Add cuDNN 9.0 (#62498)

* fix cuDNN 9 problem * remove glog
PaddlePaddle · Mar 11, 2024 · c8e8be2 · c8e8be2
1 parent f512028
commit c8e8be2
Show file tree

Hide file tree

Showing 11 changed files with 492 additions and 41 deletions.
diff --git a/paddle/fluid/operators/cudnn_rnn_cache.h b/paddle/fluid/operators/cudnn_rnn_cache.h
@@ -30,8 +30,13 @@ struct CudnnRNNCache {
   ~CudnnRNNCache() { release(); }
 
   cudnnRNNDescriptor_t rnn_desc_;
+#if CUDNN_VERSION >= 90000
+  cudnnRNNDataDescriptor_t x_desc_;
+  cudnnRNNDataDescriptor_t y_desc_;
+#else
   cudnnTensorDescriptor_t *x_desc_;
   cudnnTensorDescriptor_t *y_desc_;
+#endif
 
   cudnnTensorDescriptor_t hx_desc_;
   cudnnTensorDescriptor_t cx_desc_;
@@ -93,7 +98,37 @@ struct CudnnRNNCache {
     const auto numDirections = is_bidirec_ ? 2 : 1;
     auto cudnn_size =
         cudnn_type == CUDNN_DATA_FLOAT ? sizeof(float) : sizeof(double);
+#if CUDNN_VERSION >= 90000
+    PADDLE_ENFORCE_GPU_SUCCESS(
+        platform::dynload::cudnnCreateRNNDataDescriptor(&x_desc_));
+    PADDLE_ENFORCE_GPU_SUCCESS(
+        platform::dynload::cudnnCreateRNNDataDescriptor(&y_desc_));
+
+    std::vector<int> seq_length_array(batch_size_);
+    for (int i = 0; i < batch_size_; ++i) {
+      seq_length_array[i] = seq_length_;
+    }
 
+    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetRNNDataDescriptor(
+        x_desc_,
+        cudnn_type,
+        CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED,
+        seq_length_,
+        batch_size_,
+        input_size_,
+        reinterpret_cast<const int *>(seq_length_array.data()),
+        nullptr));
+
+    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetRNNDataDescriptor(
+        y_desc_,
+        cudnn_type,
+        CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED,
+        seq_length_,
+        batch_size_,
+        hidden_size_ * numDirections,
+        reinterpret_cast<const int *>(seq_length_array.data()),
+        nullptr));
+#else
     x_desc_ = new cudnnTensorDescriptor_t[seq_length_];
     y_desc_ = new cudnnTensorDescriptor_t[seq_length_];
     std::vector<int> dims = {batch_size_, input_size_, 1};
@@ -114,6 +149,7 @@ struct CudnnRNNCache {
       PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetTensorNdDescriptor(
           y_desc_[i], cudnn_type, 3, dims_y.data(), strides_y.data()));
     }
+#endif
 
     std::vector<int> dims_hx = {
         num_layers_ * numDirections, batch_size_, hidden_size_};
@@ -185,7 +221,24 @@ struct CudnnRNNCache {
 
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnCreateRNNDescriptor(&rnn_desc_));
-
+#if CUDNN_VERSION >= 90000
+    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetRNNDescriptor_v8(
+        rnn_desc_,
+        CUDNN_RNN_ALGO_STANDARD,
+        CUDNN_LSTM,
+        CUDNN_RNN_DOUBLE_BIAS,
+        is_bidirec_ ? CUDNN_BIDIRECTIONAL : CUDNN_UNIDIRECTIONAL,
+        CUDNN_LINEAR_INPUT,
+        cudnn_type,
+        cudnn_type,
+        CUDNN_DEFAULT_MATH,
+        input_size_,
+        hidden_size_,
+        hidden_size_,
+        num_layers_,
+        dropout_desc_,
+        CUDNN_RNN_PADDED_IO_ENABLED));
+#else
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetRNNDescriptor_v6(
         handle,
         rnn_desc_,
@@ -197,15 +250,19 @@ struct CudnnRNNCache {
         CUDNN_LSTM,
         CUDNN_RNN_ALGO_STANDARD,
         cudnn_type));
-
+#endif
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnCreateFilterDescriptor(&w_desc_));
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnCreateFilterDescriptor(&dw_desc_));
 
