sum_mkldnn_op.cc

//   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.

/*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/sum_op.h"
#include "paddle/fluid/platform/mkldnn_reuse.h"

namespace paddle {
namespace framework {
class Tensor;
}  // namespace framework
namespace platform {
class CPUDeviceContext;
class MKLDNNDeviceContext;
}  // namespace platform
}  // namespace paddle

namespace paddle {
namespace operators {

using paddle::platform::CPUDeviceContext;
using paddle::platform::MKLDNNDeviceContext;
using platform::to_void_cast;

template <typename T>
class SumMKLDNNHandler : public platform::MKLDNNHandlerT<T, dnnl::sum> {
 public:
  SumMKLDNNHandler(const MKLDNNDeviceContext& dev_ctx,
                   platform::Place cpu_place,
                   const std::vector<framework::Variable*>& in_vars,
                   framework::LoDTensor* z, const std::string& uniq_name)

      : platform::MKLDNNHandlerT<T, dnnl::sum>(
            dev_ctx, dev_ctx.GetEngine(), cpu_place,
            platform::CreateKey(framework::vectorize(z->dims()), uniq_name)),
        num_inputs_(0) {
    for (size_t i = 0; i < in_vars.size(); i++) {
      srcs_suffix_.push_back(std::string("-") + std::to_string(i));
    }

    if (!this->isCached()) {
      auto dst_tz = framework::vectorize<int64_t>(z->dims());
      auto src_tz = dst_tz;

      std::vector<mkldnn::memory::desc> srcs_md;
      for (size_t i = 0; i < in_vars.size(); i++) {
        auto& input_it = in_vars[i]->Get<framework::LoDTensor>();
        if (input_it.numel() == 0) {
          continue;
        }
        MKLDNNMemoryFormat input_format = input_it.format();
        srcs_md.push_back(mkldnn::memory::desc(
            src_tz, platform::MKLDNNGetDataType<T>(), input_format));
        ++num_inputs_;
      }
      std::vector<float> scales(num_inputs_, 1.0);

      auto dst_md = mkldnn::memory::desc(
          dst_tz, platform::MKLDNNGetDataType<T>(), MKLDNNMemoryFormat::any);

      this->AcquireForwardPrimitiveDescriptor(dst_md, scales, srcs_md);
    }
  }

  // (jczaja) sum oneDNN prim is not having .desc attribute so
  // we cannot use base AcquireForwardPrimitiveDescriptor
  void AcquireForwardPrimitiveDescriptor(
      const mkldnn::memory::desc& dst_md, const std::vector<float>& scales,
      const std::vector<mkldnn::memory::desc>& srcs_md) {
    // Sum op does not have backward so no passing from FWD to BWD is needed
    const std::string key_pd = this->key_ + "@fwd_pd";
    this->fwd_pd_ = std::static_pointer_cast<dnnl::sum::primitive_desc>(
        this->dev_ctx_.GetBlob(key_pd));
    if (this->fwd_pd_ == nullptr) {
      this->fwd_pd_.reset(new dnnl::sum::primitive_desc(dst_md, scales, srcs_md,
                                                        this->engine_));
      this->dev_ctx_.SetBlob(key_pd, this->fwd_pd_);
    }
  }

  std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
      const framework::Tensor& input, int i) {
    const T* input_data = input.data<T>();
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->src_desc(i),
                                            to_void_cast<T>(input_data),
                                            "@src_mem_p" + srcs_suffix_[i]);
  }

  using platform::MKLDNNHandlerT<T, dnnl::sum>::AcquireDstMemory;

  std::shared_ptr<mkldnn::memory> AcquireDstMemory(void) {
    return this->AcquireMemoryFromPrimitive(this->fwd_pd_->dst_desc(),
                                            "@dst_mem_p");
  }

