pyramid_hash_op.cc

/* Copyright (c) 2019 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 <xxhash.h>

#include <algorithm>
#include <cmath>

#include "paddle/fluid/framework/convert_utils.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/search_compute.h"

extern "C" {
#include "math/bloomfilter.h"
}

namespace paddle {
namespace operators {

using LoD = framework::LoD;

class PyramidHashOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "X (Tensor, MUST be Tensor<!!!_int32_!!!>) Input variable which "
             "should contain lod information.");
    AddInput("W", "W (Tensor)");
    AddInput("WhiteList", "WhiteList (Tensor)");
    AddInput("BlackList", "BlackList (Tensor)");
    AddAttr<int>("num_emb", "num_emb").SetDefault(0).EqualGreaterThan(0);
    AddAttr<int>("space_len", "space_len").SetDefault(0).EqualGreaterThan(0);
    AddAttr<int>("pyramid_layer", "pyramid_layer (must be >= 2)")
        .SetDefault(2)
        .EqualGreaterThan(2);
    AddAttr<int>("rand_len", "rand_len").SetDefault(0).EqualGreaterThan(0);
    AddAttr<float>("drop_out_percent", "drop_out_percent")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("is_training", "is_training")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<bool>("use_filter", "use_filter").SetDefault(true);
    AddAttr<int>("white_list_len", "white_list_len")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("black_list_len", "black_list_len")
        .SetDefault(0)
        .EqualGreaterThan(0);
    AddAttr<int>("seed", "seed").SetDefault(0).EqualGreaterThan(0);
    AddAttr<float>("lr", "learning rate").SetDefault(0.0).EqualGreaterThan(0.0);
    AddAttr<std::string>(
        "distribute_update_vars",
        "['PyramidHash_emb_0','Filter']"
        "Decided which params should be updated in distribute training. "
        "Used in Distribute Transpiler to create a trainer/server program.")
        .SetDefault("");
    AddOutput("Out", "Out (Tensor, default Tensor<float>) Output variable");
    AddOutput("DropPos", "Out (Tensor, Tensor<int>) Output variable");
    AddOutput("X_Temp_Out", "Out (Tensor, Tensor<int>) Output variable")
        .AsIntermediate();

    AddComment(R"DOC(
      PyramidHash

      NOTE: only support 'float32' data type now.

    )DOC");
  }
};

class PyramidHashOP : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("X"),
        true,
        platform::errors::NotFound("Input(X) of PyramidHashOP is not found."));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("W"),
        true,
        platform::errors::NotFound("Input(W) of PyramidHashOP is not found."));
    PADDLE_ENFORCE_EQ(ctx->HasOutput("Out"),
                      true,
                      platform::errors::NotFound(
                          "Output(Out) of PyramidHashOP is not found."));
    PADDLE_ENFORCE_EQ(ctx->HasOutput("DropPos"),
                      true,
                      platform::errors::NotFound(
                          "Output(DropPos) of PyramidHashOP is not found."));

    auto x_dims = ctx->GetInputDim("X");
    PADDLE_ENFORCE_EQ(x_dims.size(),
                      2,
                      platform::errors::InvalidArgument(
                          "The rank of Input(X) of PyramidHashOP is invalid. "
                          "It should be 2, but got %d",
                          x_dims.size()));

    auto w_dims = ctx->GetInputDim("W");
    PADDLE_ENFORCE_EQ(w_dims.size(),
                      2,
                      platform::errors::InvalidArgument(
                          "The rank of Input(W) of PyramidHashOP is invalid. "
                          "It should be 2, but got %d",
                          w_dims.size()));

    int space_len = ctx->Attrs().Get<int>("space_len");
    int rand_len = ctx->Attrs().Get<int>("rand_len");

