nvidia-cudnn-cu12 8.9.7.29__py3-none-win_amd64.whl → 9.1.0.70__py3-none-win_amd64.whl

nvidia/cudnn/include/{cudnn_ops_infer.h → cudnn_ops.h}

@@ -48,172 +48,41 @@
  */
 
 /*
- *  cudnn_ops_infer : cuDNN's basic definitions and inference operations.
+ *  cudnn_ops : cuDNN's basic definitions and basic operations.
  */
 
-#if !defined(CUDNN_OPS_INFER_H_)
-#define CUDNN_OPS_INFER_H_
+#if !defined(CUDNN_OPS_H_)
+#define CUDNN_OPS_H_
 
-#include <cuda_runtime.h>
 #include <stdint.h>
 
 #include "cudnn_version.h"
+#include "cudnn_graph.h"
 
 /* These version numbers are autogenerated, do not edit manually. */
-#define CUDNN_OPS_INFER_MAJOR 8
-#define CUDNN_OPS_INFER_MINOR 9
-#define CUDNN_OPS_INFER_PATCH 7
+#define CUDNN_OPS_MAJOR 9
+#define CUDNN_OPS_MINOR 1
+#define CUDNN_OPS_PATCH 0
 
-#if (CUDNN_OPS_INFER_MAJOR != CUDNN_MAJOR) || (CUDNN_OPS_INFER_MINOR != CUDNN_MINOR) || \
-    (CUDNN_OPS_INFER_PATCH != CUDNN_PATCHLEVEL)
+#if (CUDNN_OPS_MAJOR != CUDNN_MAJOR) || (CUDNN_OPS_MINOR != CUDNN_MINOR) || (CUDNN_OPS_PATCH != CUDNN_PATCHLEVEL)
 #error Version mismatch in cuDNN OPS INFER!!!
 #endif
 
-#ifndef CUDNNWINAPI
-#ifdef _WIN32
-#define CUDNNWINAPI __stdcall
-#else
-#define CUDNNWINAPI
-#endif
-#endif
-
-/* Warnings for deprecated API-s are enabled using the CUDNN_WARN_DEPRECATED macro */
-#if defined(CUDNN_WARN_DEPRECATED) && (defined(__GNUC__) || defined(__clang__))
-/* GCC, Intel C/C++, Cray C/C++, CLANG, IBM XL C/C++ little endian */
-#define CUDNN_DEPRECATED __attribute__((deprecated))
-#elif defined(CUDNN_WARN_DEPRECATED) && defined(_MSC_VER)
-/* Microsoft Visual C++ */
-#define CUDNN_DEPRECATED __declspec(deprecated)
-#elif defined(CUDNN_WARN_DEPRECATED) && (__cplusplus >= 201402L)
-/* C++14 compilers */
-#define CUDNN_DEPRECATED [[deprecated]]
-#else
-/* No support for the deprecated attribute */
-#define CUDNN_DEPRECATED
-#endif
-
 #if defined(__cplusplus)
 extern "C" {
 #endif
 
-struct cudnnContext;
-typedef struct cudnnContext *cudnnHandle_t;
-
-size_t CUDNNWINAPI
-cudnnGetVersion(void);
-
-size_t CUDNNWINAPI
-cudnnGetMaxDeviceVersion(void);
-
-/* Returns CUDA Runtime version statically linked against cudnn */
-size_t CUDNNWINAPI
-cudnnGetCudartVersion(void);
-
-/*
- * CUDNN return codes
- */
-typedef enum {
-    CUDNN_STATUS_SUCCESS                      = 0,
-    CUDNN_STATUS_NOT_INITIALIZED              = 1,
-    CUDNN_STATUS_ALLOC_FAILED                 = 2,
-    CUDNN_STATUS_BAD_PARAM                    = 3,
-    CUDNN_STATUS_INTERNAL_ERROR               = 4,
-    CUDNN_STATUS_INVALID_VALUE                = 5,
-    CUDNN_STATUS_ARCH_MISMATCH                = 6,
-    CUDNN_STATUS_MAPPING_ERROR                = 7,
-    CUDNN_STATUS_EXECUTION_FAILED             = 8,
-    CUDNN_STATUS_NOT_SUPPORTED                = 9,
-    CUDNN_STATUS_LICENSE_ERROR                = 10,
-    CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING = 11,
-    CUDNN_STATUS_RUNTIME_IN_PROGRESS          = 12,
-    CUDNN_STATUS_RUNTIME_FP_OVERFLOW          = 13,
-    CUDNN_STATUS_VERSION_MISMATCH             = 14,
-} cudnnStatus_t;
-
-/* human-readable error messages */
-const char *CUDNNWINAPI
-cudnnGetErrorString(cudnnStatus_t status);
-
-/* Forward definition in this version only */
-typedef struct cudnnRuntimeTag_t cudnnRuntimeTag_t;
-
-typedef enum {
-    CUDNN_ERRQUERY_RAWCODE     = 0,
-    CUDNN_ERRQUERY_NONBLOCKING = 1,
-    CUDNN_ERRQUERY_BLOCKING    = 2,
-} cudnnErrQueryMode_t;
-
-cudnnStatus_t CUDNNWINAPI
-cudnnQueryRuntimeError(cudnnHandle_t handle, cudnnStatus_t *rstatus, cudnnErrQueryMode_t mode, cudnnRuntimeTag_t *tag);
-
-#ifndef __LIBRARY_TYPES_H__
-
-typedef enum libraryPropertyType_t { MAJOR_VERSION, MINOR_VERSION, PATCH_LEVEL } libraryPropertyType;
-
-#endif
-
-cudnnStatus_t CUDNNWINAPI
-cudnnGetProperty(libraryPropertyType type, int *value);
-
-cudnnStatus_t CUDNNWINAPI
-cudnnCreate(cudnnHandle_t *handle);
-cudnnStatus_t CUDNNWINAPI
-cudnnDestroy(cudnnHandle_t handle);
-cudnnStatus_t CUDNNWINAPI
-cudnnSetStream(cudnnHandle_t handle, cudaStream_t streamId);
-cudnnStatus_t CUDNNWINAPI
-cudnnGetStream(cudnnHandle_t handle, cudaStream_t *streamId);
-
 /* Data structures to represent Image/Filter and the Neural Network Layer */
 typedef struct cudnnTensorStruct *cudnnTensorDescriptor_t;
-typedef struct cudnnPoolingStruct *cudnnPoolingDescriptor_t;
-typedef struct cudnnFilterStruct *cudnnFilterDescriptor_t;
+typedef struct cudnnPoolingStruct *cudnnPoolingDescriptor_t CUDNN_DEPRECATED;
+typedef struct cudnnFilterStruct *cudnnFilterDescriptor_t CUDNN_DEPRECATED;
 typedef struct cudnnLRNStruct *cudnnLRNDescriptor_t;
-typedef struct cudnnActivationStruct *cudnnActivationDescriptor_t;
+typedef struct cudnnActivationStruct *cudnnActivationDescriptor_t CUDNN_DEPRECATED;
 typedef struct cudnnSpatialTransformerStruct *cudnnSpatialTransformerDescriptor_t;
-typedef struct cudnnOpTensorStruct *cudnnOpTensorDescriptor_t;
-typedef struct cudnnReduceTensorStruct *cudnnReduceTensorDescriptor_t;
+typedef struct cudnnOpTensorStruct *cudnnOpTensorDescriptor_t CUDNN_DEPRECATED;
+typedef struct cudnnReduceTensorStruct *cudnnReduceTensorDescriptor_t CUDNN_DEPRECATED;
 typedef struct cudnnCTCLossStruct *cudnnCTCLossDescriptor_t;
-typedef struct cudnnTensorTransformStruct *cudnnTensorTransformDescriptor_t;
-/*
- * CUDNN data type
- */
-typedef enum {
-    CUDNN_DATA_FLOAT              = 0,
-    CUDNN_DATA_DOUBLE             = 1,
-    CUDNN_DATA_HALF               = 2,
-    CUDNN_DATA_INT8               = 3,
-    CUDNN_DATA_INT32              = 4,
-    CUDNN_DATA_INT8x4             = 5,
-    CUDNN_DATA_UINT8              = 6,
-    CUDNN_DATA_UINT8x4            = 7,
-    CUDNN_DATA_INT8x32            = 8,
-    CUDNN_DATA_BFLOAT16           = 9,
-    CUDNN_DATA_INT64              = 10,
-    CUDNN_DATA_BOOLEAN            = 11,
-    CUDNN_DATA_FP8_E4M3           = 12,
-    CUDNN_DATA_FP8_E5M2           = 13,
-    CUDNN_DATA_FAST_FLOAT_FOR_FP8 = 14,
-} cudnnDataType_t;
-
-/*
- * CUDNN math type
- */
-typedef enum {
-    CUDNN_DEFAULT_MATH                    = 0,
-    CUDNN_TENSOR_OP_MATH                  = 1,
-    CUDNN_TENSOR_OP_MATH_ALLOW_CONVERSION = 2,
-    CUDNN_FMA_MATH                        = 3,
-} cudnnMathType_t;
-
-/*
- * CUDNN propagate Nan
- */
-typedef enum {
-    CUDNN_NOT_PROPAGATE_NAN = 0,
-    CUDNN_PROPAGATE_NAN     = 1,
-} cudnnNanPropagation_t;
-
+typedef struct cudnnTensorTransformStruct *cudnnTensorTransformDescriptor_t CUDNN_DEPRECATED;
 /*
  * CUDNN Determinism
  */
@@ -222,19 +91,10 @@ typedef enum {
     CUDNN_DETERMINISTIC     = 1,
 } cudnnDeterminism_t;
 
