nvidia-cudnn-cu12 8.9.6.50__py3-none-win_amd64.whl → 9.0.0.312__py3-none-win_amd64.whl

nvidia/cudnn/include/{cudnn_adv_infer.h → cudnn_adv.h}

@@ -47,26 +47,24 @@
  * Users Notice.
  */
 
-/*   cudnn_adv_infer : cuDNN's advanced and experimental features.
+/*   cudnn_adv : cuDNN's advanced and experimental features.
 
 */
 
-#if !defined(CUDNN_ADV_INFER_H_)
-#define CUDNN_ADV_INFER_H_
+#if !defined(CUDNN_ADV_H_)
+#define CUDNN_ADV_H_
 
-#include <cuda_runtime.h>
 #include <stdint.h>
 
 #include "cudnn_version.h"
-#include "cudnn_ops_infer.h"
+#include "cudnn_ops.h"
 
 /* These version numbers are autogenerated, do not edit manually. */
-#define CUDNN_ADV_INFER_MAJOR 8
-#define CUDNN_ADV_INFER_MINOR 9
-#define CUDNN_ADV_INFER_PATCH 6
+#define CUDNN_ADV_MAJOR 9
+#define CUDNN_ADV_MINOR 0
+#define CUDNN_ADV_PATCH 0
 
-#if (CUDNN_ADV_INFER_MAJOR != CUDNN_MAJOR) || (CUDNN_ADV_INFER_MINOR != CUDNN_MINOR) || \
-    (CUDNN_ADV_INFER_PATCH != CUDNN_PATCHLEVEL)
+#if (CUDNN_ADV_MAJOR != CUDNN_MAJOR) || (CUDNN_ADV_MINOR != CUDNN_MINOR) || (CUDNN_ADV_PATCH != CUDNN_PATCHLEVEL)
 #error Version mismatch in cuDNN ADV INFER!!!
 #endif
 
@@ -76,6 +74,14 @@ extern "C" {
 
 /* BASIC RNN API */
 
+typedef enum {
+    CUDNN_RNN_ALGO_STANDARD               = 0,
+    CUDNN_RNN_ALGO_PERSIST_STATIC         = 1,
+    CUDNN_RNN_ALGO_PERSIST_DYNAMIC        = 2,
+    CUDNN_RNN_ALGO_PERSIST_STATIC_SMALL_H = 3,
+    CUDNN_RNN_ALGO_COUNT                  = 4,
+} cudnnRNNAlgo_t;
+
 typedef enum {
     CUDNN_FWD_MODE_INFERENCE = 0,
     CUDNN_FWD_MODE_TRAINING  = 1,
@@ -116,19 +122,13 @@ typedef enum {
     CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED = 2, /* padded, outer stride from one batch to the next */
 } cudnnRNNDataLayout_t;
 
-/* Legacy type for backward compatibility */
-typedef unsigned cudnnRNNPaddingMode_t;
-
-/* For auxFlags in cudnnSetRNNDescriptor_v8() and cudnnSetRNNPaddingMode() */
+/* For auxFlags in cudnnSetRNNDescriptor_v8() */
 #define CUDNN_RNN_PADDED_IO_DISABLED 0
 #define CUDNN_RNN_PADDED_IO_ENABLED (1U << 0)
 
 struct cudnnRNNStruct;
 typedef struct cudnnRNNStruct *cudnnRNNDescriptor_t;
 
-struct cudnnPersistentRNNPlan;
-typedef struct cudnnPersistentRNNPlan *cudnnPersistentRNNPlan_t;
-
 struct cudnnRNNDataStruct;
 typedef struct cudnnRNNDataStruct *cudnnRNNDataDescriptor_t;
 
@@ -138,6 +138,13 @@ cudnnCreateRNNDescriptor(cudnnRNNDescriptor_t *rnnDesc);
 cudnnStatus_t CUDNNWINAPI
 cudnnDestroyRNNDescriptor(cudnnRNNDescriptor_t rnnDesc);
 
+/*
+ * mathPrec in cudnnSetRNNDescriptor_v8() specifies compute precision.
+ * Compute precision is further modified by mathType that sets the
+ * preferred option for using NVIDIA Tensor Cores.  dataType specify
+ * input/output data type and weight/bias type.
+ */
+
 cudnnStatus_t CUDNNWINAPI
 cudnnSetRNNDescriptor_v8(cudnnRNNDescriptor_t rnnDesc,
                          cudnnRNNAlgo_t algo,
@@ -172,49 +179,7 @@ cudnnGetRNNDescriptor_v8(cudnnRNNDescriptor_t rnnDesc,
                          cudnnDropoutDescriptor_t *dropoutDesc,
                          uint32_t *auxFlags);
 
