sgd_op.cu

/* 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 <algorithm>
#include "paddle/fluid/operators/optimizers/sgd_op.h"
#include "paddle/fluid/platform/cuda_primitives.h"

namespace paddle {
namespace operators {

namespace {

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;
  }
}

template <typename T>
__global__ void SparseSGDFunctorKernel(const T* selected_rows,
                                       const int64_t* rows,
                                       const T* learning_rate, T* tensor_out,
                                       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]);
    }
  }
}
}  // namespace

template <typename T>
class SGDOpCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    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());

    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();
      auto& in_rows = grad->rows();

      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>();

      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()>>>(
          in_data, in_rows.CUDAData(ctx.GetPlace()), learning_rate->data<T>(),
          out_data, in_row_numel, in_rows.size());

    } else {
      PADDLE_THROW("Unsupported Variable Type of Grad");
    }
  }
};
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(sgd, ops::SGDOpCUDAKernel<float>,
                        ops::SGDOpCUDAKernel<double>);