-/* Maximum supported number of tensor dimensions */
-#define CUDNN_DIM_MAX 8
-
 /* Create an instance of a generic Tensor descriptor */
 cudnnStatus_t CUDNNWINAPI
 cudnnCreateTensorDescriptor(cudnnTensorDescriptor_t *tensorDesc);
 
-typedef enum {
-    CUDNN_TENSOR_NCHW        = 0, /* row major (wStride = 1, hStride = w) */
-    CUDNN_TENSOR_NHWC        = 1, /* feature maps interleaved ( cStride = 1 )*/
-    CUDNN_TENSOR_NCHW_VECT_C = 2, /* each image point is vector of element of C, vector length in data type */
-} cudnnTensorFormat_t;
-
 cudnnStatus_t CUDNNWINAPI
 cudnnSetTensor4dDescriptor(cudnnTensorDescriptor_t tensorDesc,
                            cudnnTensorFormat_t format,
@@ -327,18 +187,18 @@ typedef enum {
 } cudnnFoldingDirection_t;
 
 /** Create a destination descriptor for cudnnTransformTensor */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnInitTransformDest(const cudnnTensorTransformDescriptor_t transformDesc,
                        const cudnnTensorDescriptor_t srcDesc,
                        cudnnTensorDescriptor_t destDesc,
                        size_t *destSizeInBytes);
 
 /** Create an empty tensor transform descriptor */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateTensorTransformDescriptor(cudnnTensorTransformDescriptor_t *transformDesc);
 
 /** Initialize a previously created tensor transform descriptor. */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetTensorTransformDescriptor(cudnnTensorTransformDescriptor_t transformDesc,
                                   const uint32_t nbDims,
                                   const cudnnTensorFormat_t destFormat,
@@ -351,7 +211,7 @@ cudnnSetTensorTransformDescriptor(cudnnTensorTransformDescriptor_t transformDesc
  * Retrieves the values stored in a previously initialized tensor transform
  * descriptor.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetTensorTransformDescriptor(cudnnTensorTransformDescriptor_t transformDesc,
                                   uint32_t nbDimsRequested,
                                   cudnnTensorFormat_t *destFormat,
@@ -363,11 +223,11 @@ cudnnGetTensorTransformDescriptor(cudnnTensorTransformDescriptor_t transformDesc
 /**
  * Destroys a previously created tensor transform descriptor.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyTensorTransformDescriptor(cudnnTensorTransformDescriptor_t transformDesc);
 
 /* Tensor layout conversion helper (y = alpha * x + beta * y) */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnTransformTensor(cudnnHandle_t handle,
                      const void *alpha,
                      const cudnnTensorDescriptor_t xDesc,
@@ -376,7 +236,7 @@ cudnnTransformTensor(cudnnHandle_t handle,
                      const cudnnTensorDescriptor_t yDesc,
                      void *y);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnTransformTensorEx(cudnnHandle_t handle,
                        const cudnnTensorTransformDescriptor_t transDesc,
                        const void *alpha,
@@ -387,7 +247,7 @@ cudnnTransformTensorEx(cudnnHandle_t handle,
                        void *destData);
 
 /* Tensor Bias addition : C = alpha * A + beta * C  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnAddTensor(cudnnHandle_t handle,
                const void *alpha,
                const cudnnTensorDescriptor_t aDesc,
@@ -408,27 +268,27 @@ typedef enum {
     CUDNN_OP_TENSOR_NOT  = 5,
 } cudnnOpTensorOp_t;
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateOpTensorDescriptor(cudnnOpTensorDescriptor_t *opTensorDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetOpTensorDescriptor(cudnnOpTensorDescriptor_t opTensorDesc,
                            cudnnOpTensorOp_t opTensorOp,
                            cudnnDataType_t opTensorCompType,
                            cudnnNanPropagation_t opTensorNanOpt);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetOpTensorDescriptor(const cudnnOpTensorDescriptor_t opTensorDesc,
                            cudnnOpTensorOp_t *opTensorOp,
                            cudnnDataType_t *opTensorCompType,
                            cudnnNanPropagation_t *opTensorNanOpt);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyOpTensorDescriptor(cudnnOpTensorDescriptor_t opTensorDesc);
 
