mkldnn_helper.h 10.6 KB
Newer Older
1
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved.
T
tensor-tang 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15

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

16
#include <mkldnn.h>
G
gongweibao 已提交
17
#include <string>
18
#include <vector>
19
#include "paddle/fluid/framework/operator.h"
M
mozga-intel 已提交
20
#include "paddle/fluid/platform/place.h"
21

T
tensor-tang 已提交
22 23 24 25 26 27
namespace paddle {
namespace platform {

using MKLDNNStream = mkldnn::stream;
using MKLDNNEngine = mkldnn::engine;
using MKLDNNMemory = mkldnn::memory;
28
using MKLDNNMemoryDescriptor = mkldnn::memory::desc;
T
tensor-tang 已提交
29 30 31
using MKLDNNPrimitive = mkldnn::primitive;
using MKLDNNPrimitiveDesc = mkldnn::handle<mkldnn_primitive_desc_t>;

32 33 34 35 36
typedef std::unique_ptr<MKLDNNStream> MKLDNNStreamPtr;
typedef std::unique_ptr<MKLDNNEngine> MKLDNNEnginePtr;
typedef std::unique_ptr<MKLDNNMemory> MKLDNNMemoryPtr;
typedef std::unique_ptr<MKLDNNPrimitive> MKLDNNPrimitivePtr;
typedef std::unique_ptr<MKLDNNPrimitiveDesc> MKLDNNPrimitiveDescPtr;
T
tensor-tang 已提交
37

38 39 40 41 42
template <typename Type>
void* to_void_cast(const Type* t) {
  return static_cast<void*>(const_cast<Type*>(t));
}

K
Krzysztof Binias 已提交
43 44 45 46 47
template <typename Type>
void* to_void_reinterpret_cast(const Type* t) {
  return reinterpret_cast<void*>(const_cast<Type*>(t));
}

48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68
template <class Type>
using tf_desc = typename Type::desc;

template <class Type>
using tf_pd = typename Type::primitive_desc;

template <typename Type, typename Engine, typename... Args>
std::shared_ptr<tf_pd<Type>> MKLDNNFwdPrimitiveDesc(const Engine& e,
                                                    Args&&... args) {
  auto desc = tf_desc<Type>(mkldnn::prop_kind::forward, (args)...);
  auto pd = new tf_pd<Type>(desc, e);
  return std::shared_ptr<tf_pd<Type>>(pd);
}

template <typename Type, typename Engine, typename Primitive, typename... Args>
tf_pd<Type> MKLDNNBwdPrimitiveDesc(const Engine& e, const Primitive& p,
                                   Args&&... args) {
  auto desc = tf_desc<Type>(args...);
  return tf_pd<Type>(desc, e, p);
}

69 70 71 72
inline mkldnn::memory::desc MKLDNNMemDesc(const std::vector<int>& dims,
                                          mkldnn::memory::data_type data_type,
                                          mkldnn::memory::format format) {
  mkldnn::memory::dims tz = dims;
X
xiaolil1 已提交
73
  std::cout<<"this is MKLDNNMemDesc"<<"   data_type"<<data_type<<"   format"<<format<<std::endl;
74 75 76 77 78 79 80 81
  return mkldnn::memory::desc({tz}, data_type, format);
}

inline bool CanMKLDNNBeUsed(const framework::ExecutionContext& ctx) {
  bool use_mkldnn = ctx.Attr<bool>("use_mkldnn");
  return use_mkldnn && platform::is_cpu_place(ctx.GetPlace());
}

82 83 84 85 86 87 88 89 90 91
template <typename Type>
mkldnn::memory::data_type MKLDNNGetDataType() {
  return mkldnn::memory::data_undef;
}

template <>
inline mkldnn::memory::data_type MKLDNNGetDataType<float>() {
  return mkldnn::memory::f32;
}

M
mozga-intel 已提交
92 93 94 95 96 97 98 99 100 101 102 103
inline void Reorder(const mkldnn::memory& src, const mkldnn::memory& dst) {
  auto reorder_prim = mkldnn::reorder(src, dst);
  std::vector<mkldnn::primitive> pipeline;
  pipeline.push_back(reorder_prim);
  mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
}

inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) {
  return static_cast<mkldnn::memory::format>(
      memory.get_primitive_desc().desc().data.format);
}

104 105 106 107 108 109
inline mkldnn::memory::format GetMKLDNNFormat(
    const mkldnn::sum::primitive_desc& memory) {
  return static_cast<mkldnn::memory::format>(
      memory.dst_primitive_desc().desc().data.format);
}

