dgc_op.h 7.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
/* 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. */

#pragma once
#include <vector>
#include "dgc/dgc.h"
18

19
#include "paddle/fluid/framework/eigen.h"
20
#include "paddle/fluid/memory/malloc.h"
21 22
#include "paddle/fluid/operators/elementwise/elementwise_op_function.h"
#include "paddle/phi/kernels/funcs/elementwise_functor.h"
23 24 25 26 27 28

namespace paddle {
namespace operators {

inline float get_period_sparcity(const std::vector<float>& sparsity,
                                 float cur_step, float rampup_steps) {
29 30 31 32 33
  PADDLE_ENFORCE_GE(static_cast<int>(cur_step), 0,
                    platform::errors::InvalidArgument(
                        "DGC current step=%d, but it must >= 0, "
                        "please submit issue in github",
                        static_cast<int>(cur_step)));
34 35 36

  size_t idx = static_cast<int>(cur_step * sparsity.size() / rampup_steps);
  if (idx >= sparsity.size()) {
37
    idx = sparsity.size() - 1;
38 39
  }

40 41 42 43 44
  PADDLE_ENFORCE_LT(
      idx, sparsity.size(),
      platform::errors::OutOfRange(
          "sparsity index out of bounds. idx=%d >= sparsity.size=%d", idx,
          sparsity.size()));
45 46 47 48 49 50 51 52 53 54 55
  return sparsity[idx];
}

template <typename DeviceContext, typename T>
class DGCOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto u = ctx.Input<framework::Tensor>("U");
    auto v = ctx.Input<framework::Tensor>("V");
    auto g = ctx.Input<framework::Tensor>("Grad");

56 57
    auto grad_out = ctx.Output<framework::Tensor>("Grad_out");

58 59 60 61 62 63 64
    // attrs
    float m = ctx.Attr<float>("m");
    bool use_nesterov = ctx.Attr<bool>("use_nesterov");
    auto sparsity = ctx.Attr<std::vector<float>>("sparsity");
    auto rampup_begin_step = ctx.Attr<float>("rampup_begin_step");
    auto rampup_step = ctx.Attr<float>("rampup_step");

65 66 67
    // nranks
    auto nranks_tensor = ctx.Input<framework::Tensor>("nranks");
    const int nranks = static_cast<const int>(*nranks_tensor->data<float>());
68 69 70 71
    PADDLE_ENFORCE_GT(nranks, 1,
                      platform::errors::PreconditionNotMet(
                          "DGC is not useful when num_trainers <= 1. Please "
                          "use multi card or multi machine GPU"));
72

73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
    // regularization
    auto p = ctx.Input<framework::Tensor>("Param");
    float regular_coeff = ctx.Attr<float>("regular_coeff");
    int regular_type = ctx.Attr<int>("regular_type");

    auto p_e = framework::EigenVector<T>::Flatten(*p);
    auto g_e = framework::EigenVector<T>::Flatten(*g);
    auto grad_out_e = framework::EigenVector<T>::Flatten(*grad_out);

    auto& dev_ctx = ctx.template device_context<DeviceContext>();
    auto& eigen_ctx = *dev_ctx.eigen_device();

    // NOTE. In paddle, loss has divided by nranks. Because dgc_op is before
    // allreduce, so local regular_coeff need div nranks too. But now we
    // multi grad with nranks in dgc_op, in that case regular_coeff don't
    // need to /nranks, can prevent precision loss. For coeff often equal
    // with 1e-4, if nranks=32, coeff/nranks will be 3.125e-6, the numerical
    // accuracy of coeff/nranks will be too low.
    PADDLE_ENFORCE_EQ(regular_type >= 0 && regular_type <= 2, true,
                      platform::errors::InvalidArgument(
                          "DGC only support one of None|L1Decay|L2Decay "
                          "Regularization for now."));
    if (regular_type == 0) {
      grad_out_e.device(eigen_ctx) = (1.0 * nranks) * g_e;
    } else if (regular_type == 1) {
      // L1Decay. grad = grad + coeff * sign(param)
      grad_out_e.device(eigen_ctx) =
          (1.0 * nranks) * g_e + regular_coeff * p_e.sign();
    } else if (regular_type == 2) {
      // L2Decay. grad = grad + coeff * param
      grad_out_e.device(eigen_ctx) = (1.0 * nranks) * g_e + regular_coeff * p_e;
    }

