GemmConvOp.cpp 17.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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. */

15
#include "ConvOp.h"
16
#include "GemmFunctor.h"
17
#include "Im2Col.h"
18 19 20 21 22
#include "paddle/math/MemoryHandle.h"

namespace paddle {

/*
23
 * \brief Forward calculation of convolution.
24 25 26 27 28 29 30 31
 */
template <DeviceType Device>
class GemmConvFunction : public ConvFunctionBase {
public:
  void init(const FuncConfig& config) override {
    ConvFunctionBase::init(config);
  }

L
liaogang 已提交
32
  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
H
hedaoyuan 已提交
33 34 35 36 37 38
    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();
    checkShape(input, filter, output);
  }

39
  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
40 41
    CHECK_EQ(numInputs_, inputs.size());
    CHECK_EQ(numOutputs_, outputs.size());
H
hedaoyuan 已提交
42
    check(inputs, outputs);
43 44 45 46 47 48 49 50 51 52 53 54
    // TODO(hedaoyuan): Need to define some index macros,
    // to avoid useing 0 and 1.
    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();

    real beta;
    if (outputs[0].getArgType() == ADD_TO) {
      beta = 1.0;
    } else {
      beta = 0.0;
    }
55

H
hedaoyuan 已提交
56 57 58 59 60 61 62 63 64
    size_t batchSize = input[0];
    size_t inputChannels = input[1];
    size_t inputHeight = input[2];
    size_t inputWidth = input[3];
    size_t filterHeight = getFilterHeight(filter);
    size_t filterWidth = getFilterWidth(filter);
    size_t outputChannels = output[1];
    size_t outputHeight = output[2];
    size_t outputWidth = output[3];
65 66 67 68

    real* inputData = inputs[0].data<real>();
    real* filterData = inputs[1].data<real>();
    real* outputData = outputs[0].data<real>();
69
    bool needIm2col = isNeedIm2col(filter);
70

71 72 73
    TensorShape imShape =
        TensorShape({inputChannels / groups_, inputHeight, inputWidth});

74
    TensorShape colShape;
75
    real* colData = NULL;
76

77
    if (needIm2col) {
78 79 80 81 82 83 84 85
      colShape = TensorShape({inputChannels / groups_,
                              filterHeight,
                              filterWidth,
                              outputHeight,
                              outputWidth});
      resizeBuffer<Device>(colShape.getElements());
      colData = reinterpret_cast<real*>(memory_->getBuf());
    }
86

87 88
    Im2ColFunctor<kCFO, Device, real> im2col;
    size_t inputOffset = imShape.getElements();
89 90
    size_t outputOffset =
        (outputChannels / groups_) * outputHeight * outputWidth;
H
hedaoyuan 已提交
91 92
    size_t filterOffset = filter.getElements() / groups_;

93
    for (size_t i = 0; i < batchSize; i++) {
94
      for (size_t g = 0; g < groups_; g++) {
95
        if (needIm2col) {
96 97 98 99 100 101 102
          im2col(inputData + g * inputOffset,
                 imShape,
                 colData,
                 colShape,
                 strideH(),
                 strideW(),
                 paddingH(),
X
xzl 已提交
103 104 105
                 paddingW(),
                 dilationH(),
                 dilationW());
106 107
        } else {
          colData = inputData + g * inputOffset;
108
        }
H
Bug fix  
hedaoyuan 已提交
109
        int M = outputChannels / groups_;
110
        int N = outputHeight * outputWidth;
H
Bug fix  
hedaoyuan 已提交
111
        int K = inputChannels / groups_ * filterHeight * filterWidth;
H
hedaoyuan 已提交
112 113 114 115 116 117 118 119 120 121 122 123 124
        BlasGemm<Device, real>::compute(false,
                                        false,
                                        M,
                                        N,
                                        K,
                                        1.0f,
                                        filterData + g * filterOffset,
                                        K,
                                        colData,
                                        N,
                                        beta,
                                        outputData + g * outputOffset,
                                        N);
125
      }
H
hedaoyuan 已提交
126 127
      inputData += inputChannels * inputHeight * inputWidth;
      outputData += outputChannels * outputHeight * outputWidth;
128 129 130 131
    }
  }
};

H
hedaoyuan 已提交
132 133
#ifdef PADDLE_MOBILE_INFERENCE

H
hedaoyuan 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205
/*
 * \brief Forward calculation of convolution, optimized for mobile.
 */
template <DeviceType Device>
class GemmConvMobileFunction : public ConvFunctionBase {
public:
  void init(const FuncConfig& config) override {
    ConvFunctionBase::init(config);
  }

