mul_mkldnn_op.cc 16.1 KB
Newer Older
P
Physher 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64
/* 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 <string>
#include <vector>
#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/operators/mul_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
#include "paddle/fluid/platform/mkldnn_reuse.h"

namespace paddle {
namespace operators {

using framework::DataLayout;
using framework::DDim;
using framework::ExecutionContext;
using framework::Tensor;
using mkldnn::inner_product_forward;
using mkldnn::memory;
using mkldnn::prop_kind;
using mkldnn::stream;
using platform::MKLDNNDeviceContext;
using platform::to_void_cast;

template <typename XT, typename YT, typename OT>
class MulPrimitiveFactory {
 public:
  explicit MulPrimitiveFactory(const mkldnn::engine &engine)
      : engine_(engine) {}

  virtual ~MulPrimitiveFactory() {}

  virtual inner_product_forward CreateMulPrimitive(
      const Tensor *input_x, const Tensor *input_y, Tensor *output,
      const ExecutionContext &ctx) {
    /* check format and reorder if need */
    int x_num_col_dims = ctx.Attr<int>("x_num_col_dims");
    int y_num_col_dims = ctx.Attr<int>("y_num_col_dims");

    auto x_matrix = UpdateDataFormat<XT>(input_x, x_num_col_dims, ctx);
    auto y_matrix = UpdateDataFormat<YT>(input_y, y_num_col_dims, ctx);

    auto output_dim = output->dims();
    if (output_dim.size() != 2) {
      output->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
    }

    if (mul_) {
      UpdateDataPointers(ctx, output, &x_matrix);
      return *mul_;
    }

65
    auto src_desc = CreateMemDescriptor<XT>(&x_matrix, MKLDNNMemoryFormat::nc);
P
Physher 已提交
66 67
    x_input_ = CreateMemory<XT>(src_desc, &x_matrix);
    y_input_ = TransposeInputY(&y_matrix);
68
    auto dst_desc = CreateMemDescriptor<OT>(output, MKLDNNMemoryFormat::any);
P
Physher 已提交
69 70 71 72 73 74 75 76 77 78 79

    mul_ = CreateMulPrimitive(*x_input_, *y_input_, dst_desc, output, ctx);
    return *mul_;
  }

 protected:
  template <typename T>
  Tensor UpdateDataFormat(const Tensor *data, int num_col_dims,
                          const ExecutionContext &ctx) {
    Tensor x_tmp;
    Tensor data_matrix;
80 81
    MKLDNNMemoryFormat src_fmt = data->format();
    MKLDNNMemoryFormat dst_fmt;
P
Physher 已提交
82 83 84
    auto src_mdesc = CreateMemDescriptor<T>(data, src_fmt);

    if ((data->dims().size() == 4 &&
85
         src_fmt != (dst_fmt = MKLDNNMemoryFormat::nchw)) ||
P
Physher 已提交
86
        (data->dims().size() == 5 &&
87
         src_fmt != (dst_fmt = MKLDNNMemoryFormat::ncdhw))) {
P
Physher 已提交
88 89 90 91 92 93 94
      auto dst_mdesc = CreateMemDescriptor<T>(data, dst_fmt);
      x_tmp.mutable_data<T>(ctx.GetPlace(), data->memory_size());

      Reorder(src_mdesc, dst_mdesc, to_void_cast<T>(data->data<T>()),
              to_void_cast<T>(x_tmp.data<T>()));

      x_tmp.Resize(data->dims());
95
      x_tmp.set_format((MKLDNNMemoryFormat)dst_mdesc.data.format);
P
Physher 已提交
96 97 98 99 100 101 102 103 104 105 106 107 108
      data_matrix = framework::ReshapeToMatrix(x_tmp, num_col_dims);
    } else {
      data_matrix = framework::ReshapeToMatrix(*data, num_col_dims);
    }

    return data_matrix;
  }

  void UpdateDataPointers(const ExecutionContext &ctx, Tensor *out,
                          const Tensor *in) {
    x_input_->set_data_handle(to_void_cast<XT>(in->data<XT>()));
    output_->set_data_handle(out->mutable_data<OT>(ctx.GetPlace()));

