matmul_mkldnn_op.cc

/* Copyright (c) 2020 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/framework/tensor.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/mkldnn_helper.h"

namespace paddle {
namespace platform {
class MKLDNNDeviceContext;
struct CPUPlace;
}  // namespace platform
}  // namespace paddle

namespace paddle {
namespace operators {

using dnnl::memory;
using dnnl::primitive;
using framework::DataLayout;
using framework::ExecutionContext;
using platform::GetMKLDNNFormat;
using platform::MKLDNNDeviceContext;
using platform::MKLDNNGetDataType;
using platform::to_void_cast;
using Tensor = framework::Tensor;

template <typename T>
constexpr bool IsInt8() {
  return std::is_same<T, int8_t>::value || std::is_same<T, uint8_t>::value;
}

template <typename T>
constexpr bool IsBfloat16() {
  return std::is_same<T, paddle::platform::bfloat16>::value;
}

// Get row matrix shape from a vector shape. If the rank of x_dim > 1, the
// original x_dim is returned.
static framework::DDim RowMatrixDimsFromVector(const framework::DDim& x_dim) {
  return x_dim.size() > 1 ? x_dim : framework::make_ddim({1, x_dim[0]});
}

// Get column matrix shape from a vector shape. If the ran of y_dim > 1, the
// original y_dim is returned.
static framework::DDim ColumnMatrixDimsFromVector(
    const framework::DDim& y_dim) {
  return y_dim.size() > 1 ? y_dim : framework::make_ddim({y_dim[0], 1});
}

template <typename XT, typename YT, typename OT>
class MatMulFactory {
 public:
  void CreateAndExecute(const ExecutionContext& ctx) {
    SetDNNLEngine(ctx);
    if (IsInitialized()) {
      UpdateDataPointers(ctx);
      Execute();
      SetOutputFormat(ctx);
      return;
    }
    CreateMemories(ctx);
    CreatePrimitive(ctx);
    Execute();
    SetOutputFormat(ctx);
    SetInitialized();
  }

 private:
  struct MatMulDims {
    const memory::dims x_dims, y_dims, out_dims, x_strides, y_strides,
        out_strides;
  };

  void SetDNNLEngine(const ExecutionContext& ctx) {
    auto& dev_ctx =
        ctx.template device_context<platform::MKLDNNDeviceContext>();
    engine_ = dev_ctx.GetEngine();
  }

  template <typename T>
  dnnl::memory CreateMemory(const memory::dims& dims,
                            const memory::dims& strides, const T* data) {
    auto md = memory::desc(dims, MKLDNNGetDataType<T>(), strides);
    return dnnl::memory(md, engine_, to_void_cast(data));
  }

  std::vector<int64_t> Transpose(const std::vector<int64_t>& x,
                                 const std::vector<int>& axis) {
    size_t in_rank = x.size();
    size_t axis_size = axis.size();

    auto axis_set = std::set<int>(axis.begin(), axis.end());
    PADDLE_ENFORCE_EQ(axis_set.size(), axis_size,
                      platform::errors::InvalidArgument(
                          "In an axis array, elements must be unique."));

    PADDLE_ENFORCE_EQ(
        in_rank, axis_size,
        platform::errors::InvalidArgument("The input dimension's size "
                                          "should be equal to the axis's size. "
                                          "But received dimension is %d, "
                                          "axis's size is %d",
                                          in_rank, axis_size));

    PADDLE_ENFORCE_LT(*std::max_element(axis.begin(), axis.end()), axis_size,
                      platform::errors::InvalidArgument(
                          "Axis values must be ranging from 0 to (dims - 1)."));

    std::vector<int64_t> new_x(x.size());
    for (size_t i = 0; i < x.size(); i++) {
      new_x[i] = x[axis[i]];
    }
    return new_x;
  }

  std::pair<math::MatDescriptor, memory::dims> GetInputDimsAndStrides(
      const ExecutionContext& ctx, std::string input_name) {
    auto shape = ctx.Attr<std::vector<int>>("fused_reshape_" + input_name);
    auto axis = ctx.Attr<std::vector<int>>("fused_transpose_" + input_name);
    auto input_dims = ctx.Input<Tensor>(input_name)->dims();
    auto new_dims = input_dims;
    if (!shape.empty() && !axis.empty()) {
      new_dims = input_dims.reshape(shape).transpose(axis);
    }