  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();
    checkShape(input, filter, output);
  }

  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
    CHECK_EQ(numInputs_, inputs.size());
    CHECK_EQ(numOutputs_, outputs.size());
    check(inputs, outputs);
    // TODO(hedaoyuan): Need to define some index macros,
    // to avoid useing 0 and 1.
    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();

    real beta;
    if (outputs[0].getArgType() == ADD_TO) {
      beta = 1.0;
    } else {
      beta = 0.0;
    }

    size_t batchSize = input[0];
    size_t inputChannels = input[1];
    size_t inputHeight = input[2];
    size_t inputWidth = input[3];
    size_t filterHeight = getFilterHeight(filter);
    size_t filterWidth = getFilterWidth(filter);
    size_t outputChannels = output[1];
    size_t outputHeight = output[2];
    size_t outputWidth = output[3];

    real* inputData = inputs[0].data<real>();
    real* filterData = inputs[1].data<real>();
    real* outputData = outputs[0].data<real>();
    bool needIm2col = isNeedIm2col(filter);

    TensorShape imShape =
        TensorShape({inputChannels / groups_, inputHeight, inputWidth});

    TensorShape colShape;
    real* colData = NULL;

    size_t colHeight = inputChannels / groups_ * filterHeight * filterWidth;
    size_t colWidth = outputHeight * outputWidth;
    // Max col matrix height 256, Max col matrix width 1024
    size_t stepColHeight = std::min(colHeight, (size_t)256);
    size_t stepColWidth = std::min(colWidth, (size_t)2048);

    if (needIm2col) {
      colShape = TensorShape({inputChannels / groups_,
                              filterHeight,
                              filterWidth,
                              outputHeight,
                              outputWidth});

      resizeBuffer<Device>(stepColHeight * stepColWidth * sizeof(real));
      colData = reinterpret_cast<real*>(memory_->getBuf());
    }

H
hedaoyuan 已提交
206
    Im2ColMobileFunctor<real> im2col;
H
hedaoyuan 已提交
207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
    size_t inputOffset = imShape.getElements();
    size_t outputOffset =
        (outputChannels / groups_) * outputHeight * outputWidth;
    size_t filterOffset = filter.getElements() / groups_;

    int nStride = colWidth;
    int kStride = colHeight;
    for (size_t i = 0; i < batchSize; i++) {
      for (size_t g = 0; g < groups_; g++) {
        if (needIm2col) {
          real beta_ = beta;
          for (size_t colHeightStart = 0; colHeightStart < colHeight;
               colHeightStart += stepColHeight) {
            for (size_t colWidthStart = 0; colWidthStart < colWidth;
                 colWidthStart += stepColWidth) {
              int N = std::min(colWidth - colWidthStart, stepColWidth);
              int K = std::min(colHeight - colHeightStart, stepColHeight);
              // im2col
              im2col(inputData + g * inputOffset,
                     imShape,
                     colData,
                     colShape,
                     strideH(),
                     strideW(),
                     paddingH(),
                     paddingW(),
H
hedaoyuan 已提交
233 234
                     dilationH(),
                     dilationW(),
H
hedaoyuan 已提交
235 236 237 238 239 240 241
                     colHeightStart,
                     K,
                     colWidthStart,
                     N);

              // gemm
              int M = outputChannels / groups_;
H
hedaoyuan 已提交
242 243 244 245 246 247 248 249 250 251 252 253 254 255
              BlasGemm<Device, real>::compute(
                  false,
                  false,
                  M,
                  N,
                  K,
                  1.0f,
                  filterData + g * filterOffset + colHeightStart,
                  kStride,
                  colData,
                  N,
                  beta_,
                  outputData + g * outputOffset + colWidthStart,
                  nStride);
H
hedaoyuan 已提交
256 257 258 259 260 261 262
            }
            beta_ = 1.0;
          }
        } else {
          int M = outputChannels / groups_;
          int N = outputHeight * outputWidth;
          int K = inputChannels / groups_ * filterHeight * filterWidth;
H
hedaoyuan 已提交
263 264 265 266 267 268 269 270 271 272 273 274 275
          BlasGemm<Device, real>::compute(false,
                                          false,
                                          M,
                                          N,
                                          K,
                                          1.0f,
                                          filterData + g * filterOffset,
                                          K,
                                          inputData + g * inputOffset,
                                          N,
                                          beta,
                                          outputData + g * outputOffset,
                                          N);
H
hedaoyuan 已提交
276 277 278 279 280 281 282 283
        }
      }
      inputData += inputChannels * inputHeight * inputWidth;
      outputData += outputChannels * outputHeight * outputWidth;
    }
  }
};