+#if CUDNN_VERSION >= 90000
+    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnGetRNNWeightSpaceSize(
+        handle, rnn_desc_, &weights_size_));
+#else
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnGetRNNParamsSize(
         handle, rnn_desc_, x_desc_[0], &weights_size_, cudnn_type));
-
+#endif
     PADDLE_ENFORCE_EQ(
         weights_size_,
         cudnn_size * weight_numel,
@@ -220,18 +277,32 @@ struct CudnnRNNCache {
         w_desc_, cudnn_type, CUDNN_TENSOR_NCHW, 3, dim_w));
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetFilterNdDescriptor(
         dw_desc_, cudnn_type, CUDNN_TENSOR_NCHW, 3, dim_w));
-
+#if CUDNN_VERSION >= 90000
+    PADDLE_ENFORCE_GPU_SUCCESS(
+        platform::dynload::cudnnGetRNNTempSpaceSizes(handle,
+                                                     rnn_desc_,
+                                                     CUDNN_FWD_MODE_TRAINING,
+                                                     x_desc_,
+                                                     &workspace_size_,
+                                                     reserve_size_));
+#else
     PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnGetRNNWorkspaceSize(
         handle, rnn_desc_, seq_length_, x_desc_, &workspace_size_));
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnGetRNNTrainingReserveSize(
             handle, rnn_desc_, seq_length_, x_desc_, reserve_size_));
-
+#endif
     workspace_data_.Resize({static_cast<int64_t>(workspace_size_)});
     workspace_data_.mutable_data<uint8_t>(place);
   }
 
   void release() {
+#if CUDNN_VERSION >= 90000
+    PADDLE_ENFORCE_GPU_SUCCESS(
+        platform::dynload::cudnnDestroyRNNDataDescriptor(x_desc_));
+    PADDLE_ENFORCE_GPU_SUCCESS(
+        platform::dynload::cudnnDestroyRNNDataDescriptor(y_desc_));
+#else
     for (size_t i = 0; i < seq_length_; ++i) {
       PADDLE_ENFORCE_GPU_SUCCESS(
           platform::dynload::cudnnDestroyTensorDescriptor(x_desc_[i]));
@@ -241,6 +312,7 @@ struct CudnnRNNCache {
 
     delete[] x_desc_;
     delete[] y_desc_;
+#endif
 
     PADDLE_ENFORCE_GPU_SUCCESS(
         platform::dynload::cudnnDestroyTensorDescriptor(hx_desc_));

diff --git a/paddle/fluid/platform/dynload/cudnn.cc b/paddle/fluid/platform/dynload/cudnn.cc
@@ -44,6 +44,18 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R7(DEFINE_WRAP);
 CUDNN_DNN_ROUTINE_EACH_R8(DEFINE_WRAP);
 #endif
 
+#ifdef CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9
+CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(DEFINE_WRAP);
+#endif
+
+#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9
+CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(DEFINE_WRAP);
+#endif
+
+#ifdef CUDNN_DNN_ROUTINE_EACH_R9
+CUDNN_DNN_ROUTINE_EACH_R9(DEFINE_WRAP);
+#endif
+
 bool HasCUDNN() { return phi::dynload::HasCUDNN(); }
 
