Merge pull request BVLC#122 from thatguymike/v5support

Add basic v5 support
thatguymike · Mar 12, 2016 · 192d349 · 192d349
2 parents c0c2109 + 2a4fb59
commit 192d349
Show file tree

Hide file tree

Showing 9 changed files with 105 additions and 18 deletions.
diff --git a/include/caffe/neuron_layers.hpp b/include/caffe/neuron_layers.hpp
@@ -499,6 +499,7 @@ class CuDNNReLULayer : public ReLULayer<Dtype> {
   bool handles_setup_;
   cudnnTensorDescriptor_t bottom_desc_;
   cudnnTensorDescriptor_t top_desc_;
+  cudnnActivationDescriptor_t activ_desc_;
 };
 #endif
 
@@ -581,6 +582,7 @@ class CuDNNSigmoidLayer : public SigmoidLayer<Dtype> {
   bool handles_setup_;
   cudnnTensorDescriptor_t bottom_desc_;
   cudnnTensorDescriptor_t top_desc_;
+  cudnnActivationDescriptor_t activ_desc_;
 };
 #endif
 
@@ -665,6 +667,7 @@ class CuDNNTanHLayer : public TanHLayer<Dtype> {
   bool handles_setup_;
   cudnnTensorDescriptor_t bottom_desc_;
   cudnnTensorDescriptor_t top_desc_;
+  cudnnActivationDescriptor_t activ_desc_;
 };
 #endif
 

diff --git a/include/caffe/util/cudnn.hpp b/include/caffe/util/cudnn.hpp
@@ -7,6 +7,9 @@
 #include "caffe/common.hpp"
 #include "caffe/proto/caffe.pb.h"
 
+#define CUDNN_VERSION_MIN(major, minor, patch) \
+     (CUDNN_VERSION >= (major * 1000 + minor * 100 + patch))
+
 #define CUDNN_CHECK(condition) \
   do { \
     cudnnStatus_t status = condition; \
@@ -88,8 +91,13 @@ template <typename Dtype>
 inline void createFilterDesc(cudnnFilterDescriptor_t* desc,
     int n, int c, int h, int w) {
   CUDNN_CHECK(cudnnCreateFilterDescriptor(desc));
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnSetFilter4dDescriptor(*desc, dataType<Dtype>::type,
-      n, c, h, w));
+                                         CUDNN_TENSOR_NCHW, n, c, h, w));
+#else
+  CUDNN_CHECK(cudnnSetFilter4dDescriptor_v4(*desc, dataType<Dtype>::type,
+                                            CUDNN_TENSOR_NCHW, n, c, h, w));
+#endif
 }
 
 template <typename Dtype>
@@ -120,8 +128,15 @@ inline void createPoolingDesc(cudnnPoolingDescriptor_t* pool_desc,
     LOG(FATAL) << "Unknown pooling method.";
   }
   CUDNN_CHECK(cudnnCreatePoolingDescriptor(pool_desc));
-  CUDNN_CHECK(cudnnSetPooling2dDescriptor(*pool_desc, *mode, h, w,
-        pad_h, pad_w, stride_h, stride_w));
+#if CUDNN_VERSION_MIN(5, 0, 0)
+  CUDNN_CHECK(cudnnSetPooling2dDescriptor(*pool_desc, *mode,
+                                          CUDNN_PROPAGATE_NAN, h, w,
+                                          pad_h, pad_w, stride_h, stride_w));
+#else
+  CUDNN_CHECK(cudnnSetPooling2dDescriptor_v4(*pool_desc, *mode,
+                                          CUDNN_PROPAGATE_NAN, h, w,
+                                          pad_h, pad_w, stride_h, stride_w));
+#endif
 }
 