H
hedaoyuan 已提交
284 285
#endif

286 287 288 289 290 291 292 293 294 295
/*
 * \brief Backward input calculation of convolution.
 */
template <DeviceType Device>
class GemmConvGradInputFunction : public ConvFunctionBase {
public:
  void init(const FuncConfig& config) override {
    ConvFunctionBase::init(config);
  }

L
liaogang 已提交
296
  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
H
hedaoyuan 已提交
297 298 299 300 301 302
    const TensorShape& output = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& input = outputs[0].shape();
    checkShape(input, filter, output);
  }

303 304 305
  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
    CHECK_EQ(numInputs_, inputs.size());
    CHECK_EQ(numOutputs_, outputs.size());
H
hedaoyuan 已提交
306
    check(inputs, outputs);
H
hedaoyuan 已提交
307 308 309
    // Since the implementation of Col2ImFunctor is ADD_TO,
    // this function only supports ADD_TO mode.
    CHECK_EQ(outputs[0].getArgType(), ADD_TO);
310
    const TensorShape& output = inputs[0].shape();
311
    const TensorShape& filter = inputs[1].shape();
312 313 314 315 316 317
    const TensorShape& input = outputs[0].shape();

    size_t batchSize = input[0];
    size_t inputChannels = input[1];
    size_t inputHeight = input[2];
    size_t inputWidth = input[3];
H
hedaoyuan 已提交
318 319
    size_t filterHeight = getFilterHeight(filter);
    size_t filterWidth = getFilterWidth(filter);
320 321 322 323 324 325 326
    size_t outputChannels = output[1];
    size_t outputHeight = output[2];
    size_t outputWidth = output[3];

    real* outputGrad = inputs[0].data<real>();
    real* filterData = inputs[1].data<real>();
    real* inputGrad = outputs[0].data<real>();
327
    bool needIm2col = isNeedIm2col(filter);
328

329 330 331
    TensorShape imShape =
        TensorShape({inputChannels / groups_, inputHeight, inputWidth});

332
    TensorShape colShape;
333
    real* colData = NULL;
334

335
    if (needIm2col) {
336 337 338 339 340 341 342 343
      colShape = TensorShape({inputChannels / groups_,
                              filterHeight,
                              filterWidth,
                              outputHeight,
                              outputWidth});
      resizeBuffer<Device>(colShape.getElements());
      colData = reinterpret_cast<real*>(memory_->getBuf());
    }
344

345 346
    Col2ImFunctor<kCFO, Device, real> col2im;
    size_t inputOffset = imShape.getElements();
H
format  
hedaoyuan 已提交
347
    size_t outputOffset =
348 349 350 351 352 353 354 355
        (outputChannels / groups_) * outputHeight * outputWidth;
    size_t filterOffset = filter.getElements() / groups_;

    for (size_t i = 0; i < batchSize; i++) {
      for (size_t g = 0; g < groups_; g++) {
        int K = outputChannels / groups_;
        int N = outputHeight * outputWidth;
        int M = inputChannels / groups_ * filterHeight * filterWidth;
356
        real scale = 0.0f;
357 358
        if (!needIm2col) {
          colData = inputGrad + g * inputOffset;
359 360
          scale = 1.0f;
        }
H
hedaoyuan 已提交
361 362 363 364 365 366 367 368 369 370 371 372 373
        BlasGemm<Device, real>::compute(true,
                                        false,
                                        M,
                                        N,
                                        K,
                                        1.0f,
                                        filterData + g * filterOffset,
                                        M,
                                        outputGrad + g * outputOffset,
                                        N,
                                        scale,
                                        colData,
                                        N);
374
        if (needIm2col) {
375 376
          col2im(inputGrad + g * inputOffset,
                 imShape,
377
                 colData,
378 379 380 381
                 colShape,
                 strideH(),
                 strideW(),
                 paddingH(),
X
xzl 已提交
382 383 384
                 paddingW(),
                 dilationH(),
                 dilationW());
385
        }
386 387 388 389
      }
      inputGrad += inputChannels * inputHeight * inputWidth;
      outputGrad += outputChannels * outputHeight * outputWidth;
    }
390 391 392 393 394 395 396 397 398 399 400 401 402
  }
};

/*
 * \brief Backward filter calculation of convolution.
 */
template <DeviceType Device>
class GemmConvGradFilterFunction : public ConvFunctionBase {
public:
  void init(const FuncConfig& config) override {
    ConvFunctionBase::init(config);
  }

