add the mobilenet gpu acceleration, cpu is in the process

paddle/function/DepthwiseConvOp.cpp

@@ -18,11 +18,6 @@ limitations under the License. */
 
 namespace paddle {
 
-/*
- * imData = [input_channels, input_height, input_width]
- * colData = [input_channels, filter_height, filter_width,
- *            output_height, output_width]
- */
 template <class T>
 class DepthwiseConvFunctor<DEVICE_TYPE_CPU, T> {
 public:
@@ -33,6 +28,8 @@ public:
                   int outputChannels,
                   int outputHeight,
                   int outputWidth,
+                  int inputHeight,
+                  int inputWidth,
                   int filterHeight,
                   int filterWidth,
                   int strideH,
@@ -40,7 +37,7 @@ public:
                   int paddingH,
                   int paddingW,
                   T* outputData) {
-    // NO_IMPLEMENTATION
+    // TODO(zhaolong) : cpu implementation of depthwise convolution
   }
 };
 
@@ -118,8 +115,8 @@ public:
 
     size_t batchSize = input[0];
     // size_t inputChannels = input[1];
-    // size_t inputHeight = input[2];
-    // size_t inputWidth = input[3];
+    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];
@@ -139,6 +136,8 @@ public:
                   outputChannels,
                   outputHeight,
                   outputWidth,
+                  inputHeight,
+                  inputWidth,
                   filterHeight,
                   filterWidth,
                   strideH(),
@@ -233,8 +232,8 @@ public:
   }
 
   void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
-    CHECK_EQ(numInputs_, inputs.size());
-    CHECK_EQ(numOutputs_, outputs.size());
+    // CHECK_EQ(numInputs_, inputs.size());
+    // CHECK_EQ(numOutputs_, outputs.size());
     check(inputs, outputs);
     const TensorShape& output = inputs[0].shape();
     const TensorShape& input = inputs[1].shape();

paddle/function/DepthwiseConvOp.h

@@ -18,11 +18,6 @@ limitations under the License. */
 
 namespace paddle {
 
-/*
- * imData = [input_channels, input_height, input_width]
- * colData = [input_channels, filter_height, filter_width,
- *            output_height, output_width]
- */
 template <DeviceType Device, class T>
 class DepthwiseConvFunctor {
 public:
@@ -33,6 +28,8 @@ public:
                   int outputChannels,
                   int outputHeight,
                   int outputWidth,
+                  int inputHeight,
+                  int intputWidth,
                   int filterHeight,
                   int filterWidth,
                   int strideH,

paddle/gserver/layers/ConvBaseLayer.cpp

@@ -21,7 +21,8 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
                          const ParameterMap& parameterMap) {
   /* Initialize the basic parent class */
   Layer::init(layerMap, parameterMap);
-  isDeconv_ = (config_.type() == "exconv" || config_.type() == "cudnn_conv")
+  isDeconv_ = (config_.type() == "exconv" || config_.type() == "cudnn_conv" ||
+               config_.type() == "depthwise_conv")
                   ? false
                   : true;
 

paddle/gserver/layers/DepthwiseConvLayer.cpp

@@ -15,6 +15,7 @@ limitations under the License. */
 #include "DepthwiseConvLayer.h"
 #include "paddle/utils/Logging.h"
 #include "paddle/utils/Stat.h"
+#include <iostream>
 
 namespace paddle {
 
@@ -79,6 +80,7 @@ void DepthwiseConvLayer::forward(PassType passType) {
   Layer::forward(passType);
 
   size_t batchSize = inputLayers_[0]->getOutputValue()->getHeight();
+  // std::cout << "outputSize" << getOutputSize() <<std::endl;
   resetOutput(batchSize, getOutputSize());
 
   // Calculate the shape of the input, output, and filter.