-/*
- * mathPrec in cudnnSetRNNDescriptor_v6() specifies compute precision
- * compute precision is further modified by cudnnSetRNNMatrixMathType()
- * dataType in cudnnGetRNNParamsSize() and wDesc specify weight storage
- * dropout is between RNN layers, not between recurrent steps
- */
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNDescriptor_v6(cudnnHandle_t handle,
-                         cudnnRNNDescriptor_t rnnDesc,
-                         const int hiddenSize,
-                         const int numLayers,
-                         cudnnDropoutDescriptor_t dropoutDesc,
-                         cudnnRNNInputMode_t inputMode,
-                         cudnnDirectionMode_t direction,
-                         cudnnRNNMode_t cellMode,
-                         cudnnRNNAlgo_t algo,
-                         cudnnDataType_t mathPrec);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNDescriptor_v6(cudnnHandle_t handle,
-                         cudnnRNNDescriptor_t rnnDesc,
-                         int *hiddenSize,
-                         int *numLayers,
-                         cudnnDropoutDescriptor_t *dropoutDesc,
-                         cudnnRNNInputMode_t *inputMode,
-                         cudnnDirectionMode_t *direction,
-                         cudnnRNNMode_t *cellMode,
-                         cudnnRNNAlgo_t *algo,
-                         cudnnDataType_t *mathPrec);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNMatrixMathType(cudnnRNNDescriptor_t rnnDesc, cudnnMathType_t mType);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNMatrixMathType(cudnnRNNDescriptor_t rnnDesc, cudnnMathType_t *mType);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNBiasMode(cudnnRNNDescriptor_t rnnDesc, cudnnRNNBiasMode_t biasMode);
-
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNBiasMode(cudnnRNNDescriptor_t rnnDesc, cudnnRNNBiasMode_t *biasMode);
-
-cudnnStatus_t CUDNNWINAPI
 cudnnRNNSetClip_v8(cudnnRNNDescriptor_t rnnDesc,
                    cudnnRNNClipMode_t clipMode,
                    cudnnNanPropagation_t clipNanOpt,
@@ -222,71 +187,21 @@ cudnnRNNSetClip_v8(cudnnRNNDescriptor_t rnnDesc,
                    double rclip);
 
 cudnnStatus_t CUDNNWINAPI
+cudnnRNNSetClip_v9(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t clipMode, double lclip, double rclip);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnRNNGetClip_v8(cudnnRNNDescriptor_t rnnDesc,
                    cudnnRNNClipMode_t *clipMode,
                    cudnnNanPropagation_t *clipNanOpt,
                    double *lclip,
                    double *rclip);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnRNNSetClip(cudnnHandle_t handle,
-                cudnnRNNDescriptor_t rnnDesc,
-                cudnnRNNClipMode_t clipMode,
-                cudnnNanPropagation_t clipNanOpt,
-                double lclip,
-                double rclip);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnRNNGetClip(cudnnHandle_t handle,
-                cudnnRNNDescriptor_t rnnDesc,
-                cudnnRNNClipMode_t *clipMode,
-                cudnnNanPropagation_t *clipNanOpt,
-                double *lclip,
-                double *rclip);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNProjectionLayers(cudnnHandle_t handle,
-                            cudnnRNNDescriptor_t rnnDesc,
-                            const int recProjSize,
-                            const int outProjSize);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNProjectionLayers(cudnnHandle_t handle,
-                            const cudnnRNNDescriptor_t rnnDesc,
-                            int *recProjSize,
-                            int *outProjSize);
-
-/* Expensive. Creates the plan for the specific settings. */
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnCreatePersistentRNNPlan(cudnnRNNDescriptor_t rnnDesc,
-                             const int minibatch,
-                             const cudnnDataType_t dataType,
-                             cudnnPersistentRNNPlan_t *plan);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnDestroyPersistentRNNPlan(cudnnPersistentRNNPlan_t plan);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetPersistentRNNPlan(cudnnRNNDescriptor_t rnnDesc, cudnnPersistentRNNPlan_t plan);
+cudnnStatus_t CUDNNWINAPI
+cudnnRNNGetClip_v9(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t *clipMode, double *lclip, double *rclip);
 
 cudnnStatus_t CUDNNWINAPI
 cudnnBuildRNNDynamic(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, int miniBatch);
 