    PADDLE_ENFORCE_EQ(
        w_dims[0],
        space_len + rand_len,
        platform::errors::InvalidArgument(
            "The first dimension of Input(W) of PyramidHashOP is invalid. "
            "It should be space_len + rand_len, but now %d != %d + %d",
            w_dims[0],
            space_len,
            rand_len));
    PADDLE_ENFORCE_EQ(
        w_dims[1],
        1,
        platform::errors::InvalidArgument(
            "The second dimension of Input(W) of PyramidHashOP is invalid."
            " It should be 1, but got %d",
            w_dims[1]));

    int num_emb = ctx->Attrs().Get<int>("num_emb");
    PADDLE_ENFORCE_EQ(
        num_emb % rand_len,
        0,
        platform::errors::InvalidArgument(
            "The PyramidHashOP's Attr(num_emb) should mod Attr(rand_len), "
            "but num_emb is %d, rand_len is %d",
            num_emb,
            rand_len));

    int white_list_len = ctx->Attrs().Get<int>("white_list_len");
    if (white_list_len > 0) {
      PADDLE_ENFORCE_EQ(
          ctx->HasInput("WhiteList"),
          true,
          platform::errors::NotFound("Input(WhiteList) of PyramidHashOP is not "
                                     "found but white_list_len > 0."));
      auto wl_dims = ctx->GetInputDim("WhiteList");
      PADDLE_ENFORCE_EQ(
          wl_dims.size(),
          2,
          platform::errors::InvalidArgument(
              "The rank of Input(WhiteList) of PyramidHashOP is invalid."
              " It should be 2, but got %d",
              wl_dims.size()));
      PADDLE_ENFORCE_EQ(wl_dims[0],
                        white_list_len,
                        platform::errors::InvalidArgument(
                            "The first dimension of Input(WhiteList) of "
                            "PyramidHashOP is invalid."
                            " It should be equal to Attr(white_list_len) "
                            ", but first dimension is %d, white_list_len is %d",
                            wl_dims[0],
                            white_list_len));
      PADDLE_ENFORCE_EQ(wl_dims[1],
                        1,
                        platform::errors::InvalidArgument(
                            "The second dimension of Input(WhiteList) of "
                            "PyramidHashOP is invalid."
                            " It should be 1, but got %d",
                            wl_dims[1]));
    }

    int black_list_len = ctx->Attrs().Get<int>("black_list_len");
    if (black_list_len > 0) {
      PADDLE_ENFORCE_EQ(
          ctx->HasInput("BlackList"),
          true,
          platform::errors::NotFound("Input(BlackList) of PyramidHashOP is not "
                                     "found but black_list_len > 0."));
      auto bl_dims = ctx->GetInputDim("BlackList");
      PADDLE_ENFORCE_EQ(
          bl_dims.size(),
          2,
          platform::errors::InvalidArgument(
              "The rank of Input(BlackList) of PyramidHashOP is invalid."
              " It should be 2, but got %d",
              bl_dims.size()));
      PADDLE_ENFORCE_EQ(bl_dims[0],
                        black_list_len,
                        platform::errors::InvalidArgument(
                            "The first dimension of Input(BlackList) of "
                            "PyramidHashOP is invalid."
                            " It should be equal to Attr(black_list_len)"
                            ", but first dimension is %d, black_list_len is %d",
                            bl_dims[0],
                            black_list_len));
      PADDLE_ENFORCE_EQ(bl_dims[1],
                        1,
                        platform::errors::InvalidArgument(
                            "The second dimension of Input(BlackList) of "
                            "PyramidHashOP is invalid."
                            " It should be 1, but got %d",
                            bl_dims[1]));
    }

    if (ctx->IsRuntime()) {
      // something to do in runtime.
    } else {
      // compile time
      ctx->SetOutputDim("Out", phi::make_ddim({-1, num_emb}));
      ctx->SetOutputDim("X_Temp_Out", x_dims);
      ctx->ShareLoD("X", /*->*/ "Out");
    }
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "W"), ctx.GetPlace());
  }
};

template <typename DeviceContext, typename T>
class CPUPyramidHashOPKernel : public framework::OpKernel<T> {
 public:
  bool should_use_term(math::bloomfilter* _filter,
                       math::bloomfilter* _black_filter,
                       const float* word_repr,
                       int len) const {
    return (!_filter || 1 == math::bloomfilter_get(
                                 _filter, word_repr, len * sizeof(float))) &&
           (!_black_filter ||
            0 == math::bloomfilter_get(
                     _black_filter, word_repr, len * sizeof(float)));
  }