109
    if (out->format() == MKLDNNMemoryFormat::format_undef) {
A
Adam 已提交
110
      auto output_format = platform::GetMKLDNNFormat(*output_);
111
      out->set_format((MKLDNNMemoryFormat)output_format);
P
Physher 已提交
112 113 114 115 116
    }
  }

  template <typename T>
  memory::desc CreateMemDescriptor(
117
      const Tensor *tensor, MKLDNNMemoryFormat format,
P
Physher 已提交
118
      memory::data_type type = platform::MKLDNNGetDataType<T>()) {
119
    auto dims = framework::vectorize<int>(tensor->dims());
P
Physher 已提交
120 121 122 123 124
    return platform::MKLDNNMemDesc(dims, type, format);
  }

  template <typename T>
  memory::desc CreateMemDescriptor(
125
      const std::vector<int> &dims, MKLDNNMemoryFormat format,
P
Physher 已提交
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141
      memory::data_type type = platform::MKLDNNGetDataType<T>()) {
    return platform::MKLDNNMemDesc(dims, type, format);
  }

  template <typename T>
  memory CreateMemory(const memory::desc &desc, const Tensor *tensor) {
    return memory({desc, engine_}, to_void_cast<T>(tensor->data<T>()));
  }

  memory CreateDstMemory(
      const inner_product_forward::primitive_desc &mul_prim_desc,
      const ExecutionContext &ctx, Tensor *output) {
    auto dst_prim_desc = mul_prim_desc.dst_primitive_desc();
    auto buffer_size = dst_prim_desc.get_size();

    OT *output_data = output->mutable_data<OT>(ctx.GetPlace(), buffer_size);
A
Adam 已提交
142 143 144
    memory dst_mem(dst_prim_desc, to_void_cast<OT>(output_data));
    output->set_format(platform::GetMKLDNNFormat(dst_mem));
    return dst_mem;
P
Physher 已提交
145 146 147 148 149 150 151 152 153 154 155 156 157 158 159
  }

  memory Reorder(const memory::desc &src_desc, const memory::desc &dst_desc,
                 void *src_data, void *dst_data = NULL) {
    auto src_mem = memory({src_desc, engine_}, src_data);
    auto dst_mem = dst_data ? memory({dst_desc, engine_}, dst_data)
                            : memory({dst_desc, engine_});

    auto reorder = mkldnn::reorder(src_mem, dst_mem);
    stream(stream::kind::eager).submit({reorder}).wait();

    return dst_mem;
  }

  memory TransposeInputY(const Tensor *input_y) {
160
    auto dims = framework::vectorize<int>(input_y->dims());
P
Physher 已提交
161
    std::swap(dims[0], dims[1]);  // Correct output dimensions
162 163
    auto src_desc = CreateMemDescriptor<YT>(dims, MKLDNNMemoryFormat::io);
    auto dst_desc = CreateMemDescriptor<YT>(dims, MKLDNNMemoryFormat::oi);
P
Physher 已提交
164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210
    return Reorder(src_desc, dst_desc, to_void_cast<YT>(input_y->data<YT>()));
  }

  inner_product_forward CreateMulPrimitive(const memory &x_memory,
                                           const memory &y_memory,
                                           const memory::desc &dst_desc,
                                           Tensor *output,
                                           const ExecutionContext &ctx) {
    const auto y_desc = y_memory.get_primitive_desc().desc();
    const auto x_desc = x_memory.get_primitive_desc().desc();

    auto mul_prim_desc = CreateMulPrimDesc(x_desc, y_desc, dst_desc);
    output_ = CreateDstMemory(mul_prim_desc, ctx, output);

    return inner_product_forward(mul_prim_desc, x_memory, y_memory, *output_);
  }

  inner_product_forward::primitive_desc CreateMulPrimDesc(
      const memory::desc &x_desc, const memory::desc &y_desc,
      const memory::desc &dst_desc) {
    auto mul_desc = inner_product_forward::desc(prop_kind::forward, x_desc,
                                                y_desc, dst_desc);

    return inner_product_forward::primitive_desc(mul_desc, engine_);
  }

 protected:
  const mkldnn::engine &engine_;
  boost::optional<memory> x_input_;
  boost::optional<memory> y_input_;
  boost::optional<memory> output_;
  boost::optional<inner_product_forward> mul_;
};  // namespace operators

template <typename XT, typename YT, typename OT>
class QuantMulPrimitiveFactory : public MulPrimitiveFactory<XT, YT, OT> {
 public:
  using MulPrimitiveFactory<XT, YT, OT>::MulPrimitiveFactory;

  virtual inner_product_forward CreateMulPrimitive(
      const Tensor *x_input, const Tensor *y_input, Tensor *output,
      const ExecutionContext &ctx) {
    /* check data format and reorder if need */
    int x_num_col_dims = ctx.Attr<int>("x_num_col_dims");
    int y_num_col_dims = ctx.Attr<int>("y_num_col_dims");
    auto scale_y = ctx.Attr<std::vector<float>>("scale_y");