J
Jacek Czaja 已提交
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128
class MKLDNNHandler {
 public:
  MKLDNNHandler(const MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
                const std::string& base_key)
      : dev_ctx_(dev_ctx),
        engine_(engine),
        key_(base_key),
        is_reusing_(false) {}

  std::shared_ptr<mkldnn::memory> AcquireSrcMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_src_mem_p");
  }

  std::shared_ptr<mkldnn::memory> AcquireWeightsMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_weights_mem_p");
  }

129 130 131 132 133
  std::shared_ptr<mkldnn::memory> AcquireBiasMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_bias_mem_p");
  }

J
Jacek Czaja 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
  std::shared_ptr<mkldnn::memory> AcquireDstMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_dst_mem_p");
  }

  std::shared_ptr<mkldnn::memory> AcquireDiffDstMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_diff_dst_mem_p");
  }

  std::shared_ptr<mkldnn::memory> AcquireDiffSrcMemory(
      const mkldnn::memory::desc& md, void* ptr) {
    return this->AcquireMemory(md, ptr, "@user_diff_src_mem_p");
  }

  std::shared_ptr<mkldnn::memory> AcquireMemoryFromPrimitive(
      mkldnn::memory::primitive_desc mdp, void* ptr,
      const std::string& suffix) {
    auto local_key = key_ + suffix;
    auto mem_p =
        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
                   "Fail to find mem primitive in device context");
157
    //mem_p = nullptr;
J
Jacek Czaja 已提交
158 159
    if (mem_p == nullptr) {
      mem_p = std::make_shared<mkldnn::memory>(mdp, ptr);
160 161
std::cout<<"mem_p == null"<<std::endl;
//std::cout<<"mdp fmt = "<<mdp.desc().data.format<<"   mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
J
Jacek Czaja 已提交
162 163 164 165 166
      dev_ctx_.SetBlob(local_key, mem_p);
    } else {
      mem_p->set_data_handle(ptr);
      // Mark that reusing happenned. All primitives from operator instance
      // should be reused or none of them. So we check consistency
X
xiaolil1 已提交
167
std::cout<<"1 is reuse = "<<is_reusing_;
J
Jacek Czaja 已提交
168 169
      is_reusing_ = true;
    }
170
std::cout<<"mdp fmt = "<<mdp.desc().data.format<<"   mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
J
Jacek Czaja 已提交
171 172 173 174 175 176 177 178 179 180 181 182
    return mem_p;
  }

  std::shared_ptr<mkldnn::memory> AcquireMemory(const mkldnn::memory::desc& md,
                                                void* ptr,
                                                const std::string& suffix) {
    /*Generate key*/
    auto local_key = key_ + suffix;
    auto mem_p =
        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
    PADDLE_ENFORCE((mem_p != nullptr) || (is_reusing_ == false),
                   "Fail to find mem primitive in device context");
183
    //mem_p = nullptr;
J
Jacek Czaja 已提交
184
    if (mem_p == nullptr) {
185
std::cout<<"mem_p == null"<<std::endl;
J
Jacek Czaja 已提交
186 187 188 189 190 191 192
      mem_p = std::make_shared<mkldnn::memory>(
          mkldnn::memory::primitive_desc{md, engine_}, ptr);
      dev_ctx_.SetBlob(local_key, mem_p);
    } else {
      mem_p->set_data_handle(ptr);
      // Mark that reusing happenned. All primitives from operator instance
      // should be reused or none of them. So we check consistency
X
xiaolil1 已提交
193
std::cout<<"2 is reuse = "<<is_reusing_;
J
Jacek Czaja 已提交
194 195
      is_reusing_ = true;
    }
196
std::cout<<"md fmt = "<<md.data.format<<"   mem_p fmt = "<<mem_p->get_primitive_desc().desc().data.format<<std::endl;
J
Jacek Czaja 已提交
197 198 199 200
    return mem_p;
  }

