sgd_op.cu 4.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
L
liaogang 已提交
2

L
Luo Tao 已提交
3 4 5
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
L
liaogang 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
L
liaogang 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
L
liaogang 已提交
14

C
chengduo 已提交
15
#include <algorithm>
Y
Yi Wang 已提交
16
#include "paddle/fluid/operators/sgd_op.h"
D
dzhwinter 已提交
17
#include "paddle/fluid/platform/cuda_primitives.h"
Q
qijun 已提交
18 19 20 21 22

namespace paddle {
namespace operators {

namespace {
C
chengduoZH 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35

template <typename T>
__global__ void SGDKernel(const T* g, const T* p, const T* learning_rate,
                          const int num, T* p_out) {
  T lr = learning_rate[0];
  int grid_size = blockDim.x * gridDim.x;
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; i += grid_size) {
    T g_data = g[i];
    T p_data = p[i];
    p_out[i] = p_data - lr * g_data;
  }
}

C
chengduo 已提交
36
template <typename T>
Q
qijun 已提交
37 38 39
__global__ void SparseSGDFunctorKernel(const T* selected_rows,
                                       const int64_t* rows,
                                       const T* learning_rate, T* tensor_out,
C
chengduo 已提交
40 41 42 43 44 45 46 47 48 49 50
                                       int64_t row_numel, int64_t limit) {
  for (int64_t i = blockIdx.x; i < limit; i += gridDim.x) {
    const T* selected_rows_ptr = selected_rows + i * row_numel;
    T* tensor_out_ptr = tensor_out + rows[i] * row_numel;
    for (int64_t index = threadIdx.x; index < row_numel; index += blockDim.x) {
      // Since index in rows of SelectedRows can be duplicate, we have to use
      // Atomic Operation to avoid concurrent write error.
      paddle::platform::CudaAtomicAdd(
          tensor_out_ptr + index,
          -1.0 * learning_rate[0] * selected_rows_ptr[index]);
    }
Q
qijun 已提交
51 52 53 54 55
  }
}
}  // namespace

template <typename T>
C
chengduoZH 已提交
56 57 58
class SGDOpCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
C
chengduo 已提交
59 60 61 62 63 64
    const auto* param_var = ctx.InputVar("Param");
    PADDLE_ENFORCE(param_var->IsType<framework::LoDTensor>(),
                   "The Var(%s)'s type should be LoDTensor, "
                   "but the received is %s",
                   ctx.Inputs("Param").front(), param_var->Type().name());

C
chengduoZH 已提交
65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96
    auto* param = ctx.Input<framework::Tensor>("Param");
    auto* param_out = ctx.Output<framework::Tensor>("ParamOut");
    auto* learning_rate = ctx.Input<framework::Tensor>("LearningRate");

    auto* grad_var = ctx.InputVar("Grad");
    // Actually, all tensors are LoDTensor except SelectedRows.
    if (grad_var->IsType<framework::LoDTensor>()) {
      param_out->mutable_data<T>(ctx.GetPlace());
      auto* grad = ctx.Input<framework::Tensor>("Grad");
      auto* grad_data = grad->data<T>();
      auto* param_data = param->data<T>();
      auto* param_out_data = param_out->data<T>();

      int block = 512;
      int grid = (param->numel() + block - 1) / block;

      SGDKernel<T><<<grid, block, 0, ctx.cuda_device_context().stream()>>>(
          grad_data, param_data, learning_rate->data<T>(), param->numel(),
          param_out_data);

    } else if (grad_var->IsType<framework::SelectedRows>()) {
      // TODO(qijun): In Sparse SGD operator, in-place update is enforced.
      // This manual optimization brings difficulty to track data dependency.
      // It's better to find a more elegant solution.
      PADDLE_ENFORCE_EQ(param, param_out);
      auto* grad = ctx.Input<framework::SelectedRows>("Grad");

      auto in_height = grad->height();
      auto out_dims = param_out->dims();
      PADDLE_ENFORCE_EQ(in_height, out_dims[0]);

      auto& in_value = grad->value();
Y
Yu Yang 已提交
97
      auto& in_rows = grad->rows();
C
chengduoZH 已提交
98 99 100 101 102 103 104

      int64_t in_row_numel = in_value.numel() / in_rows.size();
      PADDLE_ENFORCE_EQ(in_row_numel, param_out->numel() / in_height);

      auto* in_data = in_value.data<T>();
      auto* out_data = param_out->data<T>();

C
chengduo 已提交
105 106 107 108 109 110 111
      const int kThreadsPerBlock = 256;
      int thread_x = kThreadsPerBlock;
      int max_threads = ctx.cuda_device_context().GetMaxPhysicalThreadCount();
      int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);

      SparseSGDFunctorKernel<<<max_blocks, thread_x, 0,
                               ctx.cuda_device_context().stream()>>>(
Y
Yu Yang 已提交
112
          in_data, in_rows.CUDAData(ctx.GetPlace()), learning_rate->data<T>(),
C
chengduo 已提交
113
          out_data, in_row_numel, in_rows.size());
C
chengduoZH 已提交
114 115 116 117

    } else {
      PADDLE_THROW("Unsupported Variable Type of Grad");
    }
Q
qijun 已提交
118 119 120 121
  }
};
}  // namespace operators
}  // namespace paddle
Q
Qiao Longfei 已提交
122

D
dongzhihong 已提交
123
namespace ops = paddle::operators;
C
chengduoZH 已提交
124 125
REGISTER_OP_CUDA_KERNEL(sgd, ops::SGDOpCUDAKernel<float>,
                        ops::SGDOpCUDAKernel<double>);