cudnn deconv implememtation

paddle/gserver/CMakeLists.txt

@@ -25,12 +25,16 @@ filter_test(GSERVER_HEADER)
 filter_test(GSERVER_SOURCES)
 if(NOT WITH_GPU)
     list(REMOVE_ITEM GSERVER_HEADER
+        layers/CudnnConvBaseLayer.h
         layers/CudnnConvLayer.h
+        layers/CudnnConvTransLayer.h
         layers/CudnnPoolLayer.h
         layers/CudnnBatchNormLayer.h)
 
     list(REMOVE_ITEM GSERVER_SOURCES
+        layers/CudnnConvBaseLayer.cpp
         layers/CudnnConvLayer.cpp
+        layers/CudnnConvTransLayer.cpp
         layers/CudnnPoolLayer.cpp
         layers/CudnnBatchNormLayer.cpp)
     compile_cu_as_cpp(layers/LstmCompute.cu)

paddle/gserver/layers/ConvBaseProjection.cpp

@@ -0,0 +1,204 @@
+/* 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. */
+
+#include "ConvBaseProjection.h"
+#include "paddle/utils/Stat.h"
+
+namespace paddle {
+
+ThreadLocalD<std::vector<MemoryHandle *>> ConvBaseProjection::convMem_;
+
+ConvBaseProjection::ConvBaseProjection(const ProjectionConfig &config,
+                                       ParameterPtr parameter,
+                                       bool useGpu)
+    : Projection(config, parameter, useGpu) {
+  CHECK(useGpu);  // only support GPU
+  getConvParams();
+  initCudnn();
+
+  size_t height = filterH_ * filterW_ * channels_ / groups_;
+  size_t width = numFilters_;
+  weight_.reset(new Weight(height, width, parameter));
+  weightOffset_ = height * width / groups_;
+}
+
+void ConvBaseProjection::getConvParams() {
+  const ConvConfig &conf = config_.conv_conf();
+  paddingH_ = conf.padding_y();
+  paddingW_ = conf.padding();
+
+  strideH_ = conf.stride_y();
+  strideW_ = conf.stride();
+
+  filterH_ = conf.filter_size_y();
+  filterW_ = conf.filter_size();
+
+  configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
+  configImgW_ = conf.img_size();
+
+  configOutH_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
+  configOutW_ = conf.output_x();
+
+  configChannels_ = conf.channels();
+  configNumFilters_ = config_.num_filters();
+
+  isDeconv_ = (config_.type() == "conv") ? false : true;
+
+  channels_ = (isDeconv_) ? configNumFilters_ : configChannels_;
+  numFilters_ = (isDeconv_) ? configChannels_ : configNumFilters_;
+
+  groups_ = conf.groups();
+  CHECK_EQ(channels_ % groups_, 0);
+  CHECK_EQ(numFilters_ % groups_, 0);
+}
+
+void ConvBaseProjection::initCudnn() {
+  hl_create_filter_descriptor(&filterDesc_,
+                              channels_ / groups_,
+                              numFilters_ / groups_,
+                              filterH_,
+                              filterW_);
+  hl_create_tensor_descriptor(&imageDesc_);
+  hl_create_tensor_descriptor(&outputDesc_);
+  hl_create_convolution_descriptor(&convDesc_,
+                                   imageDesc_,
+                                   filterDesc_,
+                                   paddingH_,
+                                   paddingW_,
+                                   strideH_,
+                                   strideW_);
+
+  // initialize all to default algorithms
+  fwdAlgo_ = 0;
+  bwdFilterAlgo_ = 0;
+  bwdDataAlgo_ = 0;
+  fwdLimitBytes_ = 0;
+  bwdDataLimitBytes_ = 0;
+  bwdFilterLimitBytes_ = 0;
+  workSpaceInBytes_ = 0;
+
+  batchNum_ = 0;
+  isSelectAlgo_ = false;
+}
+
+void ConvBaseProjection::reshapeTensorDesc(int batchSize) {
+  hl_tensor_reshape(imageDesc_,
+                    batchSize,
+                    channels_ / groups_,
+                    imageH_,
+                    imageW_,
+                    channels_ * imageH_ * imageW_,
+                    imageH_ * imageW_,
+                    imageW_,
+                    1);