-/* dataType in weight descriptors and input descriptors is used to describe storage */
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNWorkspaceSize(cudnnHandle_t handle,
-                         const cudnnRNNDescriptor_t rnnDesc,
-                         const int seqLength,
-                         const cudnnTensorDescriptor_t *xDesc,
-                         size_t *sizeInBytes);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNTrainingReserveSize(cudnnHandle_t handle,
-                               const cudnnRNNDescriptor_t rnnDesc,
-                               const int seqLength,
-                               const cudnnTensorDescriptor_t *xDesc,
-                               size_t *sizeInBytes);
-
 cudnnStatus_t CUDNNWINAPI
 cudnnGetRNNTempSpaceSizes(cudnnHandle_t handle,
                           cudnnRNNDescriptor_t rnnDesc,
@@ -295,38 +210,9 @@ cudnnGetRNNTempSpaceSizes(cudnnHandle_t handle,
                           size_t *workSpaceSize,
                           size_t *reserveSpaceSize);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNParamsSize(cudnnHandle_t handle,
-                      const cudnnRNNDescriptor_t rnnDesc,
-                      const cudnnTensorDescriptor_t xDesc,
-                      size_t *sizeInBytes,
-                      cudnnDataType_t dataType);
-
 cudnnStatus_t CUDNNWINAPI
 cudnnGetRNNWeightSpaceSize(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, size_t *weightSpaceSize);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNLinLayerMatrixParams(cudnnHandle_t handle,
-                                const cudnnRNNDescriptor_t rnnDesc,
-                                const int pseudoLayer,
-                                const cudnnTensorDescriptor_t xDesc,
-                                const cudnnFilterDescriptor_t wDesc,
-                                const void *w,
-                                const int linLayerID,
-                                cudnnFilterDescriptor_t linLayerMatDesc,
-                                void **linLayerMat);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNLinLayerBiasParams(cudnnHandle_t handle,
-                              const cudnnRNNDescriptor_t rnnDesc,
-                              const int pseudoLayer,
-                              const cudnnTensorDescriptor_t xDesc,
-                              const cudnnFilterDescriptor_t wDesc,
-                              const void *w,
-                              const int linLayerID,
-                              cudnnFilterDescriptor_t linLayerBiasDesc,
-                              void **linLayerBias);
-
 cudnnStatus_t CUDNNWINAPI
 cudnnGetRNNWeightParams(cudnnHandle_t handle,
                         cudnnRNNDescriptor_t rnnDesc,
@@ -339,35 +225,6 @@ cudnnGetRNNWeightParams(cudnnHandle_t handle,
                         cudnnTensorDescriptor_t bDesc,
                         void **bAddr);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnRNNForwardInference(cudnnHandle_t handle,
-                         const cudnnRNNDescriptor_t rnnDesc,
-                         const int seqLength,
-                         const cudnnTensorDescriptor_t *xDesc,
-                         const void *x,
-                         const cudnnTensorDescriptor_t hxDesc,
-                         const void *hx,
-                         const cudnnTensorDescriptor_t cxDesc,
-                         const void *cx,
-                         const cudnnFilterDescriptor_t wDesc,
-                         const void *w,
-                         const cudnnTensorDescriptor_t *yDesc,
-                         void *y,
-                         const cudnnTensorDescriptor_t hyDesc,
-                         void *hy,
-                         const cudnnTensorDescriptor_t cyDesc,
-                         void *cy,
-                         void *workSpace,
-                         size_t workSpaceSizeInBytes);
-
-/* RNN EX API */
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNPaddingMode(cudnnRNNDescriptor_t rnnDesc, unsigned paddingMode);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNPaddingMode(cudnnRNNDescriptor_t rnnDesc, unsigned *paddingMode);
-
 cudnnStatus_t CUDNNWINAPI
 cudnnCreateRNNDataDescriptor(cudnnRNNDataDescriptor_t *rnnDataDesc);
 