211 212 213 214 215 216 217 218
    // TODO(intel-minghui) : Remove the restriction that only supports Input(Y)
    // as weights
    bool enforce = std::is_same<YT, float>::value;
    PADDLE_ENFORCE(
        enforce == true,
        "Input(Y) supposed to be fp32 data type since only fp32 data type is "
        "supported in the current design of MKLDNN INT8.");

P
Physher 已提交
219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
    auto x_matrix =
        this->template UpdateDataFormat<XT>(x_input, x_num_col_dims, ctx);
    auto y_matrix =
        this->template UpdateDataFormat<YT>(y_input, y_num_col_dims, ctx);

    auto output_dim = output->dims();
    if (output_dim.size() != 2) {
      output->Resize({x_matrix.dims()[0], y_matrix.dims()[1]});
    }

    if (this->mul_) {
      this->UpdateDataPointers(ctx, output, &x_matrix);
      return *(this->mul_);
    }

234 235
    auto src_desc = this->template CreateMemDescriptor<XT>(
        &x_matrix, MKLDNNMemoryFormat::nc);
P
Physher 已提交
236 237 238 239 240 241
    this->x_input_ = this->template CreateMemory<XT>(src_desc, &x_matrix);

    const auto trans_y = this->TransposeInputY(&y_matrix);
    this->y_input_ = QuantInputY(trans_y, scale_y);

    auto dst_desc =
242
        this->template CreateMemDescriptor<OT>(output, MKLDNNMemoryFormat::any);
P
Physher 已提交
243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273

    this->mul_ = CreateMulPrimitive(*(this->x_input_), *(this->y_input_),
                                    dst_desc, output, ctx);
    return *(this->mul_);
  }

  memory ReorderWithScale(const memory::desc &src_desc,
                          const memory::desc &dst_desc, void *src_data,
                          const std::vector<float> &scale) {
    auto mask = scale.size() > 1 ? 1 : 0;
    mkldnn::primitive_attr attr;
    attr.set_output_scales(mask, scale);

    auto src_mem = memory({src_desc, this->engine_}, src_data);
    auto dst_mem = memory({dst_desc, this->engine_});

    auto reorder_pd = mkldnn::reorder::primitive_desc(
        src_mem.get_primitive_desc(), dst_mem.get_primitive_desc(), attr);

    auto reorder = mkldnn::reorder(reorder_pd, src_mem, dst_mem);
    stream(stream::kind::eager).submit({reorder}).wait();

    return dst_mem;
  }

  memory QuantInputY(memory input_y, const std::vector<float> &scale_y) {
    const auto &dims = input_y.get_primitive_desc().desc().data.dims;
    auto ndims = input_y.get_primitive_desc().desc().data.ndims;
    auto y_dims = std::vector<int>(dims, dims + ndims);

    auto user_y_desc =
274 275 276
        this->template CreateMemDescriptor<YT>(y_dims, MKLDNNMemoryFormat::oi);
    auto y_desc = this->template CreateMemDescriptor<int8_t>(
        y_dims, MKLDNNMemoryFormat::oi);
P
Physher 已提交
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341

    return ReorderWithScale(user_y_desc, y_desc, input_y.get_data_handle(),
                            scale_y);
  }

  mkldnn::primitive_attr CreateMulAttr(const ExecutionContext &ctx,
                                       bool force_fp32_output) {
    mkldnn::primitive_attr mul_attr;

    auto scale_y_data = ctx.Attr<std::vector<float>>("scale_y");
    auto scale_x_data = ctx.Attr<float>("scale_x");
    auto scale_out_data =
        force_fp32_output ? 1.0f : ctx.Attr<float>("scale_out");

    bool is_multi_channel = scale_y_data.size() > 1;
    int count = is_multi_channel ? scale_y_data.size() : 1;
    std::vector<float> output_shift_scale(count);
    for (int i = 0; i < count; i++) {
      if (scale_y_data[i] == 0.0)
        output_shift_scale[i] = scale_out_data;
      else
        output_shift_scale[i] =
            scale_out_data / (scale_x_data * scale_y_data[i]);
    }
    int mul_mask = is_multi_channel ? 1 : 0;
    mul_attr.set_output_scales(mul_mask, output_shift_scale);

    return mul_attr;
  }

  inner_product_forward CreateMulPrimitive(const memory &x_memory,
                                           const memory &y_memory,
                                           const memory::desc &dst_desc,
                                           Tensor *output,
                                           const ExecutionContext &ctx) {
    const auto x_desc = x_memory.get_primitive_desc().desc();
    const auto y_desc = y_memory.get_primitive_desc().desc();
    bool force_fp32_output = ctx.Attr<bool>("force_fp32_output");

    mkldnn::primitive_attr mul_attr = CreateMulAttr(ctx, force_fp32_output);
    auto mul_prim_desc = CreateMulPrimDesc(x_desc, y_desc, dst_desc, mul_attr);