+  hl_reset_convolution_descriptor(convDesc_,
+                                  imageDesc_,
+                                  filterDesc_,
+                                  paddingH_,
+                                  paddingW_,
+                                  strideH_,
+                                  strideW_);
+
+  // The stride between two consecutive images in ConvProjection may not be 1,
+  // for example, in the case of layer ConcatenateLayer2 with two
+  // ConvProjection, the stride is the output_size of layer ConcatenateLayer2.
+  // So the calculation of nStride is different from CudnnConvLayer.
+  // In fact, only "nStride = out_->value->getStride()" is ok.
+  //  size_t nStride = numFilters_ * outputH_ * outputW_;
+  //  if (out_->value->isContiguous()) {
+  //    CHECK_EQ(nStride, out_->value->getWidth());
+  //  } else {
+  //    nStride = out_->value->getStride();
+  //  }
+  size_t nStride = out_->value->getStride();
+
+  hl_tensor_reshape(outputDesc_,
+                    batchSize,
+                    numFilters_ / groups_,
+                    outputH_,
+                    outputW_,
+                    nStride,
+                    outputH_ * outputW_,
+                    outputW_,
+                    1);
+}
+
+void ConvBaseProjection::reshape(int batchSize) {
+  size_t width = calOutputSize();
+  CHECK_EQ(width, out_->value->getWidth());
+  if (isDeconv_) {
+    CHECK_EQ(static_cast<size_t>(configChannels_ * outputH_ * outputW_),
+             in_->value->getWidth())
+        << "Wrong input size for convolution transpose"
+        << " channels=" << configChannels_ << " outputH=" << outputH_
+        << " outputW=" << outputW_ << " inputSize=" << in_->value->getWidth();
+  } else {
+    CHECK_EQ(static_cast<size_t>(configChannels_ * imageH_ * imageW_),
+             in_->value->getWidth())
+        << "Wrong input size for convolution"
+        << " channels=" << configChannels_ << " imageH=" << imageH_
+        << " imageW=" << imageW_ << " inputSize=" << in_->value->getWidth();
+  }
+
+  isSelectAlgo_ = (batchSize == batchNum_);
+  batchNum_ = batchSize;
+
+  if (!isSelectAlgo_) {
+    reshapeTensorDesc(batchSize);
+    hl_conv_workspace(imageDesc_,
+                      outputDesc_,
+                      filterDesc_,
+                      convDesc_,
+                      &fwdAlgo_,
+                      &fwdLimitBytes_,
+                      &bwdDataAlgo_,
+                      &bwdDataLimitBytes_,
+                      &bwdFilterAlgo_,
+                      &bwdFilterLimitBytes_);
+
+    size_t maxWorkSpace = 0;
+    maxWorkSpace = std::max(fwdLimitBytes_, bwdDataLimitBytes_);
+    maxWorkSpace = std::max(maxWorkSpace, bwdFilterLimitBytes_);
+    workSpaceInBytes_ = maxWorkSpace;
+
+    VLOG(3) << getName() << " Fwd / BwdData / BwdFilter algo: " << fwdAlgo_
+            << " / " << bwdDataAlgo_ << " / " << bwdFilterAlgo_;
+  }
+
+  isSelectAlgo_ = true;
+}
+
+void *ConvBaseProjection::getSpaceBytes(size_t size) {
+  std::vector<MemoryHandle *> &convMem = *convMem_;
+  if (convMem.empty()) {
+    int numDevices = hl_get_device_count();
+    convMem.resize(numDevices);
+  }
+
+  int devId = hl_get_device();
+  MemoryHandle **localMem = &(convMem[devId]);
+  if (NULL == *localMem || size > (*localMem)->getAllocSize()) {
+    *localMem = new GpuMemoryHandle(size);
+  }
+  return (*localMem)->getBuf();
+}
+
+ConvBaseProjection::~ConvBaseProjection() {
+  hl_destroy_tensor_descriptor(imageDesc_);
+  hl_destroy_tensor_descriptor(outputDesc_);
+  hl_destroy_filter_descriptor(filterDesc_);
+  hl_destroy_convolution_descriptor(convDesc_);
+}
+
+}  // namespace paddle