 }  // namespace cudnn

diff --git a/src/caffe/layers/cudnn_conv_layer.cu b/src/caffe/layers/cudnn_conv_layer.cu
@@ -12,8 +12,8 @@
 // in presence of <80 characters rule
 #define cudnnConvFwd                       cudnnConvolutionForward
 #define cudnnConvBwdBias                   cudnnConvolutionBackwardBias
-#define cudnnConvBwdFilter                 cudnnConvolutionBackwardFilter_v3
-#define cudnnConvBwdData                   cudnnConvolutionBackwardData_v3
+#define cudnnConvBwdFilter                 cudnnConvolutionBackwardFilter
+#define cudnnConvBwdData                   cudnnConvolutionBackwardData
 
 namespace caffe {
 
@@ -58,13 +58,13 @@ namespace caffe {
         // Bias.
         if (this->bias_term_) {
           const Dtype* bias_data = this->blobs_[1]->gpu_data();
-          CUDNN_CHECK(cudnnAddTensor_v3(Caffe::cudnn_handle(),
-                                        cudnn::dataType<Dtype>::one,
-                                        bias_desc_,
-                                        bias_data + bias_offset_ * g,
-                                        cudnn::dataType<Dtype>::one,
-                                        top_descs_[i],
-                                        top_data + top_offset_ * g));
+          CUDNN_CHECK(cudnnAddTensor(Caffe::cudnn_handle(),
+                                     cudnn::dataType<Dtype>::one,
+                                     bias_desc_,
+                                     bias_data + bias_offset_ * g,
+                                     cudnn::dataType<Dtype>::one,
+                                     top_descs_[i],
+                                     top_data + top_offset_ * g));
         }
       }
 

diff --git a/src/caffe/layers/cudnn_relu_layer.cpp b/src/caffe/layers/cudnn_relu_layer.cpp
@@ -15,6 +15,9 @@ void CuDNNReLULayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
   cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
   cudnn::createTensor4dDesc<Dtype>(&top_desc_);
   handles_setup_ = true;
+  cudnnCreateActivationDescriptor(&activ_desc_);
+  cudnnSetActivationDescriptor(activ_desc_, CUDNN_ACTIVATION_RELU,
+                               CUDNN_PROPAGATE_NAN, 0.0);
 }
 
 template <typename Dtype>
@@ -34,6 +37,7 @@ CuDNNReLULayer<Dtype>::~CuDNNReLULayer() {
   // Check that handles have been setup before destroying.
   if (!handles_setup_) { return; }
 
+  cudnnDestroyActivationDescriptor(this->activ_desc_);
   cudnnDestroyTensorDescriptor(this->bottom_desc_);
   cudnnDestroyTensorDescriptor(this->top_desc_);
 }

diff --git a/src/caffe/layers/cudnn_relu_layer.cu b/src/caffe/layers/cudnn_relu_layer.cu
@@ -17,12 +17,21 @@ void CuDNNReLULayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
 
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* top_data = top[0]->mutable_gpu_data();
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationForward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_RELU,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->top_desc_, top_data));
+#else
+  CUDNN_CHECK(cudnnActivationForward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->top_desc_, top_data));
+#endif
 }
 
 template <typename Dtype>
@@ -42,13 +51,23 @@ void CuDNNReLULayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
   const Dtype* top_diff = top[0]->gpu_diff();
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationBackward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_RELU,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->top_desc_, top_data, this->top_desc_, top_diff,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->bottom_desc_, bottom_diff));
+#else
+  CUDNN_CHECK(cudnnActivationBackward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->top_desc_, top_data, this->top_desc_, top_diff,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->bottom_desc_, bottom_diff));
+#endif
 }
 
 INSTANTIATE_LAYER_GPU_FUNCS(CuDNNReLULayer);

diff --git a/src/caffe/layers/cudnn_sigmoid_layer.cpp b/src/caffe/layers/cudnn_sigmoid_layer.cpp
@@ -14,6 +14,9 @@ void CuDNNSigmoidLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
   // initialize cuDNN
   cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
   cudnn::createTensor4dDesc<Dtype>(&top_desc_);
+  cudnnCreateActivationDescriptor(&activ_desc_);
+  cudnnSetActivationDescriptor(activ_desc_, CUDNN_ACTIVATION_SIGMOID,
+                               CUDNN_PROPAGATE_NAN, 0.0);
   handles_setup_ = true;
 }
 
@@ -34,6 +37,7 @@ CuDNNSigmoidLayer<Dtype>::~CuDNNSigmoidLayer() {
   // Check that handles have been setup before destroying.
   if (!handles_setup_) { return; }
 