L
liaogang 已提交
403
  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
H
hedaoyuan 已提交
404 405 406 407 408 409
    const TensorShape& output = inputs[0].shape();
    const TensorShape& input = inputs[1].shape();
    const TensorShape& filter = outputs[0].shape();
    checkShape(input, filter, output);
  }

410 411 412
  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
    CHECK_EQ(numInputs_, inputs.size());
    CHECK_EQ(numOutputs_, outputs.size());
H
hedaoyuan 已提交
413
    check(inputs, outputs);
414
    const TensorShape& output = inputs[0].shape();
415
    const TensorShape& input = inputs[1].shape();
416 417
    const TensorShape& filter = outputs[0].shape();

418 419 420 421 422 423 424
    real beta;
    if (outputs[0].getArgType() == ADD_TO) {
      beta = 1.0;
    } else {
      beta = 0.0;
    }

425 426 427 428
    size_t batchSize = input[0];
    size_t inputChannels = input[1];
    size_t inputHeight = input[2];
    size_t inputWidth = input[3];
H
hedaoyuan 已提交
429 430
    size_t filterHeight = getFilterHeight(filter);
    size_t filterWidth = getFilterWidth(filter);
431 432 433 434 435 436 437
    size_t outputChannels = output[1];
    size_t outputHeight = output[2];
    size_t outputWidth = output[3];

    real* outputGrad = inputs[0].data<real>();
    real* inputData = inputs[1].data<real>();
    real* filterGrad = outputs[0].data<real>();
438
    bool needIm2col = isNeedIm2col(filter);
439

440 441 442
    TensorShape imShape =
        TensorShape({inputChannels / groups_, inputHeight, inputWidth});

443
    TensorShape colShape;
444
    real* colData = NULL;
445

446
    if (needIm2col) {
447 448 449 450 451 452 453 454
      colShape = TensorShape({inputChannels / groups_,
                              filterHeight,
                              filterWidth,
                              outputHeight,
                              outputWidth});
      resizeBuffer<Device>(colShape.getElements());
      colData = reinterpret_cast<real*>(memory_->getBuf());
    }
455

456 457
    Im2ColFunctor<kCFO, Device, real> im2col;
    size_t inputOffset = imShape.getElements();
458 459 460 461 462
    size_t outputOffset =
        (outputChannels / groups_) * outputHeight * outputWidth;
    size_t filterOffset = filter.getElements() / groups_;
    for (size_t i = 0; i < batchSize; i++) {
      for (size_t g = 0; g < groups_; g++) {
463
        if (needIm2col) {
464 465 466 467 468 469 470
          im2col(inputData + g * inputOffset,
                 imShape,
                 colData,
                 colShape,
                 strideH(),
                 strideW(),
                 paddingH(),
X
xzl 已提交
471 472 473
                 paddingW(),
                 dilationH(),
                 dilationW());
474 475
        } else {
          colData = inputData + g * inputOffset;
476
        }
477 478 479
        int M = outputChannels / groups_;
        int K = outputHeight * outputWidth;
        int N = inputChannels / groups_ * filterHeight * filterWidth;
H
hedaoyuan 已提交
480 481 482 483 484 485 486 487 488 489 490 491 492
        BlasGemm<Device, real>::compute(false,
                                        true,
                                        M,
                                        N,
                                        K,
                                        1.0f,
                                        outputGrad + g * outputOffset,
                                        K,
                                        colData,
                                        K,
                                        i == 0 ? beta : 1.0f,
                                        filterGrad + g * filterOffset,
                                        N);
493
      }
494 495
      inputData += inputChannels * inputHeight * inputWidth;
      outputGrad += outputChannels * outputHeight * outputWidth;
496
    }
497 498 499
  }
};

H
hedaoyuan 已提交
500 501 502
#ifdef PADDLE_MOBILE_INFERENCE
REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvMobileFunction);
#else
503
REGISTER_TYPED_FUNC(GemmConv, CPU, GemmConvFunction);
H
hedaoyuan 已提交
504
#endif
505 506
REGISTER_TYPED_FUNC(GemmConvGradInput, CPU, GemmConvGradInputFunction);
REGISTER_TYPED_FUNC(GemmConvGradFilter, CPU, GemmConvGradFilterFunction);
507
#ifdef PADDLE_WITH_CUDA
508
REGISTER_TYPED_FUNC(GemmConv, GPU, GemmConvFunction);
509 510
REGISTER_TYPED_FUNC(GemmConvGradInput, GPU, GemmConvGradInputFunction);
REGISTER_TYPED_FUNC(GemmConvGradFilter, GPU, GemmConvGradFilterFunction);
H
hedaoyuan 已提交
511
#endif
512 513

}  // namespace paddle