[Speed]implement cudnn sequence softmax cudnn (#8978)
* "add softmax cudnn functor support" * "add testing" * "refine cmakelist" * "sequence softmax forward speed up" * "add softmax grad" * "fix sequence softmax test" * "add double precision' * "fix softmax test" * "add softmax cudnn support" * "fix softmax cudnn test" * "add softmax to nn.py" * "fix compile bug" * "refine cmakelist" * "fix ci" * "fix based on comment" * "fix based on comments" * "fix ci"shanyi15-patch-2
dzhwinter
7 years ago
committed by
GitHub
parent
9b9f3f09b3
commit
128adf53cb
@ -0,0 +1,105 @@
|
|||||||
|
/* Copyright (c) 2016 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/fluid/framework/op_registry.h"
|
||||||
|
#include "paddle/fluid/operators/math/math_function.h"
|
||||||
|
#include "paddle/fluid/operators/math/softmax.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
using LoDTensor = framework::LoDTensor;
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class SequenceSoftmaxCUDNNKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||||
|
auto* x = ctx.Input<LoDTensor>("X");
|
||||||
|
auto* out = ctx.Output<LoDTensor>("Out");
|
||||||
|
|
||||||
|
auto lod = x->lod();
|
||||||
|
auto dims = x->dims();
|
||||||
|
|
||||||
|
const size_t level = lod.size() - 1;
|
||||||
|
PADDLE_ENFORCE_EQ(dims[0], static_cast<int64_t>(lod[level].back()),
|
||||||
|
"The first dimension of Input(X) should be equal to the "
|
||||||
|
"sum of all sequences' lengths.");
|
||||||
|
PADDLE_ENFORCE_EQ(dims[0], x->numel(),
|
||||||
|
"The width of each timestep in Input(X) of "
|
||||||
|
"SequenceSoftmaxOp should be 1.");
|
||||||
|
|
||||||
|
out->mutable_data<T>(ctx.GetPlace());
|
||||||
|
for (int i = 0; i < static_cast<int>(lod[level].size()) - 1; ++i) {
|
||||||
|
int start_pos = static_cast<int>(lod[level][i]);
|
||||||
|
int end_pos = static_cast<int>(lod[level][i + 1]);
|
||||||
|
Tensor x_i = x->Slice(start_pos, end_pos);
|
||||||
|
Tensor out_i = out->Slice(start_pos, end_pos);
|
||||||
|
|
||||||
|
// Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos)
|
||||||
|
framework::DDim dims_i =
|
||||||
|
// framework::make_ddim({1UL, end_pos - start_pos, 1UL, 1UL});
|
||||||
|
framework::make_ddim({1UL, end_pos - start_pos});
|
||||||
|
x_i.Resize(dims_i);
|
||||||
|
out_i.Resize(dims_i);
|
||||||
|
math::SoftmaxCUDNNFunctor<T>()(
|
||||||
|
ctx.template device_context<platform::CUDADeviceContext>(), &x_i,
|
||||||
|
&out_i);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class SequenceSoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||||
|
auto* out = ctx.Input<LoDTensor>("Out");
|
||||||
|
auto* out_grad = ctx.Input<LoDTensor>(framework::GradVarName("Out"));
|
||||||
|
auto* x = ctx.Input<LoDTensor>("X");
|
||||||
|
auto* x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
|
||||||
|
|
||||||
|
auto lod = x->lod();
|
||||||
|
const size_t level = lod.size() - 1;
|
||||||
|
|
||||||
|
x_grad->mutable_data<T>(ctx.GetPlace());
|
||||||
|
for (int i = 0; i < static_cast<int>(lod[level].size()) - 1; ++i) {
|
||||||
|
int start_pos = static_cast<int>(lod[level][i]);
|
||||||
|
int end_pos = static_cast<int>(lod[level][i + 1]);
|
||||||
|
|
||||||
|
Tensor out_i = out->Slice(start_pos, end_pos);
|
||||||
|
Tensor out_grad_i = out_grad->Slice(start_pos, end_pos);
|
||||||
|
Tensor x_grad_i = x_grad->Slice(start_pos, end_pos);
|
||||||
|
|
||||||
|
// Reshape from (end_pos - start_pos) x 1UL to 1UL x (end_pos - start_pos)
|
||||||
|
framework::DDim dims_i = framework::make_ddim({1UL, end_pos - start_pos});
|
||||||
|
out_i.Resize(dims_i);
|
||||||
|
out_grad_i.Resize(dims_i);
|
||||||
|
x_grad_i.Resize(dims_i);
|
||||||
|
math::SoftmaxGradCUDNNFunctor<T>()(
|
||||||
|
ctx.template device_context<platform::CUDADeviceContext>(), &out_i,
|
||||||
|
&out_grad_i, &x_grad_i);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OP_KERNEL(sequence_softmax, CUDNN, ::paddle::platform::CUDAPlace,
|
||||||
|
ops::SequenceSoftmaxCUDNNKernel<float>,
|
||||||
|
ops::SequenceSoftmaxCUDNNKernel<double>)
|
||||||
|
REGISTER_OP_KERNEL(sequence_softmax_grad, CUDNN, ::paddle::platform::CUDAPlace,
|
||||||
|
ops::SequenceSoftmaxGradCUDNNKernel<float>,
|
||||||
|
ops::SequenceSoftmaxGradCUDNNKernel<double>)
|
||||||
@ -0,0 +1,62 @@
|
|||||||
|
/* Copyright (c) 2018 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/fluid/operators/math/softmax.h"
|
||||||
|
#include "paddle/fluid/framework/op_registry.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace operators {
|
||||||
|
|
||||||
|
using Tensor = framework::Tensor;
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class SoftmaxCUDNNKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& context) const override {
|
||||||
|
auto* X = context.Input<Tensor>("X");
|
||||||
|
auto* Out = context.Output<Tensor>("Out");
|
||||||
|
|
||||||
|
// allocate memory on device.
|
||||||
|
Out->mutable_data<T>(context.GetPlace());
|
||||||
|
|
||||||
|
math::SoftmaxCUDNNFunctor<T>()(
|
||||||
|
context.template device_context<platform::CUDADeviceContext>(), X, Out);
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
class SoftmaxGradCUDNNKernel : public framework::OpKernel<T> {
|
||||||
|
public:
|
||||||
|
void Compute(const framework::ExecutionContext& context) const override {
|
||||||
|
auto* Out = context.Input<Tensor>("Out");
|
||||||
|
auto* dOut = context.Input<Tensor>(framework::GradVarName("Out"));
|
||||||
|
auto* dX = context.Output<Tensor>(framework::GradVarName("X"));
|
||||||
|
|
||||||
|
// allocate memory on device.
|
||||||
|
dX->mutable_data<T>(context.GetPlace());
|
||||||
|
|
||||||
|
math::SoftmaxGradCUDNNFunctor<T>()(
|
||||||
|
context.template device_context<platform::CUDADeviceContext>(), Out,
|
||||||
|
dOut, dX);
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace operators
|
||||||
|
} // namespace paddle
|
||||||
|
|
||||||
|
namespace ops = paddle::operators;
|
||||||
|
REGISTER_OP_KERNEL(softmax, CUDNN, ::paddle::platform::CUDAPlace,
|
||||||
|
ops::SoftmaxCUDNNKernel<float>);
|
||||||
|
REGISTER_OP_KERNEL(softmax_grad, CUDNN, ::paddle::platform::CUDAPlace,
|
||||||
|
ops::SoftmaxGradCUDNNKernel<float>);
|
||||||
Loading…
Reference in new issue