+  cudnnDestroyActivationDescriptor(this->activ_desc_);
   cudnnDestroyTensorDescriptor(this->bottom_desc_);
   cudnnDestroyTensorDescriptor(this->top_desc_);
 }

diff --git a/src/caffe/layers/cudnn_sigmoid_layer.cu b/src/caffe/layers/cudnn_sigmoid_layer.cu
@@ -12,12 +12,21 @@ void CuDNNSigmoidLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* top_data = top[0]->mutable_gpu_data();
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationForward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_SIGMOID,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->top_desc_, top_data));
+#else
+  CUDNN_CHECK(cudnnActivationForward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->top_desc_, top_data));
+#endif
 }
 
 template <typename Dtype>
@@ -32,13 +41,23 @@ void CuDNNSigmoidLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
   const Dtype* top_diff = top[0]->gpu_diff();
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationBackward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_SIGMOID,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->top_desc_, top_data, this->top_desc_, top_diff,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->bottom_desc_, bottom_diff));
+#else
+  CUDNN_CHECK(cudnnActivationBackward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->top_desc_, top_data, this->top_desc_, top_diff,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->bottom_desc_, bottom_diff));
+#endif
 }
 
 INSTANTIATE_LAYER_GPU_FUNCS(CuDNNSigmoidLayer);

diff --git a/src/caffe/layers/cudnn_tanh_layer.cpp b/src/caffe/layers/cudnn_tanh_layer.cpp
@@ -14,6 +14,9 @@ void CuDNNTanHLayer<Dtype>::LayerSetUp(const vector<Blob<Dtype>*>& bottom,
   // initialize cuDNN
   cudnn::createTensor4dDesc<Dtype>(&bottom_desc_);
   cudnn::createTensor4dDesc<Dtype>(&top_desc_);
+  cudnnCreateActivationDescriptor(&activ_desc_);
+  cudnnSetActivationDescriptor(activ_desc_, CUDNN_ACTIVATION_TANH,
+                               CUDNN_PROPAGATE_NAN, 0.0);
   handles_setup_ = true;
 }
 
@@ -34,6 +37,7 @@ CuDNNTanHLayer<Dtype>::~CuDNNTanHLayer() {
   // Check that handles have been setup before destroying.
   if (!handles_setup_) { return; }
 
+  cudnnDestroyActivationDescriptor(this->activ_desc_);
   cudnnDestroyTensorDescriptor(this->bottom_desc_);
   cudnnDestroyTensorDescriptor(this->top_desc_);
 }

diff --git a/src/caffe/layers/cudnn_tanh_layer.cu b/src/caffe/layers/cudnn_tanh_layer.cu
@@ -12,12 +12,21 @@ void CuDNNTanHLayer<Dtype>::Forward_gpu(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* top_data = top[0]->mutable_gpu_data();
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationForward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_TANH,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->top_desc_, top_data));
+#else
+  CUDNN_CHECK(cudnnActivationForward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->top_desc_, top_data));
+#endif
 }
 
 template <typename Dtype>
@@ -33,13 +42,23 @@ void CuDNNTanHLayer<Dtype>::Backward_gpu(const vector<Blob<Dtype>*>& top,
   const Dtype* bottom_data = bottom[0]->gpu_data();
   Dtype* bottom_diff = bottom[0]->mutable_gpu_diff();
 
+#if CUDNN_VERSION_MIN(5, 0, 0)
   CUDNN_CHECK(cudnnActivationBackward(Caffe::cudnn_handle(),
-        CUDNN_ACTIVATION_TANH,
+        activ_desc_,
         cudnn::dataType<Dtype>::one,
         this->top_desc_, top_data, this->top_desc_, top_diff,
         this->bottom_desc_, bottom_data,
         cudnn::dataType<Dtype>::zero,
         this->bottom_desc_, bottom_diff));
+#else
+  CUDNN_CHECK(cudnnActivationBackward_v4(Caffe::cudnn_handle(),
+        activ_desc_,
+        cudnn::dataType<Dtype>::one,
+        this->top_desc_, top_data, this->top_desc_, top_diff,
+        this->bottom_desc_, bottom_data,
+        cudnn::dataType<Dtype>::zero,
+        this->bottom_desc_, bottom_diff));
+#endif
 }
 
 INSTANTIATE_LAYER_GPU_FUNCS(CuDNNTanHLayer);