    this->output_ = this->CreateDstMemory(mul_prim_desc, ctx, output);

    return inner_product_forward(mul_prim_desc, x_memory, y_memory,
                                 *(this->output_));
  }

  inner_product_forward::primitive_desc CreateMulPrimDesc(
      const memory::desc &x_desc, const memory::desc &y_desc,
      const memory::desc &dst_desc, const mkldnn::primitive_attr &mul_attr) {
    const auto &mul_desc = inner_product_forward::desc(
        prop_kind::forward, x_desc, y_desc, dst_desc);

    return inner_product_forward::primitive_desc(mul_desc, mul_attr,
                                                 this->engine_);
  }
};

/* OT: output data type */
template <typename XT, typename YT, typename OT>
std::shared_ptr<MulPrimitiveFactory<XT, YT, OT>> GetPrimitiveFactory(
    const MKLDNNDeviceContext &dev_ctx, const ExecutionContext &ctx,
    const Tensor *input_x, const Tensor *input_y,
    const mkldnn::engine &mkldnn_engine, bool enable_quant) {
342 343 344 345 346
  const std::string key = platform::CreateKey(
      input_x->format(), input_x->type(),
      framework::vectorize<int>(input_x->dims()), input_y->format(),
      input_y->type(), framework::vectorize<int>(input_y->dims()),
      ctx.op().Output("Out"));
P
Physher 已提交
347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408

  auto prim_creator = std::static_pointer_cast<MulPrimitiveFactory<XT, YT, OT>>(
      dev_ctx.GetBlob(key));

  if (prim_creator == nullptr) {
    prim_creator =
        enable_quant
            ? std::make_shared<QuantMulPrimitiveFactory<XT, YT, OT>>(
                  mkldnn_engine)
            : std::make_shared<MulPrimitiveFactory<XT, YT, OT>>(mkldnn_engine);
    dev_ctx.SetBlob(key, prim_creator);
  }

  return prim_creator;
}

template <typename XT, typename YT>
inner_product_forward GetMulPrimitive(const MKLDNNDeviceContext &dev_ctx,
                                      const ExecutionContext &ctx,
                                      const Tensor *input_x,
                                      const Tensor *input_y, Tensor *output,
                                      const mkldnn::engine &mkldnn_engine) {
  bool enable_quant =
      std::is_same<XT, int8_t>::value || std::is_same<XT, uint8_t>::value;
  bool force_fp32_output = ctx.Attr<bool>("force_fp32_output");

  if (enable_quant && !force_fp32_output) {
    return GetPrimitiveFactory<XT, YT, int8_t>(dev_ctx, ctx, input_x, input_y,
                                               mkldnn_engine, enable_quant)
        ->CreateMulPrimitive(input_x, input_y, output, ctx);

  } else {
    return GetPrimitiveFactory<XT, YT, float>(dev_ctx, ctx, input_x, input_y,
                                              mkldnn_engine, enable_quant)
        ->CreateMulPrimitive(input_x, input_y, output, ctx);
  }
}

/* XT: input x data type, YT: input y data type */
template <typename XT, typename YT>
class MulMKLDNNKernel : public framework::OpKernel<XT> {
 public:
  void Compute(const ExecutionContext &ctx) const override {
    PADDLE_ENFORCE(platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

    auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
    const auto &mkldnn_engine = dev_ctx.GetEngine();

    const Tensor *x = ctx.Input<Tensor>("X");
    const Tensor *y = ctx.Input<Tensor>("Y");
    Tensor *out = ctx.Output<Tensor>("Out");
    auto out_dims = out->dims();

    auto mul = GetMulPrimitive<XT, YT>(dev_ctx, ctx, x, y, out, mkldnn_engine);

    stream(stream::kind::eager).submit({mul}).wait();

    if (out_dims.size() != 2) {
      out->Resize(out_dims);
    }
    out->set_layout(DataLayout::kMKLDNN);
409 410
    out->set_format(platform::MKLDNNFormatForSize(
        out_dims.size(), mkldnn::memory::format::nchw));
P
Physher 已提交
411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(mul, MKLDNN, ::paddle::platform::CPUPlace,
                                    U8, ops::kMULMKLDNNINT8,
                                    ops::MulMKLDNNKernel<uint8_t, float>);

REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(mul, MKLDNN, ::paddle::platform::CPUPlace,
                                    S8, ops::kMULMKLDNNINT8,
                                    ops::MulMKLDNNKernel<int8_t, float>);

REGISTER_OP_KERNEL(mul, MKLDNN, ::paddle::platform::CPUPlace,
                   ops::MulMKLDNNKernel<uint8_t, float>);