 /* Tensor operation : C = op( alpha1 * A, alpha2 * B ) + beta * C */
 /* B tensor is ignored for CUDNN_OP_TENSOR_SQRT, CUDNN_OP_TENSOR_NOT. */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnOpTensor(cudnnHandle_t handle,
               const cudnnOpTensorDescriptor_t opTensorDesc,
               const void *alpha1,
@@ -441,28 +301,13 @@ cudnnOpTensor(cudnnHandle_t handle,
               const cudnnTensorDescriptor_t cDesc,
               void *C);
 
-/*
- * CUDNN ReduceTensor op type
- */
-typedef enum {
-    CUDNN_REDUCE_TENSOR_ADD          = 0,
-    CUDNN_REDUCE_TENSOR_MUL          = 1,
-    CUDNN_REDUCE_TENSOR_MIN          = 2,
-    CUDNN_REDUCE_TENSOR_MAX          = 3,
-    CUDNN_REDUCE_TENSOR_AMAX         = 4,
-    CUDNN_REDUCE_TENSOR_AVG          = 5,
-    CUDNN_REDUCE_TENSOR_NORM1        = 6,
-    CUDNN_REDUCE_TENSOR_NORM2        = 7,
-    CUDNN_REDUCE_TENSOR_MUL_NO_ZEROS = 8,
-} cudnnReduceTensorOp_t;
-
 /*
  * CUDNN ReduceTensor indices type
  */
 typedef enum {
     CUDNN_REDUCE_TENSOR_NO_INDICES        = 0,
     CUDNN_REDUCE_TENSOR_FLATTENED_INDICES = 1,
-} cudnnReduceTensorIndices_t;
+} cudnnReduceTensorIndices_t CUDNN_DEPRECATED;
 
 /*
  * CUDNN tensor indices type size (all unsigned)
@@ -473,12 +318,12 @@ typedef enum {
     CUDNN_64BIT_INDICES = 1,
     CUDNN_16BIT_INDICES = 2,
     CUDNN_8BIT_INDICES  = 3,
-} cudnnIndicesType_t;
+} cudnnIndicesType_t CUDNN_DEPRECATED;
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateReduceTensorDescriptor(cudnnReduceTensorDescriptor_t *reduceTensorDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetReduceTensorDescriptor(cudnnReduceTensorDescriptor_t reduceTensorDesc,
                                cudnnReduceTensorOp_t reduceTensorOp,
                                cudnnDataType_t reduceTensorCompType,
@@ -486,7 +331,7 @@ cudnnSetReduceTensorDescriptor(cudnnReduceTensorDescriptor_t reduceTensorDesc,
                                cudnnReduceTensorIndices_t reduceTensorIndices,
                                cudnnIndicesType_t reduceTensorIndicesType);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetReduceTensorDescriptor(const cudnnReduceTensorDescriptor_t reduceTensorDesc,
                                cudnnReduceTensorOp_t *reduceTensorOp,
                                cudnnDataType_t *reduceTensorCompType,
@@ -494,12 +339,12 @@ cudnnGetReduceTensorDescriptor(const cudnnReduceTensorDescriptor_t reduceTensorD
                                cudnnReduceTensorIndices_t *reduceTensorIndices,
                                cudnnIndicesType_t *reduceTensorIndicesType);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyReduceTensorDescriptor(cudnnReduceTensorDescriptor_t reduceTensorDesc);
 
 /* Helper function to return the minimum size of the index space to be passed to the reduction given the input and
  * output tensors */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetReductionIndicesSize(cudnnHandle_t handle,
                              const cudnnReduceTensorDescriptor_t reduceTensorDesc,
                              const cudnnTensorDescriptor_t aDesc,
@@ -508,7 +353,7 @@ cudnnGetReductionIndicesSize(cudnnHandle_t handle,
 
 /* Helper function to return the minimum size of the workspace to be passed to the reduction given the input and output
  * tensors */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetReductionWorkspaceSize(cudnnHandle_t handle,
                                const cudnnReduceTensorDescriptor_t reduceTensorDesc,
                                const cudnnTensorDescriptor_t aDesc,
@@ -518,7 +363,7 @@ cudnnGetReductionWorkspaceSize(cudnnHandle_t handle,
 /* Tensor operation : C = reduce op( alpha * A ) + beta * C */
 /* The NaN propagation enum applies to only the min and max reduce ops; the other reduce ops propagate NaN as usual. */
 /* The indices space is ignored for reduce ops other than min or max. */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnReduceTensor(cudnnHandle_t handle,
                   const cudnnReduceTensorDescriptor_t reduceTensorDesc,
                   void *indices,
@@ -537,14 +382,14 @@ cudnnStatus_t CUDNNWINAPI
 cudnnSetTensor(cudnnHandle_t handle, const cudnnTensorDescriptor_t yDesc, void *y, const void *valuePtr);
 
 /* Scale all values of a tensor by a given factor : y[i] = alpha * y[i] */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnScaleTensor(cudnnHandle_t handle, const cudnnTensorDescriptor_t yDesc, void *y, const void *alpha);
 
 /* Create an instance of FilterStruct */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateFilterDescriptor(cudnnFilterDescriptor_t *filterDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetFilter4dDescriptor(cudnnFilterDescriptor_t filterDesc,
                            cudnnDataType_t dataType, /* image data type */
                            cudnnTensorFormat_t format,
@@ -553,7 +398,7 @@ cudnnSetFilter4dDescriptor(cudnnFilterDescriptor_t filterDesc,
                            int h,  /* height of each input filter */
                            int w); /* width of  each input filter */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetFilter4dDescriptor(const cudnnFilterDescriptor_t filterDesc,
                            cudnnDataType_t *dataType, /* image data type */
                            cudnnTensorFormat_t *format,
@@ -562,24 +407,24 @@ cudnnGetFilter4dDescriptor(const cudnnFilterDescriptor_t filterDesc,
                            int *h,  /* height of each input filter */
                            int *w); /* width of  each input filter */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetFilterNdDescriptor(cudnnFilterDescriptor_t filterDesc,
                            cudnnDataType_t dataType, /* image data type */
                            cudnnTensorFormat_t format,
                            int nbDims,
                            const int filterDimA[]);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetFilterNdDescriptor(const cudnnFilterDescriptor_t filterDesc,
                            int nbDimsRequested,
                            cudnnDataType_t *dataType, /* image data type */
                            cudnnTensorFormat_t *format,
                            int *nbDims,
                            int filterDimA[]);
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetFilterSizeInBytes(const cudnnFilterDescriptor_t filterDesc, size_t *size);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnTransformFilter(cudnnHandle_t handle,
                      const cudnnTensorTransformDescriptor_t transDesc,
                      const void *alpha,
@@ -589,7 +434,7 @@ cudnnTransformFilter(cudnnHandle_t handle,
                      const cudnnFilterDescriptor_t destDesc,
                      void *destData);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyFilterDescriptor(cudnnFilterDescriptor_t filterDesc);
 