106 107 108 109 110 111 112 113 114 115 116 117
    // current step
    auto current_step_tensor = ctx.Input<framework::Tensor>("current_step");
    const float* current_step = current_step_tensor->data<float>();

    if (static_cast<int>(*current_step) < static_cast<int>(rampup_begin_step)) {
      VLOG(10) << "current_step:" << *current_step
               << " < rampup_begin_step:" << rampup_begin_step
               << " so does't use dgc";
      return;
    }

    float ratio =
118 119 120
        1 - get_period_sparcity(
                sparsity, static_cast<float>(*current_step - rampup_begin_step),
                rampup_step);
121 122 123 124
    PADDLE_ENFORCE_GE(ratio, 0.0, platform::errors::InvalidArgument(
                                      "DGC sparsity ratio must >= 0"));
    PADDLE_ENFORCE_LT(ratio, 1.0, platform::errors::InvalidArgument(
                                      "DGC sparsity ratio must < 1"));
125 126 127 128 129 130
    int k = static_cast<int>(g->numel() * ratio);

    VLOG(10) << "m:" << m << ", use_nesterov:" << use_nesterov
             << ", rampup_begin_step:" << rampup_begin_step
             << ", rampup_step:" << rampup_step
             << ",  current_step:" << *current_step << ", ratio:" << ratio
131
             << ", k:" << k << ", nranks:" << nranks;
132 133 134 135 136 137 138 139

    auto k_out = ctx.Output<framework::Tensor>("k");
    T* k_out_data = k_out->data<T>();
    *k_out_data = k;

    auto u_out = ctx.Output<framework::Tensor>("U_out");
    auto v_out = ctx.Output<framework::Tensor>("V_out");
    auto encode_grad_out = ctx.Output<framework::Tensor>("EncodeGrad");
140
    auto gather_buff = ctx.Output<framework::Tensor>("GatherBuff");
141 142 143 144

    // FIXME(gongwb): use cublas.
    auto u_out_e = framework::EigenVector<T>::Flatten(*u_out);
    auto u_e = framework::EigenVector<T>::Flatten(*u);
145

146 147 148 149 150 151
    // calc local momentum from global momentum
    // NOTE. If grad not multi nranks, need add below code.
    // if (static_cast<int>(*current_step) ==
    //     static_cast<int>(rampup_begin_step)) {
    //   u_out_e.device(eigen_ctx) = (1.0 / nranks) * u_e;
    // }
152

153 154
    if (use_nesterov) {
      // u = m * (u + g)
155
      u_out_e.device(eigen_ctx) = m * (u_e + grad_out_e);
156 157

      // v = u + v + g
158 159
      ElementwiseComputeEx<phi::funcs::AddFunctor<T>, DeviceContext, T>(
          ctx, u, v, 0, phi::funcs::AddFunctor<T>(), v_out);
160

161 162
      ElementwiseComputeEx<phi::funcs::AddFunctor<T>, DeviceContext, T>(
          ctx, g, v, 0, phi::funcs::AddFunctor<T>(), v_out);
163 164
    } else {
      // u = m * u + g
165
      u_out_e.device(eigen_ctx) = m * u_e + grad_out_e;
166 167

      // v = u + v
168 169
      ElementwiseComputeEx<phi::funcs::AddFunctor<T>, DeviceContext, T>(
          ctx, u, v, 0, phi::funcs::AddFunctor<T>(), v_out);
170 171 172 173 174 175
    }

    T* v_out_data = v_out->mutable_data<T>(ctx.GetPlace());
    T* u_out_data = u_out->mutable_data<T>(ctx.GetPlace());
    T* encode_grad_out_data = encode_grad_out->mutable_data<T>(
        framework::DDim{2 * k}, ctx.GetPlace());
176 177
    gather_buff->mutable_data<T>(framework::DDim{2 * k * nranks},
                                 ctx.GetPlace());
178 179

    int buf_size = paddle::communication::dgc::get_buffer_size(k);
180
    auto tmp_ious_data = memory::Alloc(dev_ctx, buf_size);
181 182 183 184 185 186
    void* buf = reinterpret_cast<void*>(tmp_ious_data->ptr());

    if (!paddle::communication::dgc::k_select(
            static_cast<void*>(encode_grad_out_data), k, v_out_data,
            static_cast<int>(v_out->numel()), buf, dev_ctx.stream(),
            u_out_data)) {
187 188 189
      // TODO(weihang): owner should polish this error message
      PADDLE_THROW(platform::errors::InvalidArgument(
          "V_out numel error, V_out numel is %d.", v_out->numel()));
190 191
    }

192
    phi::funcs::SetConstant<DeviceContext, T> tset;
193 194 195 196 197
    tset(dev_ctx, grad_out, static_cast<T>(0));
  }
};
}  // namespace operators
}  // namespace paddle