    auto& MatrixDimsFromVector = input_name == "X" ? RowMatrixDimsFromVector
                                                   : ColumnMatrixDimsFromVector;
    math::MatDescriptor mat_dim =
        math::CreateMatrixDescriptor(MatrixDimsFromVector(new_dims), 0,
                                     ctx.Attr<bool>("transpose_" + input_name));

    memory::dims strides;
    if (!shape.empty()) {
      auto shape2 = input_dims.reshape(shape);
      strides.push_back(1);
      for (auto i = shape2.size() - 1; i > 0; --i) {
        strides.insert(strides.begin(), strides.front() * shape2[i]);
      }
      strides = Transpose(strides, axis);
      if (shape.size() == 4)
        strides.erase(strides.begin());
      else if (shape.size() == 2)
        strides.insert(strides.begin(), shape[0] * shape[1]);
      mat_dim.stride_ = strides[0];
      if (mat_dim.trans_) std::swap(*strides.rbegin(), *(++strides.rbegin()));
    }
    return std::make_pair(mat_dim, strides);
  }

  bool IsInputFused(const ExecutionContext& ctx) const {
    return !(ctx.Attr<std::vector<int>>("fused_reshape_X").empty() &&
             ctx.Attr<std::vector<int>>("fused_reshape_Y").empty());
  }

  bool IsOutputFused(const ExecutionContext& ctx) const {
    auto& fused_reshape_Out = ctx.Attr<std::vector<int>>("fused_reshape_Out");
    auto& fused_transpose_Out =
        ctx.Attr<std::vector<int>>("fused_transpose_Out");
    return !fused_reshape_Out.empty() && !fused_transpose_Out.empty();
  }

  void CorrectStridesWhenFloatOutputFused(const ExecutionContext& ctx,
                                          const memory::dim N, memory::dim b,
                                          memory::dims* out_strides) const {
    if (!IsInt8<OT>() && !IsBfloat16<OT>() && IsOutputFused(ctx)) {
      *out_strides = {N, b * N, 1};
    }
  }

  MatMulDims GetMatmulDims(const ExecutionContext& ctx) {
    math::MatDescriptor mat_dim_x;
    memory::dims strides_x;
    std::tie(mat_dim_x, strides_x) = GetInputDimsAndStrides(ctx, "X");
    math::MatDescriptor mat_dim_y;
    memory::dims strides_y;
    std::tie(mat_dim_y, strides_y) = GetInputDimsAndStrides(ctx, "Y");

    auto x_bs = mat_dim_x.batch_size_;
    auto y_bs = mat_dim_y.batch_size_;
    PADDLE_ENFORCE_EQ(x_bs > 0 && y_bs > 0 && x_bs != y_bs, false,
                      platform::errors::InvalidArgument(
                          "If batch sizes of X and Y are positive,"
                          "they have to be equal."));

    memory::dim out_bs = x_bs || y_bs ? std::max(x_bs, y_bs) : 1;
    const memory::dim M = mat_dim_x.height_;
    const memory::dim N = mat_dim_y.width_;
    const memory::dim K = mat_dim_x.width_;

    batch_size_ = 1;
    if (out_bs > 1 && (IsOutputFused(ctx) || IsInputFused(ctx))) {
      auto& x_dims = ctx.Input<Tensor>("X")->dims();
      auto& y_dims = ctx.Input<Tensor>("Y")->dims();
      batch_size_ = x_bs > y_bs ? x_dims[0] : y_dims[0];
      x_bs /= batch_size_;
      y_bs /= batch_size_;
      out_bs /= batch_size_;
    }
    memory::dims x_dims = {x_bs > 0 ? x_bs : 1, M, K};
    memory::dims y_dims = {y_bs > 0 ? y_bs : 1, K, N};
    memory::dims out_dims = {out_bs, M, N};

    x_offset_ = x_bs * M * K * sizeof(XT);
    y_offset_ = y_bs * K * N * sizeof(YT);
    out_offset_ = out_bs * M * N * sizeof(OT);