 /*
@@ -628,13 +473,13 @@ typedef enum {
     CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING = 1, /* count for average includes padded values */
     CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING = 2, /* count for average does not include padded values */
     CUDNN_POOLING_MAX_DETERMINISTIC             = 3
-} cudnnPoolingMode_t;
+} cudnnPoolingMode_t CUDNN_DEPRECATED;
 
 /* Create an instance of pooling descriptor */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreatePoolingDescriptor(cudnnPoolingDescriptor_t *poolingDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetPooling2dDescriptor(cudnnPoolingDescriptor_t poolingDesc,
                             cudnnPoolingMode_t mode,
                             cudnnNanPropagation_t maxpoolingNanOpt,
@@ -645,7 +490,7 @@ cudnnSetPooling2dDescriptor(cudnnPoolingDescriptor_t poolingDesc,
                             int verticalStride,
                             int horizontalStride);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetPooling2dDescriptor(const cudnnPoolingDescriptor_t poolingDesc,
                             cudnnPoolingMode_t *mode,
                             cudnnNanPropagation_t *maxpoolingNanOpt,
@@ -656,7 +501,7 @@ cudnnGetPooling2dDescriptor(const cudnnPoolingDescriptor_t poolingDesc,
                             int *verticalStride,
                             int *horizontalStride);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetPoolingNdDescriptor(cudnnPoolingDescriptor_t poolingDesc,
                             const cudnnPoolingMode_t mode,
                             const cudnnNanPropagation_t maxpoolingNanOpt,
@@ -665,7 +510,7 @@ cudnnSetPoolingNdDescriptor(cudnnPoolingDescriptor_t poolingDesc,
                             const int paddingA[],
                             const int strideA[]);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetPoolingNdDescriptor(const cudnnPoolingDescriptor_t poolingDesc,
                             int nbDimsRequested,
                             cudnnPoolingMode_t *mode,
@@ -675,13 +520,13 @@ cudnnGetPoolingNdDescriptor(const cudnnPoolingDescriptor_t poolingDesc,
                             int paddingA[],
                             int strideA[]);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetPoolingNdForwardOutputDim(const cudnnPoolingDescriptor_t poolingDesc,
                                   const cudnnTensorDescriptor_t inputTensorDesc,
                                   int nbDims,
                                   int outputTensorDimA[]);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetPooling2dForwardOutputDim(const cudnnPoolingDescriptor_t poolingDesc,
                                   const cudnnTensorDescriptor_t inputTensorDesc,
                                   int *n,
@@ -690,13 +535,13 @@ cudnnGetPooling2dForwardOutputDim(const cudnnPoolingDescriptor_t poolingDesc,
                                   int *w);
 
 /* Destroy an instance of pooling descriptor */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyPoolingDescriptor(cudnnPoolingDescriptor_t poolingDesc);
 
 /* Pooling functions: All of the form "output = alpha * Op(inputs) + beta * output" */
 
 /* Function to perform forward pooling */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnPoolingForward(cudnnHandle_t handle,
                     const cudnnPoolingDescriptor_t poolingDesc,
                     const void *alpha,
@@ -706,46 +551,33 @@ cudnnPoolingForward(cudnnHandle_t handle,
                     const cudnnTensorDescriptor_t yDesc,
                     void *y);
 
-/*
- * activation mode
- */
-typedef enum {
-    CUDNN_ACTIVATION_SIGMOID      = 0,
-    CUDNN_ACTIVATION_RELU         = 1,
-    CUDNN_ACTIVATION_TANH         = 2,
-    CUDNN_ACTIVATION_CLIPPED_RELU = 3,
-    CUDNN_ACTIVATION_ELU          = 4,
-    CUDNN_ACTIVATION_IDENTITY     = 5,
-    CUDNN_ACTIVATION_SWISH        = 6
-} cudnnActivationMode_t;
-
 /* Activation functions: All of the form "output = alpha * Op(inputs) + beta * output" */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnCreateActivationDescriptor(cudnnActivationDescriptor_t *activationDesc);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetActivationDescriptor(cudnnActivationDescriptor_t activationDesc,
                              cudnnActivationMode_t mode,
                              cudnnNanPropagation_t reluNanOpt,
                              double coef); /* ceiling for clipped RELU, alpha for ELU */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetActivationDescriptor(const cudnnActivationDescriptor_t activationDesc,
                              cudnnActivationMode_t *mode,
                              cudnnNanPropagation_t *reluNanOpt,
                              double *coef); /* ceiling for clipped RELU, alpha for ELU */
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnSetActivationDescriptorSwishBeta(cudnnActivationDescriptor_t activationDesc, double swish_beta);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnGetActivationDescriptorSwishBeta(cudnnActivationDescriptor_t activationDesc, double *swish_beta);
 
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDestroyActivationDescriptor(cudnnActivationDescriptor_t activationDesc);
 
 /* Function to perform forward activation  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnActivationForward(cudnnHandle_t handle,
                        cudnnActivationDescriptor_t activationDesc,
                        const void *alpha,
@@ -835,7 +667,7 @@ typedef enum {
      * May be faster than CUDNN_BATCHNORM_SPATIAL but imposes some limits on the range of values
      */
     CUDNN_BATCHNORM_SPATIAL_PERSISTENT = 2,
-} cudnnBatchNormMode_t;
+} cudnnBatchNormMode_t CUDNN_DEPRECATED;
 
 #define CUDNN_BN_MIN_EPSILON 0.0 /* Minimum epsilon allowed to be used in the Batch Normalization formula */
 
@@ -844,7 +676,7 @@ typedef enum {
  * scale, invVariance, bnBias, bnScale tensors. Use this tensor desc for
  * bnScaleBiasMeanVarDesc and bnScaleBiasDiffDesc in Batch Normalization forward and backward functions.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDeriveBNTensorDescriptor(cudnnTensorDescriptor_t derivedBnDesc,
                               const cudnnTensorDescriptor_t xDesc,
                               cudnnBatchNormMode_t mode);
@@ -853,7 +685,7 @@ typedef enum {
     CUDNN_BATCHNORM_OPS_BN                = 0, /* do batch normalization only */
     CUDNN_BATCHNORM_OPS_BN_ACTIVATION     = 1, /* do batchNorm, then activation */
     CUDNN_BATCHNORM_OPS_BN_ADD_ACTIVATION = 2, /* do batchNorm, then elemWiseAdd, then activation */
-} cudnnBatchNormOps_t;
+} cudnnBatchNormOps_t CUDNN_DEPRECATED;
 