@@ -395,34 +252,6 @@ cudnnGetRNNDataDescriptor(cudnnRNNDataDescriptor_t rnnDataDesc,
                           int seqLengthArray[],
                           void *paddingFill);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnRNNForwardInferenceEx(cudnnHandle_t handle,
-                           const cudnnRNNDescriptor_t rnnDesc,
-                           const cudnnRNNDataDescriptor_t xDesc,
-                           const void *x,
-                           const cudnnTensorDescriptor_t hxDesc,
-                           const void *hx,
-                           const cudnnTensorDescriptor_t cxDesc,
-                           const void *cx,
-                           const cudnnFilterDescriptor_t wDesc,
-                           const void *w,
-                           const cudnnRNNDataDescriptor_t yDesc,
-                           void *y,
-                           const cudnnTensorDescriptor_t hyDesc,
-                           void *hy,
-                           const cudnnTensorDescriptor_t cyDesc,
-                           void *cy,
-                           const cudnnRNNDataDescriptor_t kDesc, /* reserved, should pass NULL */
-                           const void *keys,                     /* reserved, should pass NULL */
-                           const cudnnRNNDataDescriptor_t cDesc, /* reserved, should pass NULL */
-                           void *cAttn,                          /* reserved, should pass NULL */
-                           const cudnnRNNDataDescriptor_t iDesc, /* reserved, should pass NULL */
-                           void *iAttn,                          /* reserved, should pass NULL */
-                           const cudnnRNNDataDescriptor_t qDesc, /* reserved, should pass NULL */
-                           void *queries,                        /* reserved, should pass NULL */
-                           void *workSpace,
-                           size_t workSpaceSizeInBytes);
-
 cudnnStatus_t CUDNNWINAPI
 cudnnRNNForward(cudnnHandle_t handle,
                 cudnnRNNDescriptor_t rnnDesc,
@@ -445,39 +274,6 @@ cudnnRNNForward(cudnnHandle_t handle,
                 size_t reserveSpaceSize,
                 void *reserveSpace);
 
-/* RNN FIND API */
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetRNNAlgorithmDescriptor(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, cudnnAlgorithmDescriptor_t algoDesc);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetRNNForwardInferenceAlgorithmMaxCount(cudnnHandle_t handle, const cudnnRNNDescriptor_t rnnDesc, int *count);
-
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnFindRNNForwardInferenceAlgorithmEx(cudnnHandle_t handle,
-                                        const cudnnRNNDescriptor_t rnnDesc,
-                                        const int seqLength,
-                                        const cudnnTensorDescriptor_t *xDesc,
-                                        const void *x,
-                                        const cudnnTensorDescriptor_t hxDesc,
-                                        const void *hx,
-                                        const cudnnTensorDescriptor_t cxDesc,
-                                        const void *cx,
-                                        const cudnnFilterDescriptor_t wDesc,
-                                        const void *w,
-                                        const cudnnTensorDescriptor_t *yDesc,
-                                        void *y,
-                                        const cudnnTensorDescriptor_t hyDesc,
-                                        void *hy,
-                                        const cudnnTensorDescriptor_t cyDesc,
-                                        void *cy,
-                                        const float findIntensity,
-                                        const int requestedAlgoCount,
-                                        int *returnedAlgoCount,
-                                        cudnnAlgorithmPerformance_t *perfResults,
-                                        void *workspace,
-                                        size_t workSpaceSizeInBytes);
-
 /* Sequence data descriptor */
 
 typedef enum {
@@ -488,17 +284,17 @@ typedef enum {
 } cudnnSeqDataAxis_t;
 
 struct cudnnSeqDataStruct;
-typedef struct cudnnSeqDataStruct *cudnnSeqDataDescriptor_t;
+typedef struct cudnnSeqDataStruct *cudnnSeqDataDescriptor_t CUDNN_DEPRECATED;
 
 #define CUDNN_SEQDATA_DIM_COUNT 4 /* dimension count */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateSeqDataDescriptor(cudnnSeqDataDescriptor_t *seqDataDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroySeqDataDescriptor(cudnnSeqDataDescriptor_t seqDataDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetSeqDataDescriptor(cudnnSeqDataDescriptor_t seqDataDesc,
                           cudnnDataType_t dataType,
                           int nbDims,
@@ -508,7 +304,7 @@ cudnnSetSeqDataDescriptor(cudnnSeqDataDescriptor_t seqDataDesc,
                           const int seqLengthArray[],
                           void *paddingFill);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetSeqDataDescriptor(const cudnnSeqDataDescriptor_t seqDataDesc,
                           cudnnDataType_t *dataType,
                           int *nbDims,
@@ -522,9 +318,6 @@ cudnnGetSeqDataDescriptor(const cudnnSeqDataDescriptor_t seqDataDesc,
 
