fusion_gru_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. */

#include "paddle/fluid/operators/fusion_gru_op.h"
#include <cstring>  // for memcpy
#include <string>
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/operators/math/cpu_vec.h"
#include "paddle/fluid/operators/math/fc_compute.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
#include "paddle/fluid/platform/cpu_info.h"

namespace paddle {
namespace operators {

void FusionGRUOp::InferShape(framework::InferShapeContext* ctx) const {
  PADDLE_ENFORCE(ctx->HasInput("X"), "Assert only one Input(X) of GRU.");
  PADDLE_ENFORCE(ctx->HasInput("WeightX"),
                 "Assert only one Input(WeightX) of GRU.");
  PADDLE_ENFORCE(ctx->HasInput("WeightH"),
                 "Assert only one Input(WeightH) of GRU.");
  PADDLE_ENFORCE(ctx->HasOutput("XX"), "Assert only one Output(XX) of GRU.");
  PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
                 "Assert only one Output(Hidden) of GRU.");

  auto x_dims = ctx->GetInputDim("X");
  PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2.");

  auto wx_dims = ctx->GetInputDim("WeightX");
  PADDLE_ENFORCE_EQ(wx_dims.size(), 2,
                    "The rank of Input(WeightX) should be 2.");
  PADDLE_ENFORCE_EQ(wx_dims[0], x_dims[1],
                    "The first dimension of Input(WeightX) "
                    "should be %d.",
                    x_dims[1]);

  int frame_size = wx_dims[1] / 3;
  auto wh_dims = ctx->GetInputDim("WeightH");
  PADDLE_ENFORCE_EQ(wh_dims.size(), 2,
                    "The rank of Input(WeightH) should be 2.");
  PADDLE_ENFORCE_EQ(wh_dims[0], frame_size,
                    "The first dimension of Input(WeightH) "
                    "should be %d.",
                    frame_size);
  PADDLE_ENFORCE_EQ(wh_dims[1], 3 * frame_size,
                    "The second dimension of Input(WeightH) "
                    "should be 3 * %d.",
                    frame_size);

  if (ctx->HasInput("H0")) {
    auto h0_dims = ctx->GetInputDim("H0");
    PADDLE_ENFORCE_EQ(h0_dims[1], frame_size,
                      "The width of H0 must be equal to frame_size.");
  }
  if (ctx->HasInput("Bias")) {
    auto b_dims = ctx->GetInputDim("Bias");
    PADDLE_ENFORCE_EQ(b_dims.size(), 2, "The rank of Input(Bias) should be 2.");
    PADDLE_ENFORCE_EQ(b_dims[0], 1,
                      "The first dimension of Input(Bias) should be 1.");
    PADDLE_ENFORCE_EQ(b_dims[1], frame_size * 3,
                      "The shape of Bias must be [1, frame_size * 3].");
  }
  framework::DDim out_dims({x_dims[0], frame_size});
  ctx->SetOutputDim("Hidden", out_dims);
  ctx->ShareLoD("X", "Hidden");
  int xx_width;
  if (ctx->Attrs().Get<bool>("use_seq")) {
    xx_width = wx_dims[1];
  } else {
    xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1];
    PADDLE_ENFORCE(ctx->HasOutput("ReorderedH0"),
                   "Assert only one Output(ReorderedH0) of GRU.");
    PADDLE_ENFORCE(ctx->HasOutput("BatchedInput"),
                   "Assert only one Output(BatchedInput) of GRU.");
    PADDLE_ENFORCE(ctx->HasOutput("BatchedOut"),
                   "Assert only one Output(BatchedOut) of GRU.");
    ctx->SetOutputDim("BatchedInput", {x_dims[0], wx_dims[1]});
    ctx->SetOutputDim("BatchedOut", out_dims);
  }
  ctx->SetOutputDim("XX", {x_dims[0], xx_width});
  ctx->ShareLoD("X", "XX");
}

framework::OpKernelType FusionGRUOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<framework::LoDTensor>("X")->type()),
      ctx.device_context());
}