 }  // namespace dynload

diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h
@@ -90,13 +90,6 @@ extern bool HasCUDNN();
   __macro(cudnnSetDropoutDescriptor);                      \
   __macro(cudnnRestoreDropoutDescriptor);                  \
   __macro(cudnnCreateRNNDescriptor);                       \
-  __macro(cudnnGetRNNParamsSize);                          \
-  __macro(cudnnGetRNNWorkspaceSize);                       \
-  __macro(cudnnGetRNNTrainingReserveSize);                 \
-  __macro(cudnnRNNForwardTraining);                        \
-  __macro(cudnnRNNBackwardData);                           \
-  __macro(cudnnRNNBackwardWeights);                        \
-  __macro(cudnnRNNForwardInference);                       \
   __macro(cudnnDestroyDropoutDescriptor);                  \
   __macro(cudnnDestroyRNNDescriptor);                      \
   __macro(cudnnSetTensorNdDescriptorEx);                   \
@@ -111,8 +104,7 @@ extern bool HasCUDNN();
   __macro(cudnnCreateActivationDescriptor);                \
   __macro(cudnnSetActivationDescriptor);                   \
   __macro(cudnnGetActivationDescriptor);                   \
-  __macro(cudnnDestroyActivationDescriptor);               \
-  __macro(cudnnSetRNNDescriptor_v6);
+  __macro(cudnnDestroyActivationDescriptor);
 CUDNN_DNN_ROUTINE_EACH(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 
 #if CUDNN_VERSION >= 7000 && CUDNN_VERSION < 8000
@@ -147,12 +139,7 @@ CUDNN_DNN_ROUTINE_EACH_R7(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #define CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7(__macro) \
   __macro(cudnnCreateRNNDataDescriptor);             \
   __macro(cudnnDestroyRNNDataDescriptor);            \
-  __macro(cudnnSetRNNDataDescriptor);                \
-  __macro(cudnnSetRNNPaddingMode);                   \
-  __macro(cudnnRNNForwardTrainingEx);                \
-  __macro(cudnnRNNBackwardDataEx);                   \
-  __macro(cudnnRNNBackwardWeightsEx);                \
-  __macro(cudnnRNNForwardInferenceEx);
+  __macro(cudnnSetRNNDataDescriptor);
 CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #endif
 
@@ -182,6 +169,39 @@ CUDNN_DNN_ROUTINE_EACH_AFTER_R7(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 CUDNN_DNN_ROUTINE_EACH_R8(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #endif
 
+#if CUDNN_VERSION < 90000
+#define CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(__macro) \
+  __macro(cudnnGetRNNParamsSize);                     \
+  __macro(cudnnGetRNNWorkspaceSize);                  \
+  __macro(cudnnGetRNNTrainingReserveSize);            \
+  __macro(cudnnSetRNNDescriptor_v6);                  \
+  __macro(cudnnRNNForwardInference);                  \
+  __macro(cudnnRNNForwardTraining);                   \
+  __macro(cudnnRNNBackwardData);                      \
+  __macro(cudnnRNNBackwardWeights);
+CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
+
+#if CUDNN_VERSION < 90000 && CUDNN_VERSION >= 7201
+#define CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(__macro) \
+  __macro(cudnnSetRNNPaddingMode);                                 \
+  __macro(cudnnRNNForwardInferenceEx);                             \
+  __macro(cudnnRNNForwardTrainingEx);                              \
+  __macro(cudnnRNNBackwardDataEx);                                 \
+  __macro(cudnnRNNBackwardWeightsEx);
+CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(
+    PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
+
+#if CUDNN_VERSION >= 90000
+#define CUDNN_DNN_ROUTINE_EACH_R9(__macro) \
+  __macro(cudnnGetRNNWeightSpaceSize);     \
+  __macro(cudnnGetRNNTempSpaceSizes);      \
+  __macro(cudnnRNNForward);                \
+  __macro(cudnnRNNBackwardData_v8);        \
+  __macro(cudnnRNNBackwardWeights_v8);
+CUDNN_DNN_ROUTINE_EACH_R9(PLATFORM_DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
 }  // namespace dynload
 }  // namespace platform
 }  // namespace paddle

diff --git a/paddle/phi/backends/dynload/cudnn.cc b/paddle/phi/backends/dynload/cudnn.cc
@@ -50,6 +50,18 @@ CUDNN_DNN_ROUTINE_EACH_R8(DEFINE_WRAP);
 CUDNN_DNN_ROUTINE_EACH_FRONTEND(DEFINE_WRAP);
 #endif
 
+#ifdef CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9
+CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(DEFINE_WRAP);
+#endif
+
+#ifdef CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9
+CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(DEFINE_WRAP);
+#endif
+
+#ifdef CUDNN_DNN_ROUTINE_EACH_R9
+CUDNN_DNN_ROUTINE_EACH_R9(DEFINE_WRAP);
+#endif
+
 bool HasCUDNN() {
   std::call_once(cudnn_dso_flag,
                  []() { cudnn_dso_handle = GetCUDNNDsoHandle(); });