 /* Multihead Attention */
 
-/* Legacy type for backward compatibility */
-typedef unsigned cudnnAttnQueryMap_t;
-
 /*
  * Multi-head attention options passed via 'attnMode' in cudnnSetAttnDescriptor().
  * Use the bitwise OR operator to combine several settings listed below.  Additional
@@ -536,15 +329,15 @@ typedef unsigned cudnnAttnQueryMap_t;
 #define CUDNN_ATTN_ENABLE_PROJ_BIASES (1U << 1)  /* use biases in attention input and output projections */
 
 struct cudnnAttnStruct;
-typedef struct cudnnAttnStruct *cudnnAttnDescriptor_t;
+typedef struct cudnnAttnStruct *cudnnAttnDescriptor_t CUDNN_DEPRECATED;
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateAttnDescriptor(cudnnAttnDescriptor_t *attnDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyAttnDescriptor(cudnnAttnDescriptor_t attnDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetAttnDescriptor(cudnnAttnDescriptor_t attnDesc,
                        unsigned attnMode,
                        int nHeads,
@@ -566,7 +359,7 @@ cudnnSetAttnDescriptor(cudnnAttnDescriptor_t attnDesc,
                        int maxBatchSize,
                        int maxBeamSize);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetAttnDescriptor(cudnnAttnDescriptor_t attnDesc,
                        unsigned *attnMode,
                        int *nHeads,
@@ -588,7 +381,7 @@ cudnnGetAttnDescriptor(cudnnAttnDescriptor_t attnDesc,
                        int *maxBatchSize,
                        int *maxBeamSize);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetMultiHeadAttnBuffers(cudnnHandle_t handle,
                              const cudnnAttnDescriptor_t attnDesc,
                              size_t *weightSizeInBytes,
@@ -608,7 +401,7 @@ typedef enum {
 