  void hash_embedding_ff(const float* hash_id,
                         int len,
                         T* top_pos,
                         const T* weights,
                         int _num_emb,
                         int _rand_len,
                         int _space_len) const {
    unsigned int pos1 = XXH32(hash_id, len * sizeof(float), 0) % _space_len;
    unsigned int pos2 =
        XXH32(hash_id, len * sizeof(float), _rand_len) % _space_len;

    for (int j = 0; j != _num_emb; j += _rand_len) {
      if (j + _rand_len < _num_emb) {
        __builtin_prefetch(weights + pos2);
        __builtin_prefetch(top_pos + j + _rand_len);
      }

      unsigned int pos3 =
          XXH32(hash_id, len * sizeof(float), j + 2 * _rand_len) % _space_len;
      memcpy(
          top_pos + j, const_cast<T*>(weights + pos1), _rand_len * sizeof(T));
      pos1 = pos2;
      pos2 = pos3;
    }
  }

  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* bottom = ctx.Input<phi::DenseTensor>("X");
    auto* _blobs_0 = ctx.Input<phi::DenseTensor>("W");
    auto* _blobs_1 = ctx.Input<phi::DenseTensor>("WhiteList");
    auto* _blobs_2 = ctx.Input<phi::DenseTensor>("BlackList");
    auto* top = ctx.Output<phi::DenseTensor>("Out");
    auto* drop_pos = ctx.Output<phi::DenseTensor>("DropPos");

    int _num_emb = ctx.Attr<int>("num_emb");
    bool use_filter = ctx.Attr<bool>("use_filter");
    int white_list_len = ctx.Attr<int>("white_list_len");
    int black_list_len = ctx.Attr<int>("black_list_len");
    int _pyramid_layer = ctx.Attr<int>("pyramid_layer");
    int _is_training = ctx.Attr<int>("is_training");
    int seed = ctx.Attr<int>("seed");
    unsigned int _seed = (unsigned int)seed;
    int _rand_len = ctx.Attr<int>("rand_len");
    int _space_len = ctx.Attr<int>("space_len");
    float _drop_out_percent = ctx.Attr<float>("drop_out_percent");

    const auto& offset = bottom->lod()[0];
    const auto* bottom_data_ori = bottom->data<int32_t>();
    auto* buff = ctx.Output<phi::DenseTensor>("X_Temp_Out");
    buff->Resize(phi::make_ddim({bottom->dims()[0], bottom->dims()[1]}));
    float* bottom_data = buff->mutable_data<float>(ctx.GetPlace());
    for (int i = 0; i < bottom->dims()[0]; i++) {
      bottom_data[i] = bottom_data_ori[i];
    }

    const auto* weights = _blobs_0->data<T>();

    std::vector<size_t> top_offset;
    top_offset.resize(offset.size());
    top_offset[0] = 0;

    math::bloomfilter* _filter = NULL;
    math::bloomfilter* _black_filter = NULL;
    if (use_filter) {
      if (white_list_len != 0) {
        _filter = (math::bloomfilter*)_blobs_1->data<float>();
        PADDLE_ENFORCE_EQ(
            math::bloomfilter_check(_filter),
            1,
            platform::errors::PreconditionNotMet(
                "The white filter is not loaded successfully, please make sure "
                "'white_list_len': %d is valid for Input(WhiteList).",
                white_list_len));
      }
      if (black_list_len != 0) {
        _black_filter = (math::bloomfilter*)_blobs_2->data<float>();
        PADDLE_ENFORCE_EQ(
            math::bloomfilter_check(_black_filter),
            1,
            platform::errors::PreconditionNotMet(
                "The black filter is not loaded successfully, please make sure "
                "'black_list_len': %d is valid for Input(BlackList).",
                black_list_len));
      }
    }

    drop_pos->Resize(phi::make_ddim(
        {bottom->dims()[0] * bottom->dims()[1] * _pyramid_layer, 1}));
    std::vector<size_t> drop_pos_offset;
    drop_pos_offset.resize(offset.size());
    drop_pos_offset[0] = 0;
    int* iter = drop_pos->mutable_data<int>(ctx.GetPlace());
    int* iter_end = iter;

