
MetaConfig config = MetaConfig()); MetaConfig config = MetaConfig());
void CrossInferMeta(const MetaTensor& x,
const MetaTensor& y,
int axis,
MetaTensor* out);
void Atan2InferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out); void Atan2InferMeta(const MetaTensor& x, const MetaTensor& y, MetaTensor* out);
void BCELossInferMeta(const MetaTensor& input, void BCELossInferMeta(const MetaTensor& input,
const MetaTensor& label, const MetaTensor& label,
......
paddle/fluid/operators/cross_op.cu → paddle/phi/kernels/cpu/cross_grad_kernel.cc
浏览文件 @ 6c358a7c
// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved. // Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
// //
// Licensed under the Apache License, Version 2.0 (the "License"); // Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License. // you may not use this file except in compliance with the License.
...@@ -12,17 +12,17 @@ ...@@ -12,17 +12,17 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/operators/cross_op.h" #include "paddle/phi/kernels/cross_grad_kernel.h"
#include "paddle/phi/kernels/impl/cross_grad_kernel_impl.h"
namespace ops = paddle::operators; #include "paddle/phi/backends/cpu/cpu_context.h"
REGISTER_OP_CUDA_KERNEL( #include "paddle/phi/core/kernel_registry.h"
cross, ops::CrossKernel<paddle::platform::CUDADeviceContext, float>,
ops::CrossKernel<paddle::platform::CUDADeviceContext, double>, PD_REGISTER_KERNEL(cross_grad,
ops::CrossKernel<paddle::platform::CUDADeviceContext, int>, CPU,
ops::CrossKernel<paddle::platform::CUDADeviceContext, int64_t>); ALL_LAYOUT,
REGISTER_OP_CUDA_KERNEL( phi::CrossGradKernel,
cross_grad, float,
ops::CrossGradKernel<paddle::platform::CUDADeviceContext, float>, double,
ops::CrossGradKernel<paddle::platform::CUDADeviceContext, double>, int,
ops::CrossGradKernel<paddle::platform::CUDADeviceContext, int>, int64_t) {}
ops::CrossGradKernel<paddle::platform::CUDADeviceContext, int64_t>);
paddle/phi/kernels/cpu/cross_kernel.cc 0 → 100644
浏览文件 @ 6c358a7c
// 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/kernels/cross_kernel.h"
#include "paddle/phi/kernels/impl/cross_kernel_impl.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(
cross, CPU, ALL_LAYOUT, phi::CrossKernel, float, double, int, int64_t) {}
paddle/phi/kernels/cross_grad_kernel.h 0 → 100644
浏览文件 @ 6c358a7c
// 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.
#pragma once
#include "paddle/phi/core/dense_tensor.h"
namespace phi {
template <typename T, typename Context>
void CrossGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& out_grad,
int axis,
DenseTensor* x_grad,
DenseTensor* y_grad);
} // namespace phi
paddle/phi/kernels/cross_kernel.h 0 → 100644
浏览文件 @ 6c358a7c
// 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.
#pragma once
#include "paddle/phi/core/dense_tensor.h"
namespace phi {
template <typename T, typename Context>
void CrossKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
int axis,
DenseTensor* out);
} // namespace phi
paddle/phi/kernels/funcs/common_shape.h
浏览文件 @ 6c358a7c
...@@ -128,5 +128,17 @@ static void GetBroadcastDims(const DDim &in_dims, ...@@ -128,5 +128,17 @@ static void GetBroadcastDims(const DDim &in_dims,
} }
} }
inline bool CheckDims(const DDim &dims_x, const DDim &dims_y) {
if (dims_x.size() != dims_y.size()) {
return false;
}
for (int i = 0; i < dims_x.size(); i++) {
if (dims_x[i] != dims_y[i]) {
return false;
}
}
return true;
}
} // namespace funcs } // namespace funcs
} // namespace phi } // namespace phi
paddle/phi/kernels/gpu/cross_grad_kernel.cu 0 → 100644
浏览文件 @ 6c358a7c
// 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/kernels/cross_grad_kernel.h"
#include "paddle/phi/kernels/impl/cross_grad_kernel_impl.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(cross_grad,
GPU,
ALL_LAYOUT,
phi::CrossGradKernel,
float,
double,
int,
int64_t) {}
paddle/phi/kernels/gpu/cross_kernel.cu 0 → 100644
浏览文件 @ 6c358a7c
// 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/kernels/cross_kernel.h"
#include "paddle/phi/kernels/impl/cross_kernel_impl.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(
cross, GPU, ALL_LAYOUT, phi::CrossKernel, float, double, int, int64_t) {}
paddle/phi/kernels/impl/cross_grad_kernel_impl.h 0 → 100644
浏览文件 @ 6c358a7c
// 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.