diff --git a/paddle/phi/backends/dynload/cudnn.h b/paddle/phi/backends/dynload/cudnn.h
@@ -103,13 +103,6 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
   __macro(cudnnSetDropoutDescriptor);                      \
   __macro(cudnnRestoreDropoutDescriptor);                  \
   __macro(cudnnCreateRNNDescriptor);                       \
-  __macro(cudnnGetRNNParamsSize);                          \
-  __macro(cudnnGetRNNWorkspaceSize);                       \
-  __macro(cudnnGetRNNTrainingReserveSize);                 \
-  __macro(cudnnRNNForwardTraining);                        \
-  __macro(cudnnRNNBackwardData);                           \
-  __macro(cudnnRNNBackwardWeights);                        \
-  __macro(cudnnRNNForwardInference);                       \
   __macro(cudnnDestroyDropoutDescriptor);                  \
   __macro(cudnnDestroyRNNDescriptor);                      \
   __macro(cudnnSetTensorNdDescriptorEx);                   \
@@ -124,8 +117,7 @@ extern void EnforceCUDNNLoaded(const char* fn_name);
   __macro(cudnnCreateActivationDescriptor);                \
   __macro(cudnnSetActivationDescriptor);                   \
   __macro(cudnnGetActivationDescriptor);                   \
-  __macro(cudnnDestroyActivationDescriptor);               \
-  __macro(cudnnSetRNNDescriptor_v6);
+  __macro(cudnnDestroyActivationDescriptor);
 CUDNN_DNN_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 
 #if CUDNN_VERSION >= 7000 && CUDNN_VERSION < 8000
@@ -159,12 +151,7 @@ CUDNN_DNN_ROUTINE_EACH_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #define CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7(__macro) \
   __macro(cudnnCreateRNNDataDescriptor);             \
   __macro(cudnnDestroyRNNDataDescriptor);            \
-  __macro(cudnnSetRNNDataDescriptor);                \
-  __macro(cudnnSetRNNPaddingMode);                   \
-  __macro(cudnnRNNForwardTrainingEx);                \
-  __macro(cudnnRNNBackwardDataEx);                   \
-  __macro(cudnnRNNBackwardWeightsEx);                \
-  __macro(cudnnRNNForwardInferenceEx);
+  __macro(cudnnSetRNNDataDescriptor);
 CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #endif
 
@@ -207,6 +194,39 @@ CUDNN_DNN_ROUTINE_EACH_R8(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 CUDNN_DNN_ROUTINE_EACH_FRONTEND(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
 #endif
 
+#if CUDNN_VERSION < 90000
+#define CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(__macro) \
+  __macro(cudnnGetRNNParamsSize);                     \
+  __macro(cudnnGetRNNWorkspaceSize);                  \
+  __macro(cudnnGetRNNTrainingReserveSize);            \
+  __macro(cudnnSetRNNDescriptor_v6);                  \
+  __macro(cudnnRNNForwardInference);                  \
+  __macro(cudnnRNNForwardTraining);                   \
+  __macro(cudnnRNNBackwardData);                      \
+  __macro(cudnnRNNBackwardWeights);
+CUDNN_DNN_ROUTINE_EACH_REMOVED_IN_E9(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
+
+#if CUDNN_VERSION < 90000 && CUDNN_VERSION >= 7201
+#define CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(__macro) \
+  __macro(cudnnSetRNNPaddingMode);                                 \
+  __macro(cudnnRNNForwardInferenceEx);                             \
+  __macro(cudnnRNNForwardTrainingEx);                              \
+  __macro(cudnnRNNBackwardDataEx);                                 \
+  __macro(cudnnRNNBackwardWeightsEx);
+CUDNN_DNN_ROUTINE_EACH_AFTER_TWO_R7_REMOVED_IN_E9(
+    DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
+
+#if CUDNN_VERSION >= 90000
+#define CUDNN_DNN_ROUTINE_EACH_R9(__macro) \
+  __macro(cudnnGetRNNWeightSpaceSize);     \
+  __macro(cudnnGetRNNTempSpaceSizes);      \
+  __macro(cudnnRNNForward);                \
+  __macro(cudnnRNNBackwardData_v8);        \
+  __macro(cudnnRNNBackwardWeights_v8);
+CUDNN_DNN_ROUTINE_EACH_R9(DECLARE_DYNAMIC_LOAD_CUDNN_WRAP)
+#endif
 }  // namespace dynload
 }  // namespace phi