 /*
  * Performs Batch Normalization during Inference:
@@ -862,7 +694,7 @@ typedef enum {
  * according to spatial or per-activation mode. Refer to cudnnBatchNormalizationForwardTraining
  * above for notes on function arguments.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnBatchNormalizationForwardInference(cudnnHandle_t handle,
                                         cudnnBatchNormMode_t mode,
                                         const void *alpha, /* alpha[0] = result blend factor */
@@ -884,16 +716,16 @@ typedef enum {
 
     /* bnScale, bnBias tensor dims are 1xCx1x1 (one value per C-dim normalized over Nx1xHxW subtensors) */
     CUDNN_NORM_PER_CHANNEL = 1,
-} cudnnNormMode_t;
+} cudnnNormMode_t CUDNN_DEPRECATED;
 
-typedef enum { CUDNN_NORM_ALGO_STANDARD = 0, CUDNN_NORM_ALGO_PERSIST = 1 } cudnnNormAlgo_t;
+typedef enum { CUDNN_NORM_ALGO_STANDARD = 0, CUDNN_NORM_ALGO_PERSIST = 1 } cudnnNormAlgo_t CUDNN_DEPRECATED;
 
 /*
  * Derives a tensor descriptor from layer data descriptor for Normalization
  * scale, invVariance, bnBias, bnScale tensors. Use this tensor desc for
  * normScaleBiasMeanVarDesc and normScaleBiasDiffDesc in Normalization forward and backward functions.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnDeriveNormTensorDescriptor(cudnnTensorDescriptor_t derivedNormScaleBiasDesc,
                                 cudnnTensorDescriptor_t derivedNormMeanVarDesc,
                                 const cudnnTensorDescriptor_t xDesc,
@@ -904,7 +736,7 @@ typedef enum {
     CUDNN_NORM_OPS_NORM                = 0, /* do normalization only */
     CUDNN_NORM_OPS_NORM_ACTIVATION     = 1, /* do Norm, then activation */
     CUDNN_NORM_OPS_NORM_ADD_ACTIVATION = 2, /* do Norm, then elemWiseAdd, then activation */
-} cudnnNormOps_t;
+} cudnnNormOps_t CUDNN_DEPRECATED;
 
 /*
  * Performs Normalization during Inference:
@@ -913,7 +745,7 @@ typedef enum {
  * according to per-channel or per-activation mode. Refer to cudnnNormalizationForwardTraining
  * above for notes on function arguments.
  */
-cudnnStatus_t CUDNNWINAPI
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
 cudnnNormalizationForwardInference(cudnnHandle_t handle,
                                    cudnnNormMode_t mode,
                                    cudnnNormOps_t normOps,
@@ -1021,11 +853,6 @@ cudnnDropoutForward(cudnnHandle_t handle,
                     void *reserveSpace,
                     size_t reserveSpaceSizeInBytes);
 
-/* TODO: remove */
-
-typedef struct cudnnAlgorithmStruct *cudnnAlgorithmDescriptor_t;
-typedef struct cudnnAlgorithmPerformanceStruct *cudnnAlgorithmPerformance_t;
-
 /* TODO: move these enums out to the appropriate submodule */
 typedef enum {
     CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM         = 0,
@@ -1060,124 +887,430 @@ typedef enum {
     CUDNN_CONVOLUTION_BWD_DATA_ALGO_COUNT             = 6
 } cudnnConvolutionBwdDataAlgo_t;
 
-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_CTC_LOSS_ALGO_DETERMINISTIC = 0, CUDNN_CTC_LOSS_ALGO_NON_DETERMINISTIC = 1 } cudnnCTCLossAlgo_t;
 
-/* TODO: remove */
-typedef struct cudnnAlgorithmUnionStruct {
-    union Algorithm {
-        cudnnConvolutionFwdAlgo_t convFwdAlgo;
-        cudnnConvolutionBwdFilterAlgo_t convBwdFilterAlgo;
-        cudnnConvolutionBwdDataAlgo_t convBwdDataAlgo;
-        cudnnRNNAlgo_t RNNAlgo;
-        cudnnCTCLossAlgo_t CTCLossAlgo;
-    } algo;
-} cudnnAlgorithm_t;
+/*
+ * \brief Cross-library version checker.
+ * 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_SUBLIBRARY_VERSION_MISMATCH if the versions are inconsistent.
+ */
+cudnnStatus_t CUDNNWINAPI
+cudnnOpsVersionCheck(void);
 
-CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnCreateAlgorithmDescriptor(cudnnAlgorithmDescriptor_t *algoDesc);
+/* Function to perform backward softmax */
+cudnnStatus_t CUDNNWINAPI
+cudnnSoftmaxBackward(cudnnHandle_t handle,
+                     cudnnSoftmaxAlgorithm_t algo,
+                     cudnnSoftmaxMode_t mode,
+                     const void *alpha,
+                     const cudnnTensorDescriptor_t yDesc,
+                     const void *y,
+                     const cudnnTensorDescriptor_t dyDesc,
+                     const void *dy,
+                     const void *beta,
+                     const cudnnTensorDescriptor_t dxDesc,
+                     void *dx);
 
+/* Function to perform backward pooling */
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetAlgorithmDescriptor(cudnnAlgorithmDescriptor_t algoDesc, cudnnAlgorithm_t algorithm);
+cudnnPoolingBackward(cudnnHandle_t handle,
+                     const cudnnPoolingDescriptor_t poolingDesc,
+                     const void *alpha,
+                     const cudnnTensorDescriptor_t yDesc,
+                     const void *y,
+                     const cudnnTensorDescriptor_t dyDesc,
+                     const void *dy,
+                     const cudnnTensorDescriptor_t xDesc,
+                     const void *x,
+                     const void *beta,
+                     const cudnnTensorDescriptor_t dxDesc,
+                     void *dx);
 