 #define CUDNN_ATTN_WKIND_COUNT 8 /* Number of attention weight/bias tensors */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetMultiHeadAttnWeights(cudnnHandle_t handle,
                              const cudnnAttnDescriptor_t attnDesc,
                              cudnnMultiHeadAttnWeightKind_t wKind,
@@ -617,7 +410,7 @@ cudnnGetMultiHeadAttnWeights(cudnnHandle_t handle,
                              cudnnTensorDescriptor_t wDesc,
                              void **wAddr);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnMultiHeadAttnForward(cudnnHandle_t handle,
                           const cudnnAttnDescriptor_t attnDesc,
                           int currIdx,
@@ -646,13 +439,233 @@ cudnnMultiHeadAttnForward(cudnnHandle_t handle,
  * This function is implemented differently in each sub-library. Each sublib
  * checks whether its own version matches that of its dependencies.
  * \returns CUDNN_STATUS_SUCCESS if the version check passes,
- *          CUDNN_STATUS_VERSION_MISMATCH if the versions are inconsistent.
+ *          CUDNN_STATUS_SUBLIBRARY_VERSION_MISMATCH if the versions are inconsistent.
  */
 cudnnStatus_t CUDNNWINAPI
-cudnnAdvInferVersionCheck(void);
+cudnnAdvVersionCheck(void);
+
+typedef enum {
+    CUDNN_WGRAD_MODE_ADD = 0, /* add partial gradients to wgrad output buffers */
+    CUDNN_WGRAD_MODE_SET = 1, /* write partial gradients to wgrad output buffers */
+} cudnnWgradMode_t;
+
+cudnnStatus_t CUDNNWINAPI
+cudnnRNNBackwardData_v8(cudnnHandle_t handle,
+                        cudnnRNNDescriptor_t rnnDesc,
+                        const int32_t devSeqLengths[],
+                        cudnnRNNDataDescriptor_t yDesc,
+                        const void *y,
+                        const void *dy,
+                        cudnnRNNDataDescriptor_t xDesc,
+                        void *dx,
+                        cudnnTensorDescriptor_t hDesc,
+                        const void *hx,
+                        const void *dhy,
+                        void *dhx,
+                        cudnnTensorDescriptor_t cDesc,
+                        const void *cx,
+                        const void *dcy,
+                        void *dcx,
+                        size_t weightSpaceSize,
+                        const void *weightSpace,
+                        size_t workSpaceSize,
+                        void *workSpace,
+                        size_t reserveSpaceSize,
+                        void *reserveSpace);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnRNNBackwardWeights_v8(cudnnHandle_t handle,
+                           cudnnRNNDescriptor_t rnnDesc,
+                           cudnnWgradMode_t addGrad,
+                           const int32_t devSeqLengths[],
+                           cudnnRNNDataDescriptor_t xDesc,
+                           const void *x,
+                           cudnnTensorDescriptor_t hDesc,
+                           const void *hx,
+                           cudnnRNNDataDescriptor_t yDesc,
+                           const void *y,
+                           size_t weightSpaceSize,
+                           void *dweightSpace,
+                           size_t workSpaceSize,
+                           void *workSpace,
+                           size_t reserveSpaceSize,
+                           void *reserveSpace);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnMultiHeadAttnBackwardData(cudnnHandle_t handle,
+                               const cudnnAttnDescriptor_t attnDesc,
+                               const int loWinIdx[],
+                               const int hiWinIdx[],
+                               const int devSeqLengthsDQDO[],
+                               const int devSeqLengthsDKDV[],
+                               const cudnnSeqDataDescriptor_t doDesc,
+                               const void *dout,
+                               const cudnnSeqDataDescriptor_t dqDesc,
+                               void *dqueries,
+                               const void *queries,
+                               const cudnnSeqDataDescriptor_t dkDesc,
+                               void *dkeys,
+                               const void *keys,
+                               const cudnnSeqDataDescriptor_t dvDesc,
+                               void *dvalues,
+                               const void *values,
+                               size_t weightSizeInBytes,
+                               const void *weights,
+                               size_t workSpaceSizeInBytes,
+                               void *workSpace,
+                               size_t reserveSpaceSizeInBytes,
+                               void *reserveSpace);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnMultiHeadAttnBackwardWeights(cudnnHandle_t handle,
+                                  const cudnnAttnDescriptor_t attnDesc,
+                                  cudnnWgradMode_t addGrad,
+                                  const cudnnSeqDataDescriptor_t qDesc,
+                                  const void *queries,
+                                  const cudnnSeqDataDescriptor_t kDesc,
+                                  const void *keys,
+                                  const cudnnSeqDataDescriptor_t vDesc,
+                                  const void *values,
+                                  const cudnnSeqDataDescriptor_t doDesc,
+                                  const void *dout,
+                                  size_t weightSizeInBytes,
+                                  const void *weights,
+                                  void *dweights,
+                                  size_t workSpaceSizeInBytes,
+                                  void *workSpace,
+                                  size_t reserveSpaceSizeInBytes,
+                                  void *reserveSpace);
+
+/*
+* CTC (Connectionist Temporal Classification) loss descriptor create/destory/set/get functions
+*/
+/* Input normalization mode for loss function */
+typedef enum {
+    CUDNN_LOSS_NORMALIZATION_NONE    = 0,
+    CUDNN_LOSS_NORMALIZATION_SOFTMAX = 1,
+} cudnnLossNormalizationMode_t;
+
+cudnnStatus_t CUDNNWINAPI
+cudnnCreateCTCLossDescriptor(cudnnCTCLossDescriptor_t *ctcLossDesc);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnSetCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t compType);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnSetCTCLossDescriptorEx(cudnnCTCLossDescriptor_t ctcLossDesc,