#pragma once
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/phi/core/dense_tensor.h"
namespace phi {
template <typename T, typename Context>
void CrossGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& out_grad,
int axis,
DenseTensor* x_grad,
DenseTensor* y_grad) {
auto& input_x = x;
auto& input_y = y;
auto& input_out_grad = out_grad;
auto* output_x_grad = x_grad;
auto* output_y_grad = y_grad;
int dim = axis;
auto input_x_dims = input_x.dims();
if (dim != DDim::kMaxRank) {
PADDLE_ENFORCE_EQ(
dim < input_x_dims.size() && dim >= (0 - input_x_dims.size()),
true,
errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
input_x_dims.size(),
input_x_dims.size() - 1,
dim));
if (dim < 0) {
dim += input_x_dims.size();
}
PADDLE_ENFORCE_EQ(
input_x_dims[dim] == 3,
true,
errors::InvalidArgument(
"Input(X/Y).dims[dim] must be equal to 3. But received: "
"Input(X/Y).dims[dim] = [%d].",
input_x_dims[dim]));
} else {
for (auto i = 0; i < input_x_dims.size(); i++) {
if (input_x_dims[i] == 3) {
dim = i;
break;
}
}
PADDLE_ENFORCE_EQ(
dim == DDim::kMaxRank,
false,
errors::InvalidArgument("There must be at least one dimension 'd' "
"so that Input(X/Y).dims()[d] is equal to 3. "
"But received: Input(X/Y).dims() == [%s].",
input_x_dims));
}
auto outer_loops = 1;
for (auto i = 0; i < dim; i++) {
outer_loops *= input_x_dims[i];
}
auto slice_size = 1;
for (auto i = dim + 1; i < input_x_dims.size(); i++) {
slice_size *= input_x_dims[i];
}
std::vector<T> input_x_vec, input_y_vec, input_dout_vec;
paddle::framework::TensorToVector(input_x, dev_ctx, &input_x_vec);
paddle::framework::TensorToVector(input_y, dev_ctx, &input_y_vec);
paddle::framework::TensorToVector(input_out_grad, dev_ctx, &input_dout_vec);
std::vector<T> out_dx_vec(output_x_grad->numel());
std::vector<T> out_dy_vec(output_y_grad->numel());
dev_ctx.template Alloc<T>(output_x_grad);
dev_ctx.template Alloc<T>(output_y_grad);
for (auto i = 0; i < outer_loops; i++) {
for (auto j = 0; j < 3; j++) {
auto dst_pos = (3 * i + j) * slice_size;
auto in_pos1 = (3 * i + ((j + 1) % 3)) * slice_size;
auto in_pos2 = (3 * i + ((j + 2) % 3)) * slice_size;
for (auto k = 0; k < slice_size; k++) {
out_dx_vec[dst_pos + k] =
input_dout_vec[in_pos2 + k] * input_y_vec[in_pos1 + k] -
input_dout_vec[in_pos1 + k] * input_y_vec[in_pos2 + k];
out_dy_vec[dst_pos + k] =
input_dout_vec[in_pos1 + k] * input_x_vec[in_pos2 + k] -
input_dout_vec[in_pos2 + k] * input_x_vec[in_pos1 + k];
}
}
}
paddle::framework::TensorFromVector(out_dx_vec, dev_ctx, output_x_grad);
paddle::framework::TensorFromVector(out_dy_vec, dev_ctx, output_y_grad);
output_x_grad->Resize(input_x_dims);
output_y_grad->Resize(input_x_dims);
}
} // namespace phi
paddle/phi/kernels/impl/cross_kernel_impl.h 0 → 100644
浏览文件 @ 6c358a7c
// 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.