+/* Function to perform backward activation  */
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetAlgorithmDescriptor(const cudnnAlgorithmDescriptor_t algoDesc, cudnnAlgorithm_t *algorithm);
+cudnnActivationBackward(cudnnHandle_t handle,
+                        cudnnActivationDescriptor_t activationDesc,
+                        const void *alpha,
+                        const cudnnTensorDescriptor_t yDesc,
+                        const void *y,
+                        const cudnnTensorDescriptor_t dyDesc,
+                        const void *dy,
+                        const cudnnTensorDescriptor_t xDesc,
+                        const void *x,
+                        const void *beta,
+                        const cudnnTensorDescriptor_t dxDesc,
+                        void *dx);
+
+/* LRN cross-channel backward computation. Double parameters cast to tensor data type */
+cudnnStatus_t CUDNNWINAPI
+cudnnLRNCrossChannelBackward(cudnnHandle_t handle,
+                             cudnnLRNDescriptor_t normDesc,
+                             cudnnLRNMode_t lrnMode,
+                             const void *alpha,
+                             const cudnnTensorDescriptor_t yDesc,
+                             const void *y,
+                             const cudnnTensorDescriptor_t dyDesc,
+                             const void *dy,
+                             const cudnnTensorDescriptor_t xDesc,
+                             const void *x,
+                             const void *beta,
+                             const cudnnTensorDescriptor_t dxDesc,
+                             void *dx);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnDivisiveNormalizationBackward(cudnnHandle_t handle,
+                                   cudnnLRNDescriptor_t normDesc,
+                                   cudnnDivNormMode_t mode,
+                                   const void *alpha,
+                                   const cudnnTensorDescriptor_t xDesc, /* same desc for x, means, dy, temp, temp2 */
+                                   const void *x,
+                                   const void *means, /* if NULL, means are assumed to be zero */
+                                   const void *dy,
+                                   void *temp,
+                                   void *temp2,
+                                   const void *beta,
+                                   const cudnnTensorDescriptor_t dXdMeansDesc, /* same desc for dx, dMeans */
+                                   void *dx,                                   /* output x differential */
+                                   void *dMeans); /* output means differential, can be NULL */
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnCopyAlgorithmDescriptor(const cudnnAlgorithmDescriptor_t src, cudnnAlgorithmDescriptor_t dest);
+cudnnGetBatchNormalizationForwardTrainingExWorkspaceSize(cudnnHandle_t handle,
+                                                         cudnnBatchNormMode_t mode,
+                                                         cudnnBatchNormOps_t bnOps,
+                                                         const cudnnTensorDescriptor_t xDesc,
+                                                         const cudnnTensorDescriptor_t zDesc,
+                                                         const cudnnTensorDescriptor_t yDesc,
+                                                         const cudnnTensorDescriptor_t bnScaleBiasMeanVarDesc,
+                                                         const cudnnActivationDescriptor_t activationDesc,
+                                                         size_t *sizeInBytes);
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnDestroyAlgorithmDescriptor(cudnnAlgorithmDescriptor_t algoDesc);
+cudnnGetBatchNormalizationBackwardExWorkspaceSize(cudnnHandle_t handle,
+                                                  cudnnBatchNormMode_t mode,
+                                                  cudnnBatchNormOps_t bnOps,
+                                                  const cudnnTensorDescriptor_t xDesc,
+                                                  const cudnnTensorDescriptor_t yDesc,
+                                                  const cudnnTensorDescriptor_t dyDesc,
+                                                  const cudnnTensorDescriptor_t dzDesc,
+                                                  const cudnnTensorDescriptor_t dxDesc,
+                                                  const cudnnTensorDescriptor_t dBnScaleBiasDesc,
+                                                  const cudnnActivationDescriptor_t activationDesc,
+                                                  size_t *sizeInBytes);
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnCreateAlgorithmPerformance(cudnnAlgorithmPerformance_t *algoPerf, int numberToCreate);
-
+cudnnGetBatchNormalizationTrainingExReserveSpaceSize(cudnnHandle_t handle,
+                                                     cudnnBatchNormMode_t mode,
+                                                     cudnnBatchNormOps_t bnOps,
+                                                     const cudnnActivationDescriptor_t activationDesc,
+                                                     const cudnnTensorDescriptor_t xDesc,
+                                                     size_t *sizeInBytes);
+
+/* Computes y = BN(x). Also accumulates moving averages of mean and inverse variances */
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSetAlgorithmPerformance(cudnnAlgorithmPerformance_t algoPerf,
-                             cudnnAlgorithmDescriptor_t algoDesc,
-                             cudnnStatus_t status,
-                             float time,
-                             size_t memory);
-
+cudnnBatchNormalizationForwardTraining(
+    cudnnHandle_t handle,
+    cudnnBatchNormMode_t mode,
+
+    const void *alpha, /* alpha[0] = result blend factor */
+    const void *beta,  /* beta[0] = dest layer blend factor */
+
+    const cudnnTensorDescriptor_t xDesc,
+    const void *x, /* NxCxHxW */
+    const cudnnTensorDescriptor_t yDesc,
+    void *y, /* NxCxHxW */
+
+    /* Shared desc for the next 6 tensors in the argument list.
+       Data type to be set as follows:
+       type = (typeOf(x) == double) ? double : float
+       Dimensions for this descriptor depend on normalization mode
+       - Spatial Normalization : tensors are expected to have dims 1xCx1x1
+        (normalization is performed across NxHxW)
+       - Per-Activation Normalization : tensors are expected to have dims of 1xCxHxW
+        (normalization is performed across N) */
+    const cudnnTensorDescriptor_t bnScaleBiasMeanVarDesc,
+
+    /* 'Gamma' and 'Beta' respectively in Ioffe and Szegedy's paper's notation */
+    const void *bnScale,
+    const void *bnBias,
+
+    /* MUST use factor=1 in the very first call of a complete training cycle.
+       Use a factor=1/(1+n) at N-th call to the function to get
+       Cumulative Moving Average (CMA) behavior
+       CMA[n] = (x[1]+...+x[n])/n
+       Since CMA[n+1] = (n*CMA[n]+x[n+1])/(n+1) =
+       ((n+1)*CMA[n]-CMA[n])/(n+1) + x[n+1]/(n+1) =
+       CMA[n]*(1-1/(n+1)) + x[n+1]*1/(n+1) */
+    double exponentialAverageFactor,
+
+    /* Used in Training phase only.
+       runningMean = newMean*factor + runningMean*(1-factor) */
+    void *resultRunningMean,
+    /* Output in training mode, input in inference. Is the moving average
+       of  variance[x] (factor is applied in the same way as for runningMean) */
+    void *resultRunningVariance,
+
+    /* Has to be >= CUDNN_BN_MIN_EPSILON. Should be the same in forward and backward functions. */
+    double epsilon,
+
+    /* Optionally save intermediate results from the forward pass here
+       - can be reused to speed up backward pass. NULL if unused */
+    void *resultSaveMean,
+    void *resultSaveInvVariance);
+
+/* Computes y = relu(BN(x) + z). Also accumulates moving averages of mean and inverse variances */
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetAlgorithmPerformance(const cudnnAlgorithmPerformance_t algoPerf,
-                             cudnnAlgorithmDescriptor_t *algoDesc,
-                             cudnnStatus_t *status,
-                             float *time,
-                             size_t *memory);
-
+cudnnBatchNormalizationForwardTrainingEx(
+    cudnnHandle_t handle,
+    cudnnBatchNormMode_t mode,
+    cudnnBatchNormOps_t bnOps,
+
+    const void *alpha, /* alpha[0] = result blend factor */
+    const void *beta,  /* beta[0] = dest layer blend factor */
+
+    const cudnnTensorDescriptor_t xDesc,
+    const void *xData,
+    const cudnnTensorDescriptor_t zDesc,
+    const void *zData,
+    const cudnnTensorDescriptor_t yDesc,
+    void *yData,
+
+    const cudnnTensorDescriptor_t bnScaleBiasMeanVarDesc,
+    const void *bnScale,
+    const void *bnBias,
+
+    double exponentialAverageFactor,
+    void *resultRunningMean,
+    void *resultRunningVariance,
+
+    /* Has to be >= CUDNN_BN_MIN_EPSILON. Should be the same in forward and backward functions. */
+    double epsilon,
+
+    /* Optionally save intermediate results from the forward pass here
+       - can be reused to speed up backward pass. NULL if unused */
+    void *resultSaveMean,
+    void *resultSaveInvVariance,
+
+    cudnnActivationDescriptor_t activationDesc,
+    void *workspace,
+    size_t workSpaceSizeInBytes,
+    void *reserveSpace,
+    size_t reserveSpaceSizeInBytes);
+
+/* Performs backward pass of Batch Normalization layer. Returns x gradient,
+* bnScale gradient and bnBias gradient */
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnDestroyAlgorithmPerformance(cudnnAlgorithmPerformance_t *algoPerf, int numberToDestroy);
+cudnnBatchNormalizationBackward(cudnnHandle_t handle,
+                                cudnnBatchNormMode_t mode,
+                                const void *alphaDataDiff,
+                                const void *betaDataDiff,
+                                const void *alphaParamDiff,
+                                const void *betaParamDiff,
+                                const cudnnTensorDescriptor_t xDesc, /* same desc for x, dx, dy */
+                                const void *x,
+                                const cudnnTensorDescriptor_t dyDesc,
+                                const void *dy,
+                                const cudnnTensorDescriptor_t dxDesc,
+                                void *dx,
+                                /* Shared tensor desc for the 4 tensors below */
+                                const cudnnTensorDescriptor_t dBnScaleBiasDesc,
+                                const void *bnScale, /* bnBias doesn't affect backpropagation */
+                                /* scale and bias diff are not backpropagated below this layer */
+                                void *dBnScaleResult,
+                                void *dBnBiasResult,
+                                /* Same epsilon as forward pass */
+                                double epsilon,
+
+                                /* Optionally cached intermediate results from
+                                   forward pass */
+                                const void *savedMean,
+                                const void *savedInvVariance);
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnGetAlgorithmSpaceSize(cudnnHandle_t handle, cudnnAlgorithmDescriptor_t algoDesc, size_t *algoSpaceSizeInBytes);
+cudnnBatchNormalizationBackwardEx(cudnnHandle_t handle,
+                                  cudnnBatchNormMode_t mode,
+                                  cudnnBatchNormOps_t bnOps,
+
+                                  const void *alphaDataDiff,
+                                  const void *betaDataDiff,
+                                  const void *alphaParamDiff,
+                                  const void *betaParamDiff,
+                                  const cudnnTensorDescriptor_t xDesc,
+                                  const void *xData,
+                                  const cudnnTensorDescriptor_t yDesc,
+                                  const void *yData,
+                                  const cudnnTensorDescriptor_t dyDesc,
+                                  const void *dyData,
+                                  const cudnnTensorDescriptor_t dzDesc,
+                                  void *dzData,
+                                  const cudnnTensorDescriptor_t dxDesc,
+                                  void *dxData,
+
+                                  /* Shared tensor desc for the 4 tensors below */
+                                  const cudnnTensorDescriptor_t dBnScaleBiasDesc,
+                                  const void *bnScaleData,
+                                  const void *bnBiasData, /* needed if there is activation */
+                                  void *dBnScaleData,
+                                  void *dBnBiasData,
+                                  double epsilon, /* Same epsilon as forward pass */
+
+                                  /* Optionally cached intermediate results from
+                                     forward pass */
+                                  const void *savedMean,
+                                  const void *savedInvVariance,
+                                  cudnnActivationDescriptor_t activationDesc,
+                                  void *workSpace,
+                                  size_t workSpaceSizeInBytes,
+                                  void *reserveSpace,
+                                  size_t reserveSpaceSizeInBytes);
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnSaveAlgorithm(cudnnHandle_t handle,
-                   cudnnAlgorithmDescriptor_t algoDesc,
-                   void *algoSpace,
-                   size_t algoSpaceSizeInBytes);
+cudnnGetNormalizationForwardTrainingWorkspaceSize(cudnnHandle_t handle,
+                                                  cudnnNormMode_t mode,
+                                                  cudnnNormOps_t normOps,
+                                                  cudnnNormAlgo_t algo,
+                                                  const cudnnTensorDescriptor_t xDesc,
+                                                  const cudnnTensorDescriptor_t zDesc,
+                                                  const cudnnTensorDescriptor_t yDesc,
+                                                  const cudnnTensorDescriptor_t normScaleBiasDesc,
+                                                  const cudnnActivationDescriptor_t activationDesc,
+                                                  const cudnnTensorDescriptor_t normMeanVarDesc,
+                                                  size_t *sizeInBytes,
+                                                  int groupCnt); /* Place hold for future work, should be set to 1 now*/
 
 CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
-cudnnRestoreAlgorithm(cudnnHandle_t handle,
-                      void *algoSpace,
-                      size_t algoSpaceSizeInBytes,
-                      cudnnAlgorithmDescriptor_t algoDesc);
-
-typedef enum {
-    CUDNN_SEV_FATAL   = 0,
-    CUDNN_SEV_ERROR   = 1,
-    CUDNN_SEV_WARNING = 2,
-    CUDNN_SEV_INFO    = 3,
-} cudnnSeverity_t;
-
-/* Message masks to be used with cudnnSetCallback() */
-#define CUDNN_SEV_ERROR_EN (1U << CUDNN_SEV_ERROR)
-#define CUDNN_SEV_WARNING_EN (1U << CUDNN_SEV_WARNING)
-#define CUDNN_SEV_INFO_EN (1U << CUDNN_SEV_INFO)
-
-/* struct containing useful informaiton for each API call */
-typedef struct cudnnDebugStruct {
-    unsigned cudnn_version;
-    cudnnStatus_t cudnnStatus;
-    unsigned time_sec;      /* epoch time in seconds */
-    unsigned time_usec;     /* microseconds part of epoch time */
-    unsigned time_delta;    /* time since start in seconds */
-    cudnnHandle_t handle;   /* cudnn handle */
-    cudaStream_t stream;    /* cuda stream ID */
-    unsigned long long pid; /* process ID */
-    unsigned long long tid; /* thread ID */
-    int cudaDeviceId;       /* CUDA device ID */
-    int reserved[15];       /* reserved for future use */
-} cudnnDebug_t;
-
-typedef void (*cudnnCallback_t)(cudnnSeverity_t sev, void *udata, const cudnnDebug_t *dbg, const char *msg);
+cudnnGetNormalizationBackwardWorkspaceSize(cudnnHandle_t handle,
+                                           cudnnNormMode_t mode,
+                                           cudnnNormOps_t normOps,
+                                           cudnnNormAlgo_t algo,
+                                           const cudnnTensorDescriptor_t xDesc,
+                                           const cudnnTensorDescriptor_t yDesc,
+                                           const cudnnTensorDescriptor_t dyDesc,
+                                           const cudnnTensorDescriptor_t dzDesc,
+                                           const cudnnTensorDescriptor_t dxDesc,
+                                           const cudnnTensorDescriptor_t dNormScaleBiasDesc,
+                                           const cudnnActivationDescriptor_t activationDesc,
+                                           const cudnnTensorDescriptor_t normMeanVarDesc,
+                                           size_t *sizeInBytes,
+                                           int groupCnt); /* Place hold for future work, should be set to 1 now*/
 