    for (size_t i = 0; i < top_offset.size() - 1; ++i) {
      int w = offset[i + 1] - offset[i];
      int nsentense_with_pyramid = 0;
      if (w < 2) {
        nsentense_with_pyramid = 0;
      } else {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (should_use_term(_filter,
                                _black_filter,
                                (const float*)(bottom_data + offset[i] + l),
                                ilayer + 1)) {
              if (_is_training != 0) {
                unsigned int rand_val = rand_r(&_seed);
                float rate = static_cast<float>(rand_val) / (RAND_MAX);
                *(iter_end++) = (rate < _drop_out_percent ? 0 : 1);
              } else {
                *(iter_end++) = 1;
              }
            } else {
              *(iter_end++) = 0;
            }
          }
        }
        nsentense_with_pyramid = std::count(iter, iter_end, 1);
        iter = iter_end;
      }
      drop_pos_offset[i + 1] = drop_pos_offset[i] + nsentense_with_pyramid;
      top_offset[i + 1] =
          top_offset[i] +
          (nsentense_with_pyramid == 0 ? 1 : nsentense_with_pyramid);
    }

    int top_l = top_offset[top_offset.size() - 1];

    framework::LoD top_lod;
    top_lod.push_back(top_offset);
    top->set_lod(top_lod);
    top->Resize(phi::make_ddim({top_l, _num_emb}));
    auto* top_data = top->mutable_data<T>(ctx.GetPlace());

    framework::LoD drop_pos_lod;
    drop_pos_lod.push_back(drop_pos_offset);
    drop_pos->set_lod(drop_pos_lod);

    iter = drop_pos->mutable_data<int>(ctx.GetPlace());
    int top_counter = 0;
    for (size_t i = 0; i < offset.size() - 1; ++i) {
      int w_drop = drop_pos_offset[i + 1] - drop_pos_offset[i];
      int w = offset[i + 1] - offset[i];
      if (w_drop == 0) {
        if (w >= 2) {
          for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w;
               ++ilayer) {
            for (int l = 0; l < w - ilayer; ++l) {
              iter++;
            }
          }
        }
        auto* top_pos = top_data + top_counter++ * _num_emb;
        memset(top_pos, 0, _num_emb * sizeof(T));
        continue;
      }
      if (w >= 2) {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (*(iter++) == 0) {
              // do nothing
            } else {
              auto* top_pos = top_data + top_counter++ * _num_emb;
              hash_embedding_ff((const float*)(bottom_data + offset[i] + l),
                                ilayer + 1,
                                top_pos,
                                weights,
                                _num_emb,
                                _rand_len,
                                _space_len);
            }
          }
        }
      }
    }
    if (iter != iter_end) {
      exit(1);
    }
    auto weight_type = framework::TransToProtoVarType(_blobs_0->dtype());
    if (_is_training == 0 && weight_type != framework::proto::VarType::INT8) {
      axpy_noadd(top_data,
                 top_data,
                 top->dims()[0] * top->dims()[1],
                 _drop_out_percent);
    }
  }
};

class PyramidHashOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE_EQ(ctx->HasInput("X"),
                      true,
                      platform::errors::NotFound(
                          "Input(X) of PyramidHashOpGrad is not found."));
    PADDLE_ENFORCE_EQ(ctx->HasInput("W"),
                      true,
                      platform::errors::NotFound(
                          "Input(W) of PyramidHashOpGrad is not found."));
    PADDLE_ENFORCE_EQ(ctx->HasInput("DropPos"),
                      true,
                      platform::errors::NotFound(
                          "Input(DropPos) of PyramidHashOpGrad is not found."));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput("X_Temp_Out"),
        true,
        platform::errors::NotFound(
            "Input(X_Temp_Out) of PyramidHashOpGrad is not found."));
    PADDLE_ENFORCE_EQ(
        ctx->HasInput(framework::GradVarName("Out")),
        true,
        platform::errors::NotFound(
            "Input(Out@Grad) of PyramidHashOpGrad is not found."));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "W"), ctx.GetPlace());
  }
};

template <typename T>
class PyramidHashGradOpMaker : public framework::SingleGradOpMaker<T> {
 public:
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;