  inline int GetNumInputs(void) { return num_inputs_; }

 protected:
  // isCached need to be overloaded as base one works on key_common
  bool isCached() {
    const std::string key_pd = this->key_ + "@fwd_pd";
    this->fwd_pd_ = std::static_pointer_cast<dnnl::sum::primitive_desc>(
        this->dev_ctx_.GetBlob(key_pd));

    const std::string key_p = this->key_ + "@fwd_p";
    return (this->dev_ctx_.GetBlob(key_p) != nullptr);
  }

 private:
  int num_inputs_;
  std::vector<std::string> srcs_suffix_;
};

template <typename T>
class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true,
                      paddle::platform::errors::PreconditionNotMet(
                          "Operator DNNL Sum must use CPUPlace"));
    auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    auto in_vars = ctx.MultiInputVar("X");

    PADDLE_ENFORCE_NE(in_vars.empty(), true, platform::errors::InvalidArgument(
                                                 "Input variable is empty."));
    auto& input0 = in_vars[0]->Get<LoDTensor>();
    LoDTensor* output = ctx.Output<LoDTensor>("Out");

    bool in_place = (input0.numel() > 0) && input0.IsSharedBufferWith(*output);

    SumMKLDNNHandler<T> handler(dev_ctx, ctx.GetPlace(), in_vars, output,
                                ctx.OutputName("Out"));

    // Create list of SRC MEMs
    std::vector<std::shared_ptr<mkldnn::memory>> srcs_mem;
    srcs_mem.reserve(handler.GetNumInputs());
    int input_index = 0;
    for (size_t i = 0; i < in_vars.size(); i++) {
      auto& input_it = in_vars[i]->Get<framework::LoDTensor>();
      if (input_it.numel() == 0) {
        continue;
      }
      srcs_mem.push_back(handler.AcquireSrcMemory(input_it, input_index));
      ++input_index;
    }

    auto dst_mem = in_place ? handler.AcquireDstMemory()
                            : handler.AcquireDstMemory(output);

    auto sum_p = handler.AcquireForwardPrimitive();

    std::unordered_map<int, mkldnn::memory> args;
    for (size_t i = 0; i < srcs_mem.size(); ++i) {
      args.insert({MKLDNN_ARG_MULTIPLE_SRC + i, *(srcs_mem[i])});
    }
    args.insert({MKLDNN_ARG_DST, *dst_mem});

    mkldnn::stream astream(dev_ctx.GetEngine());
    sum_p->execute(astream, args);
    astream.wait();

    // For in-place execution which sum does not have we need to fake it
    // so from oneDNN dst memory we reorder data into input
    if (in_place) {
      const std::string reorder_key = platform::CreateKey(
          framework::vectorize(output->dims()), ctx.OutputName("Out") + "-I");

      auto& in_out = in_vars[0]->Get<framework::LoDTensor>();
      auto output_tz = framework::vectorize<int64_t>(output->dims());
      platform::ReorderMKLDNNHandler reorder_handler(
          output_tz, output->type(), framework::ToMKLDNNDataType(in_out.type()),
          dev_ctx, dev_ctx.GetEngine(), reorder_key);

      auto target_mem = reorder_handler.AcquireDstMemory(
          output, in_out.format(), ctx.GetPlace());

      auto reorder_p = reorder_handler.AcquireReorder(target_mem, dst_mem);
      {
        platform::RecordEvent record_reorder("int_reorder",
                                             platform::EventRole::kUniqueOp);
        reorder_p->execute(astream, *dst_mem, *target_mem);
        astream.wait();
      }
    }
    output->set_layout(framework::DataLayout::kMKLDNN);
    output->set_format(platform::GetMKLDNNFormat(*dst_mem));
  }
};

}  // namespace operators
}  // namespace paddle

REGISTER_OP_KERNEL(
    sum, MKLDNN, ::paddle::platform::CPUPlace,
    paddle::operators::SumMKLDNNOpKernel<paddle::platform::bfloat16>,
    paddle::operators::SumMKLDNNOpKernel<float>);