-cudnnStatus_t CUDNNWINAPI
-cudnnSetCallback(unsigned mask, void *udata, cudnnCallback_t fptr);
-
-cudnnStatus_t CUDNNWINAPI
-cudnnGetCallback(unsigned *mask, void **udata, cudnnCallback_t *fptr);
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnGetNormalizationTrainingReserveSpaceSize(cudnnHandle_t handle,
+                                              cudnnNormMode_t mode,
+                                              cudnnNormOps_t normOps,
+                                              cudnnNormAlgo_t algo,
+                                              const cudnnActivationDescriptor_t activationDesc,
+                                              const cudnnTensorDescriptor_t xDesc,
+                                              size_t *sizeInBytes,
+                                              int groupCnt); /* Place hold for future work, should be set to 1 now*/
+
+/* Computes y = relu(Norm(x) + z). Also accumulates moving averages of mean and inverse variances */
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnNormalizationForwardTraining(cudnnHandle_t handle,
+                                  cudnnNormMode_t mode,
+                                  cudnnNormOps_t normOps,
+                                  cudnnNormAlgo_t algo,
+                                  const void *alpha, /* alpha[0] = result blend factor */
+                                  const void *beta,  /* beta[0] = dest layer blend factor */
+                                  const cudnnTensorDescriptor_t xDesc,
+                                  const void *xData,
+                                  const cudnnTensorDescriptor_t normScaleBiasDesc,
+                                  const void *normScale,
+                                  const void *normBias,
+                                  double exponentialAverageFactor,
+                                  const cudnnTensorDescriptor_t normMeanVarDesc,
+                                  void *resultRunningMean,
+                                  void *resultRunningVariance,
+                                  /* Has to be >= 0. Should be the same in forward and backward functions. */
+                                  double epsilon,
+                                  /* Optionally save intermediate results from the forward pass here
+                                     - can be reused to speed up backward pass. NULL if unused */
+                                  void *resultSaveMean,
+                                  void *resultSaveInvVariance,
+                                  cudnnActivationDescriptor_t activationDesc,
+                                  const cudnnTensorDescriptor_t zDesc,
+                                  const void *zData,
+                                  const cudnnTensorDescriptor_t yDesc,
+                                  void *yData,
+                                  void *workspace,
+                                  size_t workSpaceSizeInBytes,
+                                  void *reserveSpace,
+                                  size_t reserveSpaceSizeInBytes,
+                                  int groupCnt); /* Place hold for future work, should be set to 1 now*/
 
-/*
- * \brief Cross-library version checker.
- * 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.
- */
-cudnnStatus_t CUDNNWINAPI
-cudnnOpsInferVersionCheck(void);
+CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI
+cudnnNormalizationBackward(cudnnHandle_t handle,
+                           cudnnNormMode_t mode,
+                           cudnnNormOps_t normOps,
+                           cudnnNormAlgo_t algo,
+                           const void *alphaDataDiff,
+                           const void *betaDataDiff,
+                           const void *alphaParamDiff,
+                           const void *betaParamDiff,
+                           const cudnnTensorDescriptor_t xDesc,
+                           const void *xData,
+                           const cudnnTensorDescriptor_t yDesc,
+                           const void *yData,
+                           const cudnnTensorDescriptor_t dyDesc,
+                           const void *dyData,
+                           const cudnnTensorDescriptor_t dzDesc,
+                           void *dzData,
+                           const cudnnTensorDescriptor_t dxDesc,
+                           void *dxData,
+                           /* Shared tensor desc for the 4 tensors below */
+                           const cudnnTensorDescriptor_t dNormScaleBiasDesc,
+                           const void *normScaleData,
+                           const void *normBiasData, /* needed if there is activation */
+                           void *dNormScaleData,
+                           void *dNormBiasData,
+                           double epsilon, /* Same epsilon as forward pass */
+                           const cudnnTensorDescriptor_t normMeanVarDesc,
+                           /* Optionally cached intermediate results from
+                              forward pass */
+                           const void *savedMean,
+                           const void *savedInvVariance,
+                           cudnnActivationDescriptor_t activationDesc,
+                           void *workSpace,
+                           size_t workSpaceSizeInBytes,
+                           void *reserveSpace,
+                           size_t reserveSpaceSizeInBytes,
+                           int groupCnt); /* Place hold for future work, should be set to 1 now*/
+
+cudnnStatus_t CUDNNWINAPI
+cudnnSpatialTfGridGeneratorBackward(cudnnHandle_t handle,
+                                    const cudnnSpatialTransformerDescriptor_t stDesc,
+                                    const void *dgrid,
+                                    void *dtheta);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnSpatialTfSamplerBackward(cudnnHandle_t handle,
+                              cudnnSpatialTransformerDescriptor_t stDesc,
+                              const void *alpha,
+                              const cudnnTensorDescriptor_t xDesc,
+                              const void *x,
+                              const void *beta,
+                              const cudnnTensorDescriptor_t dxDesc,
+                              void *dx,
+                              const void *alphaDgrid,
+                              const cudnnTensorDescriptor_t dyDesc,
+                              const void *dy,
+                              const void *grid,
+                              const void *betaDgrid,
+                              void *dgrid);
+
+cudnnStatus_t CUDNNWINAPI
+cudnnDropoutBackward(cudnnHandle_t handle,
+                     const cudnnDropoutDescriptor_t dropoutDesc,
+                     const cudnnTensorDescriptor_t dydesc,
+                     const void *dy,
+                     const cudnnTensorDescriptor_t dxdesc,
+                     void *dx,
+                     void *reserveSpace,
+                     size_t reserveSpaceSizeInBytes);
 
 #if defined(__cplusplus)
 }
 #endif
 
-#endif /* CUDNN_OPS_INFER_H_ */
+#endif /* CUDNN_OPS_H_ */