 protected:
  void Apply(GradOpPtr<T> op_desc_ptr) const override {
    op_desc_ptr->SetType("pyramid_hash_grad");
    op_desc_ptr->SetInput("X", this->Input("X"));
    op_desc_ptr->SetInput("W", this->Input("W"));
    op_desc_ptr->SetInput("DropPos", this->Output("DropPos"));
    op_desc_ptr->SetInput("X_Temp_Out", this->Output("X_Temp_Out"));

    op_desc_ptr->SetInput(framework::GradVarName("Out"),
                          this->OutputGrad("Out"));
    op_desc_ptr->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op_desc_ptr->SetAttrMap(this->Attrs());
  }
};

template <typename DeviceContext, typename T>
class CPUPyramidHashOPGradKernel : public framework::OpKernel<T> {
 public:
  void hash_embedding_bp(const T* hash_id,
                         int len,
                         const T* top_pos,
                         T* weights,
                         T mlr,
                         int _num_emb,
                         int _rand_len,
                         int _space_len) const {
    for (int j = 0; j != _num_emb; j += _rand_len) {
      unsigned int pos = XXH32(hash_id, len * sizeof(T), j) % _space_len;
      axpy(top_pos + j, weights + pos, _rand_len, mlr);
    }
  }

  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* bottom = ctx.Input<phi::DenseTensor>("X");
    auto* _blobs = ctx.Input<phi::DenseTensor>("W");
    auto* drop_pos = ctx.Input<phi::DenseTensor>("DropPos");
    auto* top = ctx.Input<phi::DenseTensor>(framework::GradVarName("Out"));

    int _num_emb = ctx.Attr<int>("num_emb");
    float _lr = ctx.Attr<float>("lr");
    int _rand_len = ctx.Attr<int>("rand_len");
    int _space_len = ctx.Attr<int>("space_len");
    int _pyramid_layer = ctx.Attr<int>("pyramid_layer");

    auto* buff = ctx.Input<phi::DenseTensor>("X_Temp_Out");
    auto* bottom_data = buff->data<T>();

    int _slot_len = bottom->dims()[0];
    if (static_cast<size_t>(_slot_len) == bottom->lod()[0].size() - 1 &&
        std::count(bottom_data, bottom_data + _slot_len, -1) == _slot_len) {
      return;
    }

    auto& offset = bottom->lod()[0];
    auto& drop_pos_offset = drop_pos->lod()[0];

    const auto* top_diff = top->data<T>();
    // in-place update weight, so need const_cast
    T* weights = const_cast<T*>(_blobs->data<T>());
    T mlr = -1.0 * _lr;

    const int* iter = drop_pos->data<int>();
    int top_counter = 0;
    for (size_t i = 0; i < offset.size() - 1; ++i) {
      int w = offset[i + 1] - offset[i];
      int w_drop = drop_pos_offset[i + 1] - drop_pos_offset[i];
      if (w_drop == 0) {
        top_counter++;
      }
      if (w > 1) {
        for (int ilayer = 1; ilayer < _pyramid_layer && ilayer < w; ++ilayer) {
          for (int l = 0; l < w - ilayer; ++l) {
            if (*(iter++) == 0) {
              // do nothing
            } else {
              const T* top_pos = top_diff + top_counter++ * _num_emb;
              hash_embedding_bp((const T*)(bottom_data + offset[i] + l),
                                ilayer + 1,
                                top_pos,
                                weights,
                                mlr,
                                _num_emb,
                                _rand_len,
                                _space_len);
            }
          }
        }
      } else {
        // do nothing
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plt = paddle::platform;
namespace frm = paddle::framework;
REGISTER_OPERATOR(pyramid_hash,
                  ops::PyramidHashOP,
                  ops::PyramidHashOpMaker,
                  ops::PyramidHashGradOpMaker<paddle::framework::OpDesc>,
                  ops::PyramidHashGradOpMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(pyramid_hash_grad, ops::PyramidHashOpGrad);

REGISTER_OP_CPU_KERNEL(pyramid_hash,
                       ops::CPUPyramidHashOPKernel<phi::CPUContext, float>,
                       ops::CPUPyramidHashOPKernel<phi::CPUContext, int8_t>);
REGISTER_OP_CPU_KERNEL(pyramid_hash_grad,
                       ops::CPUPyramidHashOPGradKernel<phi::CPUContext, float>);