  std::shared_ptr<mkldnn::memory> AcquireMemory(
G
gongweibao 已提交
201 202
      mkldnn::memory::primitive_desc& mpd,       // NOLINT
      mkldnn::memory::primitive_desc& user_mpd,  // NOLINT
J
Jacek Czaja 已提交
203
      const std::shared_ptr<mkldnn::memory> user_memory_p,
K
Krzysztof Binias 已提交
204 205
      const std::string& suffix,
      std::vector<mkldnn::primitive>& pipeline,  // NOLINT
X
xiaolil1 已提交
206 207 208 209
      bool is_persistent = false,
      bool is_INT8 = false,
      std::vector<float> scale_data = {1.0f},
      int mask = 0) {
J
Jacek Czaja 已提交
210 211 212 213 214 215 216 217 218 219 220 221 222
    // create reorder primitive if the input format is not the preferred one
    auto local_key = key_ + suffix;
    auto key_reorder_p = key_ + suffix + "reorder_p";

    auto target_memory_p =
        std::static_pointer_cast<mkldnn::memory>(dev_ctx_.GetBlob(local_key));
    PADDLE_ENFORCE((target_memory_p != nullptr) || (is_reusing_ == false),
                   "Fail to find mem primitive in device context");
    if (target_memory_p == nullptr) {
      target_memory_p = user_memory_p;
      std::shared_ptr<mkldnn::primitive> reorder_p;
      if (mpd != user_mpd) {
        target_memory_p = std::make_shared<mkldnn::memory>(mpd);
X
xiaolil1 已提交
223 224
        auto reorder_p =
            std::make_shared<mkldnn::reorder>(*user_memory_p, *target_memory_p);
X
xiaolil1 已提交
225 226 227 228 229 230 231 232 233
        if(is_INT8){
            mkldnn::primitive_attr attri;
            attri.set_output_scales(mask, scale_data);

            auto reorder_pd = std::shared_ptr<mkldnn::reorder::primitive_desc>(
                    new mkldnn::reorder::primitive_desc(mpd, user_mpd, attri));
            auto reorder_p =
                std::shared_ptr<mkldnn::reorder>(new mkldnn::reorder(*reorder_pd, *user_memory_p, *target_memory_p));
        }
J
Jacek Czaja 已提交
234 235 236 237
        dev_ctx_.SetBlob(key_reorder_p, reorder_p);
        pipeline.push_back(*reorder_p);
      }
      dev_ctx_.SetBlob(local_key, target_memory_p);
K
Krzysztof Binias 已提交
238
    } else if (!is_persistent) {
J
Jacek Czaja 已提交
239 240 241 242 243 244
      // Make reorder if needed
      auto reorder_p = std::static_pointer_cast<mkldnn::reorder>(
          dev_ctx_.GetBlob(key_reorder_p));
      if (reorder_p != nullptr) {
        pipeline.push_back(*reorder_p);
      }
X
xiaolil1 已提交
245
std::cout<<"3 is reuse = "<<is_reusing_;
J
Jacek Czaja 已提交
246 247 248 249 250
      is_reusing_ = true;
    }
    return target_memory_p;
  }

G
gongweibao 已提交
251
  static std::string GetHash(mkldnn::memory::dims& operand_dims,  // NOLINT
J
Jacek Czaja 已提交
252 253
                             const std::string& suffix) {
    return dims2str(operand_dims) + suffix;
254
  }
255 256 257 258 259 260 261 262

 protected:
  static std::string dims2str(const mkldnn::memory::dims& operand_dims) {
    std::string dstr = "";
    for (size_t i = 0; i < operand_dims.size(); ++i) {
      dstr += std::to_string(operand_dims[i]) + "-";
    }
    return dstr;
263
  }
J
Jacek Czaja 已提交
264 265 266 267 268 269 270 271

 protected:
  const MKLDNNDeviceContext& dev_ctx_;
  mkldnn::engine engine_;
  std::string key_;
  bool is_reusing_;
};

272 273 274 275 276 277 278 279 280 281
inline mkldnn::memory::format MKLDNNFormatForSize(
    size_t dims_size, mkldnn::memory::format data_format) {
  if (dims_size == 1) {
    return mkldnn::memory::format::x;
  } else if (dims_size == 2) {
    return mkldnn::memory::format::nc;
  }
  return data_format;
}

282 283 284 285 286 287 288 289 290 291 292 293
inline mkldnn::memory::format data_format_to_memory_format(
    const std::string& data_format) {
  switch (framework::StringToDataLayout(data_format)) {
    case framework::DataLayout::kNHWC:
      return mkldnn::memory::format::nhwc;
    case framework::DataLayout::kNCHW:
      return mkldnn::memory::format::nchw;
    default:
      return mkldnn::memory::format::any;
  }
}

T
tensor-tang 已提交
294 295
}  // namespace platform
}  // namespace paddle