Merge pull request #7414 from wanghaoshuang/warpctc
Adapt warpctc grad op for gradient checkingadd_depthwiseConv_op_gpu
whs
7 years ago
committed by
GitHub
commit
448fee3db4
@ -0,0 +1,46 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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/operators/math/sequence_scale.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace math {
|
||||
|
||||
template <typename T>
|
||||
class ScaleLoDTensorFunctor<platform::CPUDeviceContext, T> {
|
||||
public:
|
||||
void operator()(const platform::CPUDeviceContext& context,
|
||||
framework::LoDTensor& seq, const T* scales) {
|
||||
const size_t level = 0;
|
||||
auto lod = seq.lod();
|
||||
const size_t num_seq = lod[level].size() - 1;
|
||||
size_t seq_width = seq.dims()[1];
|
||||
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
|
||||
|
||||
T* seq_data = seq.mutable_data<T>(context.GetPlace());
|
||||
for (size_t i = 0; i < num_seq; ++i) {
|
||||
for (size_t j = lod[level][i] * seq_width;
|
||||
j < lod[level][i + 1] * seq_width; ++j) {
|
||||
seq_data[j] *= scales[i];
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template class ScaleLoDTensorFunctor<platform::CPUDeviceContext, float>;
|
||||
|
||||
} // namespace math
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,57 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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/operators/math/sequence_scale.h"
|
||||
#include "paddle/platform/cuda_helper.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace math {
|
||||
|
||||
using platform::PADDLE_CUDA_NUM_THREADS;
|
||||
|
||||
template <typename T, int BlockSize>
|
||||
__global__ void SequenceScaleKernel(T* seq, size_t* lod, const T* scales,
|
||||
const size_t seq_width) {
|
||||
for (int i = threadIdx.x;
|
||||
i < (lod[blockIdx.x + 1] - lod[blockIdx.x]) * seq_width;
|
||||
i += BlockSize) {
|
||||
int idx = lod[blockIdx.x] * seq_width + i;
|
||||
seq[idx] *= scales[blockIdx.x];
|
||||
}
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
class ScaleLoDTensorFunctor<platform::CUDADeviceContext, T> {
|
||||
public:
|
||||
void operator()(const platform::CUDADeviceContext& context,
|
||||
framework::LoDTensor& seq, const T* scales) {
|
||||
const size_t level = 0;
|
||||
auto lod = seq.lod();
|
||||
const size_t num_seq = lod[level].size() - 1;
|
||||
const size_t seq_width = seq.numel() / seq.dims()[0];
|
||||
framework::LoD abs_offset_lod = framework::ToAbsOffset(lod);
|
||||
T* seq_data = seq.mutable_data<T>(context.GetPlace());
|
||||
|
||||
SequenceScaleKernel<T, PADDLE_CUDA_NUM_THREADS><<<
|
||||
num_seq, PADDLE_CUDA_NUM_THREADS, 0, context.stream()>>>(
|
||||
seq_data, abs_offset_lod[level].data(), scales, seq_width);
|
||||
}
|
||||
};
|
||||
|
||||
template class ScaleLoDTensorFunctor<platform::CUDADeviceContext, float>;
|
||||
|
||||
} // namespace math
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,55 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
|
||||
|
||||
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/framework/lod_tensor.h"
|
||||
#include "paddle/platform/device_context.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace math {
|
||||
|
||||
/*
|
||||
* \brief Scale a sequence.
|
||||
*
|
||||
* All sequences will be padded to the same length and stored in a transposed
|
||||
* shape.
|
||||
* Example:
|
||||
* Given:
|
||||
* seq = (s0, s0, s0, s0; s1, s1; s2, s2, s2; s3)
|
||||
* scales = (2, 3, 4, 5)
|
||||
* then:
|
||||
* result = (2*s0, 2*s0, 2*s0, 2*s0; 3*s1, 3*s1; 4*s2, 4*s2, 4*s2; 5*s3)
|
||||
|
||||
*
|
||||
* \param context Device context of this functor.
|
||||
* \param seq LoDTensor which is stored in sequence format, the shape
|
||||
* is [total_sequence_length, sequence_width] where
|
||||
* total_sequence_length is the sum of all sequences'
|
||||
* length.
|
||||
* \param scales Array<T>. The i-th sequence will be scaled by scales[i].
|
||||
* \param num_seq Number of sequence
|
||||
*
|
||||
*/
|
||||
template <typename DeviceContext, typename T>
|
||||
class ScaleLoDTensorFunctor {
|
||||
public:
|
||||
void operator()(const DeviceContext& context, framework::LoDTensor& seq,
|
||||
const T* scales);
|
||||
};
|
||||
|
||||
} // namespace math
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
Loading…
Reference in new issue