paddle/gserver/layers/ConvBaseProjection.h

@@ -0,0 +1,162 @@
+/* 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. */
+
+#pragma once
+
+#include "Projection.h"
+#include "paddle/math/MathUtils.h"
+
+namespace paddle {
+
+/**
+ * @brief Base class for ConvProjection and ConvTransProjection.
+ */
+class ConvBaseProjection : public Projection {
+public:
+  /**
+   * Constructor.
+   */
+  ConvBaseProjection(const ProjectionConfig& config,
+                     ParameterPtr parameter,
+                     bool useGpu);
+
+  ~ConvBaseProjection();
+
+protected:
+  void getConvParams();
+  void initCudnn();
+
+  void reshapeTensorDesc(int batchSize);
+  void reshape(int batchSize);
+
+  size_t calOutputSize() {
+    if (isDeconv_) {
+      outputH_ = in_->getFrameHeight();
+      outputW_ = in_->getFrameWidth();
+      if (outputH_ == 0) outputH_ = configOutH_;
+      if (outputW_ == 0) outputW_ = configOutW_;
+      imageH_ = imageSize(outputH_,
+                          filterH_,
+                          paddingH_,
+                          strideH_,
+                          /* caffeMode */ true);
+
+      imageW_ = imageSize(outputW_,
+                          filterW_,
+                          paddingW_,
+                          strideW_,
+                          /* caffeMode */ true);
+
+      const_cast<Argument*>(out_)->setFrameHeight(imageH_);
+      const_cast<Argument*>(out_)->setFrameWidth(imageW_);
+
+      inputOffset_ = (configChannels_ / groups_) * outputH_ * outputW_;
+      outputOffset_ = (configNumFilters_ / groups_) * imageH_ * imageW_;
+      return imageH_ * imageW_ * configNumFilters_;
+    } else {
+      imageH_ = in_->getFrameHeight();
+      imageW_ = in_->getFrameWidth();
+      if (imageH_ == 0) imageH_ = configImgH_;
+      if (imageW_ == 0) imageW_ = configImgW_;
+      outputH_ = outputSize(imageH_,
+                            filterH_,
+                            paddingH_,
+                            strideH_,
+                            /* caffeMode */ true);
+      outputW_ = outputSize(imageW_,
+                            filterW_,
+                            paddingW_,
+                            strideW_,
+                            /* caffeMode */ true);
+
+      const_cast<Argument*>(out_)->setFrameHeight(outputH_);
+      const_cast<Argument*>(out_)->setFrameWidth(outputW_);
+
+      inputOffset_ = (configChannels_ / groups_) * imageH_ * imageW_;
+      outputOffset_ = (configNumFilters_ / groups_) * outputH_ * outputW_;
+      return outputH_ * outputW_ * configNumFilters_;
+    }
+  }
+
+  static void* getSpaceBytes(size_t size);
+
+  /// True if it's deconv projection layer, false if it's ConvProjection layer
+  bool isDeconv_;
+  /// imageH_ and imageW_ / outputH_ and outputW_
+  /// is calculated from the input layer.
+  int imageH_, imageW_;
+  int outputH_, outputW_;
+  /// configImgH_ and configImgW_ / configOutH_ and configOutW_
+  /// is obtained from config.
+  int configImgH_, configImgW_;
+  int configOutH_, configOutW_;
+  /// channels_ and numFilters_ are defined in terms of convolution semantics
+  int channels_, numFilters_;
+  /// configChannels and configNumFilters_ are obtained from config
+  /// For Conv they are the same as channels_ and numFilters
+  /// For ConvTrans they are opposite to channels_ and numFilters
+  int configChannels_, configNumFilters_;
+  int paddingH_, paddingW_;
+  int strideH_, strideW_;
+  int filterH_, filterW_;
+  /// One group offset of input data.
+  int inputOffset_;
+  /// One group offset of output data.
+  int outputOffset_;
+  /// One group offset of weight.
+  int weightOffset_;
+  int groups_;
+
+  /// Cudnn tensor descriptor for input.
+  hl_tensor_descriptor imageDesc_;
+  /// Cudnn tensor descriptor for output.
+  hl_tensor_descriptor outputDesc_;
+  /// Cudnn tensor descriptor for filter.
+  hl_filter_descriptor filterDesc_;
+  /// Cudnn tensor descriptor for a convolution operation.
+  hl_convolution_descriptor convDesc_;
+
+  /// Record the algorithm for forward convolution, which is obtained by cudnn
+  /// api to search the best suited algorithm.
+  int fwdAlgo_;
+  /// Record the algorithm for computing convolution gradient with respect to
+  /// filter coefficients.
+  int bwdFilterAlgo_;
+  /// Record the algorithm for computing convolution gradient with respect to
+  /// the output.
+  int bwdDataAlgo_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// forward convolution with the specified algo.
+  size_t fwdLimitBytes_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// backwardFilter with the specified algo.
+  size_t bwdDataLimitBytes_;
+  /// Amount of GPU memory needed as workspace to be able to execute a
+  /// backwardData with the specified algo.
+  size_t bwdFilterLimitBytes_;
+  /// Size of total work space.
+  size_t workSpaceInBytes_;
+
+  /// Whether to call cuDNN api to choose conv algorithm.
+  bool isSelectAlgo_;
+  /// batchNum is used to record batch size. If the batch size is changed,
+  /// the selection algorithm will be called.
+  int batchNum_;
+  bool bias_;
+
+  std::unique_ptr<Weight> weight_;
+  static ThreadLocalD<std::vector<MemoryHandle*>> convMem_;
+};
+
+}  // namespace paddle