void FusionGRUOpMaker::Make() {
  AddInput("X",
           "(LoDTensor) the input is a LodTensor, which support "
           "variable-time length input sequence. The underlying tensor in "
           "this LoDTensor is a matrix with shape (T X M), where T is the "
           "total time steps in this mini-batch, M is the dim size of x.");
  AddInput("H0",
           "(Tensor, optional) The initial hidden state is an optional "
           "input. This is a tensor with shape (N x D), where N is the "
           "batch size, D is the hidden size.")
      .AsDispensable();
  AddInput("WeightX",
           "(Tensor) The FC weight with shape (M x 3D),"
           "where M is the dim size of x, D is the hidden size. ");
  AddInput("WeightH",
           "(Tensor) (D x 3D) Same as GRUOp, where D is the hidden size. "
           "This weight is not exactly D x 3D as: {W_update, W_reset, W_state}"
           "Acutally they are D x 2D and D x D two part weights."
           "{W_update, W_reset; W_state}"
           "{D x (D + D); D x D}");
  AddInput("Bias",
           "(Tensor, optional) (1 x 3D)."
           "Almost same as GRUOp."
           "Note: if have FC bias it should be added on this bias.")
      .AsDispensable();
  AddOutput("ReorderedH0", "(Tensor) (N x D), which N is the min-batch size.")
      .AsIntermediate();
  AddOutput("XX",
            "(LoDTensor) the result after X * WeightX (size is T x 3D)"
            " or batched_X (size is T x M), this will be automatically chosen,"
            " where T is the total time steps in this mini-batch,"
            " D is the hidden size, M is the dim size of x input.")
      .AsIntermediate();
  AddOutput("BatchedInput",
            "(LoDTensor) This is the batched result of input X"
            "or the batched result after fc, shape (T x 3D)")
      .AsIntermediate();
  AddOutput("BatchedOut", "(LoDTensor) (T X D) save batched hidden.")
      .AsIntermediate();
  AddOutput("Hidden", "(LoDTensor) (T x D) Same as GRUOp");
  AddAttr<std::string>("activation",
                       "(string, default tanh) "
                       "The activation type used for output candidate {h}_t.")
      .SetDefault("tanh");
  AddAttr<std::string>(
      "gate_activation",
      "(string, default sigmoid) "
      "The activation type used in update gate and reset gate.")
      .SetDefault("sigmoid");
  AddAttr<bool>("is_reverse",
                "(bool, defalut: False) "
                "whether to compute reversed GRU.")
      .SetDefault(false);
  AddAttr<bool>("use_seq",
                "(bool, defalut: True) "
                "whether to use seq mode to compute GRU.")
      .SetDefault(true);
  AddComment(R"DOC(
The Fusion complete GRU Operator.
This operator fuse the fully-connected operator into GRU, 
more details can refer to GRU op.
)DOC");
}

template <typename T>
class FusionGRUKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    if (ctx.Attr<bool>("use_seq")) {
      SeqCompute(ctx);
    } else {
      BatchCompute(ctx);
    }
  }

#define INIT_VEC_FUNC                                                     \
  std::function<void(const int, const T *, T *)> act_gate, act_state;     \
  std::function<void(const int, const T*, const T*, const T*, T*)> cross; \
  auto& act_gate_str = ctx.Attr<std::string>("gate_activation");          \
  auto& act_state_str = ctx.Attr<std::string>("activation");              \
  if (platform::jit::MayIUse(platform::jit::avx)) {                       \
    math::VecActivations<T, platform::jit::avx> act_functor;              \
    act_gate = act_functor(act_gate_str);                                 \
    act_state = act_functor(act_state_str);                               \
    cross = math::vec_cross<T, platform::jit::avx>;                       \
  } else {                                                                \
    math::VecActivations<T, platform::jit::isa_any> act_functor;          \
    act_gate = act_functor(act_gate_str);                                 \
    act_state = act_functor(act_state_str);                               \
    cross = math::vec_cross<T, platform::jit::isa_any>;                   \
  }

#define INIT_BASE_INPUT_OUTPUT                        \
  auto* h0 = ctx.Input<Tensor>("H0");                 \
  auto* wx = ctx.Input<Tensor>("WeightX");            \
  auto* wh = ctx.Input<Tensor>("WeightH");            \
  auto* bias = ctx.Input<Tensor>("Bias");             \
  auto* xx = ctx.Output<LoDTensor>("XX");             \
  auto* hidden_out = ctx.Output<LoDTensor>("Hidden"); \
  bool is_reverse = ctx.Attr<bool>("is_reverse");