    // Translate transA and transB
    if (strides_x.empty())
      strides_x = !ctx.Attr<bool>("transpose_X") ? memory::dims{M * K, K, 1}
                                                 : memory::dims{M * K, 1, M};
    if (strides_y.empty())
      strides_y = !ctx.Attr<bool>("transpose_Y") ? memory::dims{N * K, N, 1}
                                                 : memory::dims{N * K, 1, K};
    memory::dims out_strides = memory::dims{M * N, N, 1};

    CorrectStridesWhenFloatOutputFused(ctx, N, out_bs, &out_strides);

    return {x_dims, y_dims, out_dims, strides_x, strides_y, out_strides};
  }

  void CreateMemories(const ExecutionContext& ctx) {
    auto matmul_dims = GetMatmulDims(ctx);

    x_mem_ = CreateMemory<XT>(matmul_dims.x_dims, matmul_dims.x_strides,
                              ctx.Input<Tensor>("X")->data<XT>());
    y_mem_ = CreateMemory<YT>(matmul_dims.y_dims, matmul_dims.y_strides,
                              ctx.Input<Tensor>("Y")->data<YT>());
    out_mem_ = CreateMemory<OT>(
        matmul_dims.out_dims, matmul_dims.out_strides,
        ctx.Output<Tensor>("Out")->mutable_data<OT>(ctx.GetPlace()));
  }

  float ComputeOutputScale(const ExecutionContext& ctx) {
    float scale_x = ctx.Attr<float>("Scale_x");
    float scale_y = ctx.Attr<float>("Scale_y");
    bool force_fp32_out = ctx.Attr<bool>("force_fp32_output");
    float scale_out = force_fp32_out ? 1.f : ctx.Attr<float>("Scale_out");
    float alpha = ctx.Attr<float>("alpha");
    return alpha * scale_out / (scale_x * scale_y);
  }

  void CreatePrimitive(const ExecutionContext& ctx) {
    dnnl::primitive_attr attr;
    float scale_out = ComputeOutputScale(ctx);
    if (scale_out != 1.0f) {
      constexpr unsigned tensor_wide_scale = 0;
      attr.set_output_scales(tensor_wide_scale, {scale_out});
    }

    auto matmul_d = dnnl::matmul::desc(x_mem_.get_desc(), y_mem_.get_desc(),
                                       out_mem_.get_desc());
    auto matmul_pd = dnnl::matmul::primitive_desc(matmul_d, attr, engine_);
    matmul_prim_ = dnnl::matmul(matmul_pd);
  }

  void Execute() {
    dnnl::stream stream(engine_);

    void* x_ptr = x_mem_.get_data_handle();
    void* y_ptr = y_mem_.get_data_handle();
    void* out_ptr = out_mem_.get_data_handle();
    for (uint16_t i = 0; i < batch_size_; i++) {
      x_mem_.set_data_handle(x_ptr);
      y_mem_.set_data_handle(y_ptr);
      out_mem_.set_data_handle(out_ptr);
      matmul_prim_.execute(stream, {
                                       {MKLDNN_ARG_SRC, x_mem_},
                                       {MKLDNN_ARG_WEIGHTS, y_mem_},
                                       {MKLDNN_ARG_DST, out_mem_},
                                   });
      x_ptr = static_cast<char*>(x_ptr) + x_offset_;
      y_ptr = static_cast<char*>(y_ptr) + y_offset_;
      out_ptr = static_cast<char*>(out_ptr) + out_offset_;
    }
    stream.wait();
  }

  void SetOutputFormat(const ExecutionContext& ctx) {
    using platform::MKLDNNFormatForSize;
    auto* out = ctx.Output<Tensor>("Out");
    auto format =
        MKLDNNFormatForSize(out->dims().size(), MKLDNNMemoryFormat::nchw);
    out->set_format(format);
    out->set_layout(DataLayout::kMKLDNN);
  }

  void UpdateDataPointers(const ExecutionContext& ctx) {
    auto* x = ctx.Input<Tensor>("X");
    auto* y = ctx.Input<Tensor>("Y");
    auto* out = ctx.Output<Tensor>("Out");
    x_mem_.set_data_handle(to_void_cast(x->data<XT>()));
    y_mem_.set_data_handle(to_void_cast(y->data<YT>()));
    out_mem_.set_data_handle(out->mutable_data<OT>(ctx.GetPlace()));
  }