#pragma once
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/common_shape.h"
namespace phi {
template <typename T, typename Context>
void CrossKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
int axis,
DenseTensor* out) {
auto& input_x = x;
auto& input_y = y;
auto* output = out;
int dim = axis;
auto input_x_dims = input_x.dims();
auto input_y_dims = input_y.dims();
bool dims_match = phi::funcs::CheckDims(input_x_dims, input_y_dims);
PADDLE_ENFORCE_EQ(
dims_match,
true,
phi::errors::InvalidArgument("The 'shape' of Input(X) should be equal to "
"the 'shape' of Input(Y). But received "
"Input(X).dimensions = [%s], "
"Input(Y).dimensions = [%s]",
input_x_dims,
input_x_dims));
if (dim != DDim::kMaxRank) {
PADDLE_ENFORCE_EQ(
dim < input_x_dims.size() && dim >= (0 - input_x_dims.size()),
true,
phi::errors::OutOfRange(
"Attr(dim) is out of range, It's expected "
"to be in range of [-%d, %d]. But received Attr(dim) = %d.",
input_x_dims.size(),
input_x_dims.size() - 1,
dim));
if (dim < 0) {
dim += input_x_dims.size();
}
PADDLE_ENFORCE_EQ(
input_x_dims[dim] == 3,
true,
phi::errors::InvalidArgument(
"Input(X/Y).dims[dim] must be equal to 3. But received: "
"Input(X/Y).dims[dim] = [%d].",
input_x_dims[dim]));
} else {
for (auto i = 0; i < input_x_dims.size(); i++) {
if (input_x_dims[i] == 3) {
dim = i;
break;
}
}
PADDLE_ENFORCE_EQ(dim == DDim::kMaxRank,
false,
phi::errors::InvalidArgument(
"There must be at least one dimension 'd' so that "
"Input(X/Y).dims()[d] is equal to 3. "
"But received: Input(X/Y).dims() == [%s].",
input_x_dims));
}
auto outer_loops = 1;
for (auto i = 0; i < dim; i++) {
outer_loops *= input_x_dims[i];
}
auto slice_size = 1;
for (auto i = dim + 1; i < input_x_dims.size(); i++) {
slice_size *= input_x_dims[i];
}
std::vector<T> input_x_vec, input_y_vec;
paddle::framework::TensorToVector(input_x, dev_ctx, &input_x_vec);
paddle::framework::TensorToVector(input_y, dev_ctx, &input_y_vec);
std::vector<T> out_vec(output->numel());
dev_ctx.template Alloc<T>(output);
for (auto i = 0; i < outer_loops; i++) {
for (auto j = 0; j < 3; j++) {
auto dst_pos = (3 * i + j) * slice_size;
auto in_pos1 = (3 * i + ((j + 1) % 3)) * slice_size;
auto in_pos2 = (3 * i + ((j + 2) % 3)) * slice_size;
for (auto k = 0; k < slice_size; k++) {
out_vec[dst_pos + k] =
input_x_vec[in_pos1 + k] * input_y_vec[in_pos2 + k] -
input_x_vec[in_pos2 + k] * input_y_vec[in_pos1 + k];
}
}
}
paddle::framework::TensorFromVector(out_vec, dev_ctx, output);
output->Resize(input_x_dims);
}
} // namespace phi
paddle/phi/ops/compat/cross_sig.cc 0 → 100644
浏览文件 @ 6c358a7c
// 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 CrossOpArgumentMapping(const ArgumentMappingContext& ctx) {
return KernelSignature("cross", {"X", "Y"}, {"dim"}, {"Out"});
}
KernelSignature CrossGradOpArgumentMapping(const ArgumentMappingContext& ctx) {
return KernelSignature("cross_grad",
{"X", "Y", GradVarName("Out")},
{"dim"},
{GradVarName("X"), GradVarName("Y")});
}
} // namespace phi
PD_REGISTER_ARG_MAPPING_FN(cross, phi::CrossOpArgumentMapping);
PD_REGISTER_ARG_MAPPING_FN(cross_grad, phi::CrossGradOpArgumentMapping);
Markdown is supported
0% 或 .
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 或