#define INIT_BASE_SIZES                  \
  auto x_dims = x->dims();   /* T x M*/  \
  auto wh_dims = wh->dims(); /* D x 3D*/ \
  const int total_T = x_dims[0];         \
  const int M = x_dims[1];               \
  const int D = wh_dims[0];              \
  const int D3 = wh_dims[1];             \
  const int D2 = D * 2;

  void SeqCompute(const framework::ExecutionContext& ctx) const {
    using DeviceContext = paddle::platform::CPUDeviceContext;
    auto* x = ctx.Input<LoDTensor>("X");
    INIT_BASE_INPUT_OUTPUT
    INIT_BASE_SIZES
    INIT_VEC_FUNC

    auto x_lod = x->lod();
    const int N = x_lod[0].size() - 1;
    const T* x_data = x->data<T>();
    const T* h0_data = h0 ? h0->data<T>() : nullptr;
    const T* wx_data = wx->data<T>();
    const T* wh_data = wh->data<T>();
    const T* wh_state_data = wh_data + D * D2;
    T* xx_data = xx->mutable_data<T>(ctx.GetPlace());
    T* hidden_out_data = hidden_out->mutable_data<T>(ctx.GetPlace());

    auto blas = math::GetBlas<DeviceContext, T>(ctx);
    math::FCCompute<DeviceContext, T>(blas, total_T, D3, M, x_data, wx_data,
                                      xx_data,
                                      bias ? bias->data<T>() : nullptr);

    int xx_offset = D3;
    int gate_offset = D;
    if (is_reverse) {
      const int offset = (total_T - 1) * D;
      xx_data = xx_data + offset * 3;
      hidden_out_data = hidden_out_data + offset;
      xx_offset = -D3;
      gate_offset = -D;
    }
    auto move_step = [&]() {
      xx_data = xx_data + xx_offset;
      hidden_out_data = hidden_out_data + gate_offset;
    };
    for (int i = 0; i < N; ++i) {
      int bid = is_reverse ? N - 1 - i : i;
      int seq_len = x_lod[0][bid + 1] - x_lod[0][bid];
      const T* prev_hidden_data = nullptr;
      int tstart = 0;
      if (h0_data) {
        prev_hidden_data = h0_data + bid * D;
      } else {
        // W: {W_update, W_reset; W_state}
        // update gate
        act_gate(D, xx_data, xx_data);
        // state gate
        act_state(D, xx_data + D2, xx_data + D2);
        // out = a*b
        blas.VMUL(D, xx_data, xx_data + D2, hidden_out_data);
        // save prev
        prev_hidden_data = hidden_out_data;
        tstart = 1;
        move_step();
      }
      for (int step = tstart; step < seq_len; ++step) {
        // gemm prev * (Wu + Wr)
        blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D2, D, static_cast<T>(1),
                  prev_hidden_data, D, wh_data, D2, static_cast<T>(1), xx_data,
                  D3);
        act_gate(D2, xx_data, xx_data);
        // rt = rt*ht_1 inplace result
        blas.VMUL(D, prev_hidden_data, xx_data + D, hidden_out_data);

        // gemm rt * Ws
        blas.GEMM(CblasNoTrans, CblasNoTrans, 1, D, D, static_cast<T>(1),
                  hidden_out_data, D, wh_state_data, D, static_cast<T>(1),
                  xx_data + D2, D3);
        act_state(D, xx_data + D2, xx_data + D2);
        // out = zt*ht~ + (1-zt)*ht_1
        cross(D, xx_data, xx_data + D2, prev_hidden_data, hidden_out_data);
        // save prev
        prev_hidden_data = hidden_out_data;
        move_step();
      }
    }
  }

  void BatchCompute(const framework::ExecutionContext& ctx) const {
    using DeviceContext = paddle::platform::CPUDeviceContext;
    auto* x = ctx.Input<LoDTensor>("X");
    if (x->lod()[0].size() == 2) {
      SeqCompute(ctx);
      return;
    }
    INIT_BASE_INPUT_OUTPUT
    INIT_BASE_SIZES
    INIT_VEC_FUNC

    auto* reordered_h0 = ctx.Output<Tensor>("ReorderedH0");
    auto* batched_input = ctx.Output<LoDTensor>("BatchedInput");
    auto* batched_out = ctx.Output<LoDTensor>("BatchedOut");

    const T* x_data = x->data<T>();
    const T* wx_data = wx->data<T>();
    const T* wh_data = wh->data<T>();
    T* xx_data = xx->mutable_data<T>(ctx.GetPlace());
    T* batched_input_data = batched_input->mutable_data<T>(ctx.GetPlace());
    T* batched_out_data = batched_out->mutable_data<T>(ctx.GetPlace());
    hidden_out->mutable_data<T>(ctx.GetPlace());