  // If initialized, x memory should've been already initialized
  bool IsInitialized() { return initialized_; }

  void SetInitialized() { initialized_ = true; }

 private:
  struct memory_offsets {
    size_t x_offset;
    size_t y_offset;
    size_t out_offset;
  };

  dnnl::engine engine_;
  dnnl::memory x_mem_;
  dnnl::memory y_mem_;
  dnnl::memory out_mem_;
  dnnl::matmul matmul_prim_;
  uint32_t x_offset_;
  uint32_t y_offset_;
  uint32_t out_offset_;
  uint16_t batch_size_;
  bool initialized_ = false;
};

template <typename XT, typename YT, typename OT>
static std::shared_ptr<MatMulFactory<XT, YT, OT>> GetPrimitiveFactory(
    const ExecutionContext& ctx) {
  const auto& out_name = ctx.OutputName("Out");
  const auto& dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
  const auto batch_size = ctx.Input<Tensor>("X")->dims()[0];
  std::string key = platform::CreateKey(dev_ctx, batch_size, out_name);
  key = platform::ExtendKeyWithThreadInfoIfNeeded(dev_ctx, key);

  auto factory =
      std::static_pointer_cast<MatMulFactory<XT, YT, OT>>(dev_ctx.GetBlob(key));
  if (factory == nullptr) {
    factory = std::make_shared<MatMulFactory<XT, YT, OT>>();
    dev_ctx.SetBlob(key, factory);
  }

  return factory;
}

// Choose appropriate primitive factory implementation based on inferred
// output type (uint8, int8 or float).
template <typename XT, typename YT>
static void ExecuteMatMul(const ExecutionContext& ctx) {
  constexpr bool is_int8 = IsInt8<XT>();
  constexpr bool is_bfloat16 = IsBfloat16<XT>();
  const bool force_fp32_output = ctx.Attr<bool>("force_fp32_output");
  constexpr bool fuse_relu = false;  // TODO(intel): Enable eltwise fuses
  if (force_fp32_output || ((!is_int8) && (!is_bfloat16))) {
    GetPrimitiveFactory<XT, YT, float>(ctx)->CreateAndExecute(ctx);
  } else if (is_bfloat16) {
    GetPrimitiveFactory<XT, YT, paddle::platform::bfloat16>(ctx)
        ->CreateAndExecute(ctx);
  } else if (fuse_relu) {
    GetPrimitiveFactory<XT, YT, uint8_t>(ctx)->CreateAndExecute(ctx);
  } else {
    GetPrimitiveFactory<XT, YT, int8_t>(ctx)->CreateAndExecute(ctx);
  }
}

template <typename T>
class DNNLMatMulKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    if (ctx.HasAttr("head_number")) {
      PADDLE_ENFORCE_EQ(
          ctx.Attr<int>("head_number"), 1,
          platform::errors::Unimplemented(
              "DNNL matmul doesn't support multiple heads. Expected "
              "head_number=1. But received `head_number` is %d",
              ctx.Attr<int>("head_number")));
    }
    platform::MKLDNNDeviceContext::tls().log_lib_version();
    ExecuteMatMul<T, T>(ctx);
  }
};
}  // namespace operators
}  // namespace paddle
namespace ops = paddle::operators;

REGISTER_OP_KERNEL(matmul, MKLDNN, ::paddle::platform::CPUPlace,
                   ops::DNNLMatMulKernel<float>,
                   ops::DNNLMatMulKernel<paddle::platform::bfloat16>,
                   ops::DNNLMatMulKernel<int8_t>,
                   ops::DNNLMatMulKernel<uint8_t>);