+                            cudnnDataType_t compType,
+                            cudnnLossNormalizationMode_t normMode,
+                            cudnnNanPropagation_t gradMode);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnSetCTCLossDescriptor_v8(cudnnCTCLossDescriptor_t ctcLossDesc,
+                             cudnnDataType_t compType,
+                             cudnnLossNormalizationMode_t normMode,
+                             cudnnNanPropagation_t gradMode,
+                             int maxLabelLength);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnSetCTCLossDescriptor_v9(cudnnCTCLossDescriptor_t ctcLossDesc,
+                             cudnnDataType_t compType,
+                             cudnnLossNormalizationMode_t normMode,
+                             cudnnCTCGradMode_t ctcGradMode,
+                             int maxLabelLength);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t *compType);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossDescriptorEx(cudnnCTCLossDescriptor_t ctcLossDesc,
+                            cudnnDataType_t *compType,
+                            cudnnLossNormalizationMode_t *normMode,
+                            cudnnNanPropagation_t *gradMode);
+
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossDescriptor_v8(cudnnCTCLossDescriptor_t ctcLossDesc,
+                             cudnnDataType_t *compType,
+                             cudnnLossNormalizationMode_t *normMode,
+                             cudnnNanPropagation_t *gradMode,
+                             int *maxLabelLength);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossDescriptor_v9(cudnnCTCLossDescriptor_t ctcLossDesc,
+                             cudnnDataType_t *compType,
+                             cudnnLossNormalizationMode_t *normMode,
+                             cudnnCTCGradMode_t *ctcGradMode,
+                             int *maxLabelLength);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnDestroyCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc);
+
+/* return the ctc costs and gradients, given the probabilities and labels */
+cudnnStatus_t CUDNNWINAPI
+cudnnCTCLoss(
+    cudnnHandle_t handle,
+    const cudnnTensorDescriptor_t
+        probsDesc,     /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the
+                          mini batch size, A is the alphabet size)  */
+    const void *probs, /* probabilities after softmax, in GPU memory */
+    const int hostLabels[],                      /* labels, in CPU memory */
+    const int hostLabelLengths[],                /* the length of each label, in CPU memory */
+    const int hostInputLengths[],                /* the lengths of timing steps in each batch, in CPU memory */
+    void *costs,                                 /* the returned costs of CTC, in GPU memory */
+    const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A */
+    void *gradients,         /* the returned CTC gradients, in GPU memory, to compute costs only, set it to NULL */
+    cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */
+    cudnnCTCLossDescriptor_t ctcLossDesc,
+    void *workspace,              /* pointer to the workspace, in GPU memory */
+    size_t workSpaceSizeInBytes); /* size of the workspace */
+
+/* return the ctc costs and gradients, given the probabilities and labels */
+cudnnStatus_t CUDNNWINAPI
+cudnnCTCLoss_v8(
+    cudnnHandle_t handle,
+    cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */
+    cudnnCTCLossDescriptor_t ctcLossDesc,
+    const cudnnTensorDescriptor_t
+        probsDesc,     /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the
+                          mini batch size, A is the alphabet size)  */
+    const void *probs, /* probabilities after softmax, in GPU memory */
+    const int labels[],                          /* labels, in GPU memory */
+    const int labelLengths[],                    /* the length of each label, in GPU memory */
+    const int inputLengths[],                    /* the lengths of timing steps in each batch, in GPU memory */
+    void *costs,                                 /* the returned costs of CTC, in GPU memory */
+    const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A */
+    void *gradients,             /* the returned CTC gradients, in GPU memory, to compute costs only, set it to NULL */
+    size_t workSpaceSizeInBytes, /* size of the workspace */
+    void *workspace);            /* pointer to the workspace, in GPU memory */
+
+/* return the workspace size needed for ctc */
+cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossWorkspaceSize(
+    cudnnHandle_t handle,
+    const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the
+                                                timing steps, N is the mini batch size, A is the alphabet size) */
+    const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the
+                                                    dimensions are T,N,A. To compute costs
+                                                    only, set it to NULL */
+    const int *labels,                           /* labels, in CPU memory */
+    const int *labelLengths,                     /* the length of each label, in CPU memory */
+    const int *inputLengths,                     /* the lengths of timing steps in each batch, in CPU memory */
+    cudnnCTCLossAlgo_t algo,                     /* algorithm selected, supported now 0 and 1 */
+    cudnnCTCLossDescriptor_t ctcLossDesc,
+    size_t *sizeInBytes); /* pointer to the returned workspace size */
+
+/* return the workspace size needed for ctc */
+cudnnStatus_t CUDNNWINAPI
+cudnnGetCTCLossWorkspaceSize_v8(
+    cudnnHandle_t handle,
+    cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */
+    cudnnCTCLossDescriptor_t ctcLossDesc,
+    const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the
+                                                timing steps, N is the mini batch size, A is the alphabet size) */
+    const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the
+                                                    dimensions are T,N,A. To compute costs
+                                                    only, set it to NULL */
+    size_t *sizeInBytes);                        /* pointer to the returned workspace size */
 
 #if defined(__cplusplus)
 }
 #endif
 
-#endif /* CUDNN_ADV_INFER_H_ */
+#endif /* CUDNN_ADV_H_ */