    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    auto blas = math::GetBlas<DeviceContext, T>(dev_ctx);
    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
    if (M > D3) {
      math::FCCompute<DeviceContext, T>(blas, total_T, D3, M, x_data, wx_data,
                                        xx_data,
                                        bias ? bias->data<T>() : nullptr);
      to_batch(dev_ctx, *xx, batched_input, true, is_reverse);
    } else {
      to_batch(dev_ctx, *x, xx, true, is_reverse);
      batched_input->set_lod(xx->lod());
      math::FCCompute<DeviceContext, T>(blas, total_T, D3, M, xx_data, wx_data,
                                        batched_input_data,
                                        bias ? bias->data<T>() : nullptr);
    }

    auto batched_lod = batched_input->lod();
    const auto& seq_order = batched_lod[2];
    const int max_bs = seq_order.size();
    reordered_h0->Resize({max_bs, D});

    int tstart = 0;
    T* prev_hidden_data = nullptr;
    if (h0) {
      // reorder h0
      T* reordered_h0_data = reordered_h0->mutable_data<T>(ctx.GetPlace());
      const T* h0_data = h0->data<T>();
      prev_hidden_data = reordered_h0_data;
      size_t sz = sizeof(T) * D;
      for (int i = 0; i < max_bs; ++i) {
        std::memcpy(reordered_h0_data, h0_data + seq_order[i] * D, sz);
        reordered_h0_data += D;
      }
    } else {
      // compute without h0
      T* cur_in_data = batched_input_data;
      T* cur_out_data = batched_out_data;
      // W: {W_update, W_reset; W_state}
      for (int i = 0; i < max_bs; ++i) {
        // update gate
        act_gate(D, cur_in_data, cur_in_data);
        // state gate
        act_state(D, cur_in_data + D2, cur_in_data + D2);
        // out = a*b
        blas.VMUL(D, cur_in_data, cur_in_data + D2, cur_out_data);
        // add offset
        cur_in_data += D3;
        cur_out_data += D;
      }
      tstart = 1;
      prev_hidden_data = batched_out_data;
    }
    // Then start from next
    const T* wh_state_data = wh_data + D * D2;
    const auto& batch_starts = batched_lod[0];
    const int max_seq_len = batch_starts.size() - 1;
    batched_input_data = batched_input_data + tstart * max_bs * D3;
    batched_out_data = batched_out_data + tstart * max_bs * D;
    for (int step = tstart; step < max_seq_len; ++step) {
      const int cur_bs = batch_starts[step + 1] - batch_starts[step];
      // gemm prev * (Wu + Wr)
      blas.GEMM(CblasNoTrans, CblasNoTrans, cur_bs, D2, D, static_cast<T>(1),
                prev_hidden_data, D, wh_data, D2, static_cast<T>(1),
                batched_input_data, D3);

      T* cur_batched_data = batched_input_data;
      T* cur_out_data = batched_out_data;
      T* cur_prev_hidden_data = prev_hidden_data;
      for (int i = 0; i < cur_bs; ++i) {
        act_gate(D2, cur_batched_data, cur_batched_data);
        // rt = rt*ht_1 inplace result
        blas.VMUL(D, cur_prev_hidden_data, cur_batched_data + D, cur_out_data);

        cur_batched_data += D3;
        cur_prev_hidden_data += D;
        cur_out_data += D;
      }

      cur_batched_data = batched_input_data;
      cur_out_data = batched_out_data;
      blas.GEMM(CblasNoTrans, CblasNoTrans, cur_bs, D, D, static_cast<T>(1),
                cur_out_data, D, wh_state_data, D, static_cast<T>(1),
                cur_batched_data + D2, D3);

      cur_prev_hidden_data = prev_hidden_data;
      for (int i = 0; i < cur_bs; ++i) {
        // ht~ = act_state(...)
        act_state(D, cur_batched_data + D2, cur_batched_data + D2);
        // out = zt*ht~ + (1-zt)*ht_1
        cross(D, cur_batched_data, cur_batched_data + D2, cur_prev_hidden_data,
              cur_out_data);

        cur_batched_data += D3;
        cur_prev_hidden_data += D;
        cur_out_data += D;
      }
      prev_hidden_data = batched_out_data;
      batched_out_data = cur_out_data;
      batched_input_data = cur_batched_data;
    }

    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
    batched_out->set_lod(batched_lod);
    to_seq(dev_ctx, *batched_out, hidden_out);
  }
#undef INIT_VEC_FUNC
#undef INIT_BASE_SIZES
#undef INIT_BASE_INPUT_OUTPUT
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OPERATOR(fusion_gru, ops::FusionGRUOp, ops::FusionGRUOpMaker,
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OP_CPU_KERNEL(fusion_gru, ops::FusionGRUKernel<float>,
                       ops::FusionGRUKernel<double>);