/* * Copyright 2014-2023 NVIDIA Corporation. All rights reserved. * * NOTICE TO LICENSEE: * * This source code and/or documentation ("Licensed Deliverables") are * subject to NVIDIA intellectual property rights under U.S. and * international Copyright laws. * * These Licensed Deliverables contained herein is PROPRIETARY and * CONFIDENTIAL to NVIDIA and is being provided under the terms and * conditions of a form of NVIDIA software license agreement by and * between NVIDIA and Licensee ("License Agreement") or electronically * accepted by Licensee. Notwithstanding any terms or conditions to * the contrary in the License Agreement, reproduction or disclosure * of the Licensed Deliverables to any third party without the express * written consent of NVIDIA is prohibited. * * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE. IT IS * PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND. * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY, * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE. * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE * OF THESE LICENSED DELIVERABLES. * * U.S. Government End Users. These Licensed Deliverables are a * "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT * 1995), consisting of "commercial computer software" and "commercial * computer software documentation" as such terms are used in 48 * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government * only as a commercial end item. Consistent with 48 C.F.R.12.212 and * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all * U.S. Government End Users acquire the Licensed Deliverables with * only those rights set forth herein. * * Any use of the Licensed Deliverables in individual and commercial * software must include, in the user documentation and internal * comments to the code, the above Disclaimer and U.S. Government End * Users Notice. */ /** * @file cudnn_cnn.h * @brief cuDNN CNN library - convolution, algorithm selection, and fused operations. Most APIs deprecated. * * @since cuDNN 9.0.0 */ /* * cudnn_cnn : cuDNN's basic definitions and CNN functions. */ #if !defined(CUDNN_CNN_H_) #define CUDNN_CNN_H_ #pragma once #include #include "cudnn_version.h" #include "cudnn_ops.h" /* These version numbers are autogenerated, do not edit manually. */ #define CUDNN_CNN_MAJOR 9 #define CUDNN_CNN_MINOR 24 #define CUDNN_CNN_PATCH 0 #if (CUDNN_CNN_MAJOR != CUDNN_MAJOR) || (CUDNN_CNN_MINOR != CUDNN_MINOR) || (CUDNN_CNN_PATCH != CUDNN_PATCHLEVEL) #error Version mismatch in cuDNN CNN INFER!!! #endif #if defined(__cplusplus) extern "C" { #endif /** * @brief Opaque descriptor for a convolution operation. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnConvolutionStruct *cudnnConvolutionDescriptor_t CUDNN_DEPRECATED; /** * @brief Performance results for forward convolution algorithm selection. * * Contains timing, memory usage, and determinism information for a given * forward convolution algorithm. Returned by algorithm search functions. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnConvolutionFwdAlgoPerfStruct { cudnnConvolutionFwdAlgo_t algo; /**< The forward convolution algorithm. */ cudnnStatus_t status; /**< Status returned when running this algorithm. */ float time; /**< Execution time in milliseconds. */ size_t memory; /**< Workspace memory required in bytes. */ cudnnDeterminism_t determinism; /**< Whether the algorithm is deterministic. */ cudnnMathType_t mathType; /**< Math type used by the algorithm. */ int reserved[3]; /**< Reserved for future use. */ } cudnnConvolutionFwdAlgoPerf_t CUDNN_DEPRECATED; /** * @brief Create an instance of convolution descriptor. * * @param[out] convDesc Pointer to receive the newly created convolution descriptor. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Create an instance of convolution descriptor */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateConvolutionDescriptor(cudnnConvolutionDescriptor_t *convDesc); /** * @brief Destroy an instance of convolution descriptor. * * @param[in] convDesc The convolution descriptor to destroy. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Destroy an instance of convolution descriptor */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroyConvolutionDescriptor(cudnnConvolutionDescriptor_t convDesc); /** * @brief Set the math type for a convolution descriptor. * * @param[in,out] convDesc The convolution descriptor. * @param[in] mathType The math type to set. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetConvolutionMathType(cudnnConvolutionDescriptor_t convDesc, cudnnMathType_t mathType); /** * @brief Get the math type from a convolution descriptor. * * @param[in] convDesc The convolution descriptor. * @param[out] mathType Pointer to receive the math type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionMathType(cudnnConvolutionDescriptor_t convDesc, cudnnMathType_t *mathType); /** * @brief Set the group count for a convolution descriptor. * * @param[in,out] convDesc The convolution descriptor. * @param[in] groupCount The number of groups for grouped convolution. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetConvolutionGroupCount(cudnnConvolutionDescriptor_t convDesc, int groupCount); /** * @brief Get the group count from a convolution descriptor. * * @param[in] convDesc The convolution descriptor. * @param[out] groupCount Pointer to receive the group count. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionGroupCount(cudnnConvolutionDescriptor_t convDesc, int *groupCount); /** * @brief Set the reorder type for a convolution descriptor. * * @param[in,out] convDesc The convolution descriptor. * @param[in] reorderType The reorder type to set. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetConvolutionReorderType(cudnnConvolutionDescriptor_t convDesc, cudnnReorderType_t reorderType); /** * @brief Get the reorder type from a convolution descriptor. * * @param[in] convDesc The convolution descriptor. * @param[out] reorderType Pointer to receive the reorder type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionReorderType(cudnnConvolutionDescriptor_t convDesc, cudnnReorderType_t *reorderType); /** * @brief Set a 2D convolution descriptor with padding, stride, dilation, mode, and compute type. * * @param[in,out] convDesc The convolution descriptor to initialize. * @param[in] pad_h Zero-padding height. * @param[in] pad_w Zero-padding width. * @param[in] u Vertical filter stride. * @param[in] v Horizontal filter stride. * @param[in] dilation_h Filter dilation in the vertical dimension. * @param[in] dilation_w Filter dilation in the horizontal dimension. * @param[in] mode Convolution mode (cross-correlation or convolution). * @param[in] computeType Data type for convolution computation. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetConvolution2dDescriptor(cudnnConvolutionDescriptor_t convDesc, int pad_h, /* zero-padding height */ int pad_w, /* zero-padding width */ int u, /* vertical filter stride */ int v, /* horizontal filter stride */ int dilation_h, /* filter dilation in the vertical dimension */ int dilation_w, /* filter dilation in the horizontal dimension */ cudnnConvolutionMode_t mode, cudnnDataType_t computeType); /** * @brief Get the parameters of a 2D convolution descriptor. * * @param[in] convDesc The convolution descriptor to query. * @param[out] pad_h Pointer to receive zero-padding height. * @param[out] pad_w Pointer to receive zero-padding width. * @param[out] u Pointer to receive vertical filter stride. * @param[out] v Pointer to receive horizontal filter stride. * @param[out] dilation_h Pointer to receive filter dilation in the vertical dimension. * @param[out] dilation_w Pointer to receive filter dilation in the horizontal dimension. * @param[out] mode Pointer to receive convolution mode. * @param[out] computeType Pointer to receive compute data type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolution2dDescriptor(const cudnnConvolutionDescriptor_t convDesc, int *pad_h, /* zero-padding height */ int *pad_w, /* zero-padding width */ int *u, /* vertical filter stride */ int *v, /* horizontal filter stride */ int *dilation_h, /* filter dilation in the vertical dimension */ int *dilation_w, /* filter dilation in the horizontal dimension */ cudnnConvolutionMode_t *mode, cudnnDataType_t *computeType); /** * @brief Set an N-dimensional convolution descriptor. * * @param[in,out] convDesc The convolution descriptor to initialize. * @param[in] arrayLength Number of dimensions (nbDims-2 size). * @param[in] padA Array of zero-padding values per dimension. * @param[in] filterStrideA Array of filter strides per dimension. * @param[in] dilationA Array of dilation values per dimension. * @param[in] mode Convolution mode. * @param[in] computeType Data type for convolution computation. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetConvolutionNdDescriptor(cudnnConvolutionDescriptor_t convDesc, int arrayLength, /* nbDims-2 size */ const int padA[], const int filterStrideA[], const int dilationA[], cudnnConvolutionMode_t mode, cudnnDataType_t computeType); /* convolution data type */ /** * @brief Get the parameters of an N-dimensional convolution descriptor. * * @param[in] convDesc The convolution descriptor to query. * @param[in] arrayLengthRequested Maximum number of dimensions to retrieve. * @param[out] arrayLength Pointer to receive the actual number of dimensions. * @param[out] padA Array to receive zero-padding values. * @param[out] strideA Array to receive stride values. * @param[out] dilationA Array to receive dilation values. * @param[out] mode Pointer to receive convolution mode. * @param[out] computeType Pointer to receive compute data type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to return the dimensions of the output tensor given a convolution descriptor */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionNdDescriptor(const cudnnConvolutionDescriptor_t convDesc, int arrayLengthRequested, int *arrayLength, int padA[], int strideA[], int dilationA[], cudnnConvolutionMode_t *mode, cudnnDataType_t *computeType); /* convolution data type */ /** * @brief Compute the output dimensions of a 2D convolution. * * @param[in] convDesc The convolution descriptor. * @param[in] inputTensorDesc Descriptor for the input tensor. * @param[in] filterDesc Descriptor for the filter. * @param[out] n Pointer to receive the output batch size. * @param[out] c Pointer to receive the output channels. * @param[out] h Pointer to receive the output height. * @param[out] w Pointer to receive the output width. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolution2dForwardOutputDim(const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t inputTensorDesc, const cudnnFilterDescriptor_t filterDesc, int *n, int *c, int *h, int *w); /** * @brief Compute the output dimensions of an N-dimensional convolution. * * @param[in] convDesc The convolution descriptor. * @param[in] inputTensorDesc Descriptor for the input tensor. * @param[in] filterDesc Descriptor for the filter. * @param[in] nbDims Number of dimensions. * @param[out] tensorOuputDimA Array to receive the output dimensions. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to return the dimensions of the output tensor given a convolution descriptor */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionNdForwardOutputDim(const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t inputTensorDesc, const cudnnFilterDescriptor_t filterDesc, int nbDims, int tensorOuputDimA[]); /** * @brief Get the maximum number of forward convolution algorithms available. * * @param[in] handle The cuDNN handle. * @param[out] count Pointer to receive the maximum algorithm count. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* helper function to provide the convolution forward algo that fit best the requirement */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionForwardAlgorithmMaxCount(cudnnHandle_t handle, int *count); /** * @brief Get forward convolution algorithm recommendations without executing them. * * Returns a list of algorithms sorted by expected performance. Does not * require a workspace or run actual convolutions. * * @param[in] handle The cuDNN handle. * @param[in] srcDesc Descriptor for the input tensor. * @param[in] filterDesc Descriptor for the filter. * @param[in] convDesc The convolution descriptor. * @param[in] destDesc Descriptor for the output tensor. * @param[in] requestedAlgoCount Maximum number of algorithms to return. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionForwardAlgorithm_v7(cudnnHandle_t handle, const cudnnTensorDescriptor_t srcDesc, const cudnnFilterDescriptor_t filterDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t destDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionFwdAlgoPerf_t *perfResults); /** * @brief Find the best forward convolution algorithm by running benchmarks. * * Executes all applicable algorithms and returns performance results * sorted by execution time. Does not require user-allocated workspace. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] wDesc Descriptor for the filter. * @param[in] convDesc The convolution descriptor. * @param[in] yDesc Descriptor for the output tensor. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionForwardAlgorithm(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const cudnnFilterDescriptor_t wDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t yDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionFwdAlgoPerf_t *perfResults); /** * @brief Find the best forward convolution algorithm by running benchmarks with user-provided buffers. * * Similar to cudnnFindConvolutionForwardAlgorithm but uses caller-provided * data buffers and workspace. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] wDesc Descriptor for the filter. * @param[in] w Pointer to filter data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] yDesc Descriptor for the output tensor. * @param[out] y Pointer to output data in device memory. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionForwardAlgorithmEx(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnFilterDescriptor_t wDesc, const void *w, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t yDesc, void *y, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionFwdAlgoPerf_t *perfResults, void *workSpace, size_t workSpaceSizeInBytes); /** * @brief Perform the Im2Col transform for convolution. * * Rearranges image data into a column matrix suitable for matrix multiplication * based convolution. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] wDesc Descriptor for the filter. * @param[in] convDesc The convolution descriptor. * @param[out] colBuffer Pointer to the output column buffer in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnIm2Col(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnFilterDescriptor_t wDesc, const cudnnConvolutionDescriptor_t convDesc, void *colBuffer); /** * @brief Reorder filter and bias data for optimized convolution execution. * * Rearranges filter and optionally bias data into a layout optimized for * the specified reorder type. * * @param[in] handle The cuDNN handle. * @param[in] filterDesc Descriptor for the filter. * @param[in] reorderType The reorder type to apply. * @param[in] filterData Pointer to source filter data in device memory. * @param[out] reorderedFilterData Pointer to destination filter data in device memory. * @param[in] reorderBias Non-zero to also reorder bias data. * @param[in] biasData Pointer to source bias data in device memory (may be NULL). * @param[out] reorderedBiasData Pointer to destination bias data in device memory (may be NULL). * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnReorderFilterAndBias(cudnnHandle_t handle, const cudnnFilterDescriptor_t filterDesc, cudnnReorderType_t reorderType, const void *filterData, void *reorderedFilterData, int reorderBias, const void *biasData, void *reorderedBiasData); /** * @brief Get the minimum workspace size required for a forward convolution algorithm. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] wDesc Descriptor for the filter. * @param[in] convDesc The convolution descriptor. * @param[in] yDesc Descriptor for the output tensor. * @param[in] algo The forward convolution algorithm. * @param[out] sizeInBytes Pointer to receive the required workspace size. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to return the minimum size of the workspace to be passed to the convolution given an algo*/ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionForwardWorkspaceSize(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const cudnnFilterDescriptor_t wDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t yDesc, cudnnConvolutionFwdAlgo_t algo, size_t *sizeInBytes); /* Convolution functions: All of the form "output = alpha * Op(inputs) + beta * output" */ /** * @brief Execute a forward convolution. * * Computes: y = alpha * conv(x, w) + beta * y * * @param[in] handle The cuDNN handle. * @param[in] alpha Pointer to scaling factor for the convolution result. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] wDesc Descriptor for the filter. * @param[in] w Pointer to filter data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] algo The forward convolution algorithm to use. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @param[in] beta Pointer to scaling factor for the prior output. * @param[in] yDesc Descriptor for the output tensor. * @param[in,out] y Pointer to output data in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Function to perform the forward pass for batch convolution */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnConvolutionForward(cudnnHandle_t handle, const void *alpha, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnFilterDescriptor_t wDesc, const void *w, const cudnnConvolutionDescriptor_t convDesc, cudnnConvolutionFwdAlgo_t algo, void *workSpace, size_t workSpaceSizeInBytes, const void *beta, const cudnnTensorDescriptor_t yDesc, void *y); /** * @brief Execute a fused convolution with bias and activation. * * Computes: y = Act( alpha1 * conv(x) + alpha2 * z + bias ) * * @param[in] handle The cuDNN handle. * @param[in] alpha1 Pointer to scaling factor for the convolution result. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] wDesc Descriptor for the filter. * @param[in] w Pointer to filter data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] algo The forward convolution algorithm to use. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @param[in] alpha2 Pointer to scaling factor for the residual input z. * @param[in] zDesc Descriptor for the residual input tensor. * @param[in] z Pointer to residual data in device memory. * @param[in] biasDesc Descriptor for the bias tensor. * @param[in] bias Pointer to bias data in device memory. * @param[in] activationDesc Descriptor for the activation operation. * @param[in] yDesc Descriptor for the output tensor. * @param[in,out] y Pointer to output data in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Fused conv/bias/activation operation : y = Act( alpha1 * conv(x) + alpha2 * z + bias ) */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnConvolutionBiasActivationForward(cudnnHandle_t handle, const void *alpha1, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnFilterDescriptor_t wDesc, const void *w, const cudnnConvolutionDescriptor_t convDesc, cudnnConvolutionFwdAlgo_t algo, void *workSpace, size_t workSpaceSizeInBytes, const void *alpha2, const cudnnTensorDescriptor_t zDesc, const void *z, const cudnnTensorDescriptor_t biasDesc, const void *bias, const cudnnActivationDescriptor_t activationDesc, const cudnnTensorDescriptor_t yDesc, void *y); /* helper function to provide the convolution backward data algo that fit best the requirement */ /** * @brief Performance results for backward data convolution algorithm selection. * * Contains timing, memory usage, and determinism information for a given * backward data convolution algorithm. Returned by algorithm search functions. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnConvolutionBwdDataAlgoPerfStruct { cudnnConvolutionBwdDataAlgo_t algo; /**< The backward data convolution algorithm. */ cudnnStatus_t status; /**< Status returned when running this algorithm. */ float time; /**< Execution time in milliseconds. */ size_t memory; /**< Workspace memory required in bytes. */ cudnnDeterminism_t determinism; /**< Whether the algorithm is deterministic. */ cudnnMathType_t mathType; /**< Math type used by the algorithm. */ int reserved[3]; /**< Reserved for future use. */ } cudnnConvolutionBwdDataAlgoPerf_t CUDNN_DEPRECATED; /** * @brief Get the maximum number of backward data convolution algorithms available. * * @param[in] handle The cuDNN handle. * @param[out] count Pointer to receive the maximum algorithm count. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardDataAlgorithmMaxCount(cudnnHandle_t handle, int *count); /** * @brief Find the best backward data convolution algorithm by running benchmarks. * * @param[in] handle The cuDNN handle. * @param[in] wDesc Descriptor for the filter. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] dxDesc Descriptor for the gradient input tensor. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionBackwardDataAlgorithm(cudnnHandle_t handle, const cudnnFilterDescriptor_t wDesc, const cudnnTensorDescriptor_t dyDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t dxDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdDataAlgoPerf_t *perfResults); /** * @brief Find the best backward data convolution algorithm with user-provided buffers. * * @param[in] handle The cuDNN handle. * @param[in] wDesc Descriptor for the filter. * @param[in] w Pointer to filter data in device memory. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] dy Pointer to gradient output data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] dxDesc Descriptor for the gradient input tensor. * @param[out] dx Pointer to gradient input data in device memory. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionBackwardDataAlgorithmEx(cudnnHandle_t handle, const cudnnFilterDescriptor_t wDesc, const void *w, const cudnnTensorDescriptor_t dyDesc, const void *dy, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t dxDesc, void *dx, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdDataAlgoPerf_t *perfResults, void *workSpace, size_t workSpaceSizeInBytes); /** * @brief Get backward data convolution algorithm recommendations without executing them. * * @param[in] handle The cuDNN handle. * @param[in] filterDesc Descriptor for the filter. * @param[in] diffDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] gradDesc Descriptor for the gradient input tensor. * @param[in] requestedAlgoCount Maximum number of algorithms to return. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardDataAlgorithm_v7(cudnnHandle_t handle, const cudnnFilterDescriptor_t filterDesc, const cudnnTensorDescriptor_t diffDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t gradDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdDataAlgoPerf_t *perfResults); /* * convolution algorithm (which requires potentially some workspace) */ /** * @brief Get the minimum workspace size required for a backward data convolution algorithm. * * @param[in] handle The cuDNN handle. * @param[in] wDesc Descriptor for the filter. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] dxDesc Descriptor for the gradient input tensor. * @param[in] algo The backward data convolution algorithm. * @param[out] sizeInBytes Pointer to receive the required workspace size. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to return the minimum size of the workspace to be passed to the convolution given an algo*/ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardDataWorkspaceSize(cudnnHandle_t handle, const cudnnFilterDescriptor_t wDesc, const cudnnTensorDescriptor_t dyDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t dxDesc, cudnnConvolutionBwdDataAlgo_t algo, size_t *sizeInBytes); /** * @brief Execute a backward data convolution (compute gradient with respect to input data). * * Computes: dx = alpha * dconv(w, dy) + beta * dx * * @param[in] handle The cuDNN handle. * @param[in] alpha Pointer to scaling factor for the convolution result. * @param[in] wDesc Descriptor for the filter. * @param[in] w Pointer to filter data in device memory. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] dy Pointer to gradient output data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] algo The backward data convolution algorithm. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @param[in] beta Pointer to scaling factor for the prior gradient input. * @param[in] dxDesc Descriptor for the gradient input tensor. * @param[in,out] dx Pointer to gradient input data in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnConvolutionBackwardData(cudnnHandle_t handle, const void *alpha, const cudnnFilterDescriptor_t wDesc, const void *w, const cudnnTensorDescriptor_t dyDesc, const void *dy, const cudnnConvolutionDescriptor_t convDesc, cudnnConvolutionBwdDataAlgo_t algo, void *workSpace, size_t workSpaceSizeInBytes, const void *beta, const cudnnTensorDescriptor_t dxDesc, void *dx); /** * @brief Calculate folding descriptors for backward data convolution (dgrad). * * Computes the folded descriptors needed for tensor transform operations * used in backward data gradient computation. * * @param[in] handle The cuDNN handle. * @param[in] filterDesc Descriptor for the filter. * @param[in] diffDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] gradDesc Descriptor for the gradient input tensor. * @param[in] transformFormat The tensor format for the transform. * @param[out] foldedFilterDesc Descriptor for the folded filter. * @param[out] paddedDiffDesc Descriptor for the padded gradient output. * @param[out] foldedConvDesc Descriptor for the folded convolution. * @param[out] foldedGradDesc Descriptor for the folded gradient input. * @param[out] filterFoldTransDesc Transform descriptor for filter folding. * @param[out] diffPadTransDesc Transform descriptor for diff padding. * @param[out] gradFoldTransDesc Transform descriptor for gradient folding. * @param[out] gradUnfoldTransDesc Transform descriptor for gradient unfolding. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to calculate folding descriptors for dgrad */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetFoldedConvBackwardDataDescriptors(const cudnnHandle_t handle, const cudnnFilterDescriptor_t filterDesc, const cudnnTensorDescriptor_t diffDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnTensorDescriptor_t gradDesc, const cudnnTensorFormat_t transformFormat, cudnnFilterDescriptor_t foldedFilterDesc, cudnnTensorDescriptor_t paddedDiffDesc, cudnnConvolutionDescriptor_t foldedConvDesc, cudnnTensorDescriptor_t foldedGradDesc, cudnnTensorTransformDescriptor_t filterFoldTransDesc, cudnnTensorTransformDescriptor_t diffPadTransDesc, cudnnTensorTransformDescriptor_t gradFoldTransDesc, cudnnTensorTransformDescriptor_t gradUnfoldTransDesc); /* cudnnFusedOps... */ /** * @brief Opaque descriptor for fused operations constant parameter pack. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ struct cudnnFusedOpsConstParamStruct; /** * @brief Handle to a fused operations constant parameter pack. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnFusedOpsConstParamStruct *cudnnFusedOpsConstParamPack_t CUDNN_DEPRECATED; /** * @brief Opaque descriptor for fused operations variant parameter pack. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ struct cudnnFusedOpsVariantParamStruct; /** * @brief Handle to a fused operations variant parameter pack. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnFusedOpsVariantParamStruct *cudnnFusedOpsVariantParamPack_t CUDNN_DEPRECATED; /** * @brief Opaque descriptor for a fused operations execution plan. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ struct cudnnFusedOpsPlanStruct; /** * @brief Handle to a fused operations execution plan. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnFusedOpsPlanStruct *cudnnFusedOpsPlan_t CUDNN_DEPRECATED; /** * @brief Enumeration of fused operation sequences. * * Specifies which sequence of operations to fuse together for optimized execution. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef enum { /* each op in [ ] can be disabled by passing NULL ptr */ /* [per channel scale], [per channel bias], [activation], convolution, [generate BN stats] */ CUDNN_FUSED_SCALE_BIAS_ACTIVATION_CONV_BNSTATS = 0, /**< Fused scale, bias, activation, convolution, and optional BN stats generation. */ /* [per channel scale], [per channel bias], [activation], convolutionBackwardWeights */ CUDNN_FUSED_SCALE_BIAS_ACTIVATION_WGRAD = 1, /**< Fused scale, bias, activation, and backward weight convolution. */ /* utility for BN training in BN-conv fusion */ /* computes the equivalent scale and bias from ySum ySqSum and learned scale, bias */ /* optionally update running stats and generate saved stats */ CUDNN_FUSED_BN_FINALIZE_STATISTICS_TRAINING = 2, /**< Finalize batch normalization statistics for training. */ /* utility for BN inference in BN-conv fusion */ /* computes the equivalent scale and bias from learned running stats and learned scale, bias */ CUDNN_FUSED_BN_FINALIZE_STATISTICS_INFERENCE = 3, /**< Finalize batch normalization statistics for inference. */ /* reserved for future use: convolution, [per channel scale], [per channel bias], [residual add], [activation] */ CUDNN_FUSED_CONV_SCALE_BIAS_ADD_ACTIVATION = 4, /**< Fused convolution, scale, bias, residual add, and activation. */ /* reserved for future use: [per channel scale], [per channel bias], [residual add], activation, bitmask */ CUDNN_FUSED_SCALE_BIAS_ADD_ACTIVATION_GEN_BITMASK = 5, /**< Fused scale, bias, residual add, activation, and bitmask generation. */ /* reserved for future use */ CUDNN_FUSED_DACTIVATION_FORK_DBATCHNORM = 6, /**< Fused backward activation fork and backward batch normalization. */ } cudnnFusedOps_t CUDNN_DEPRECATED; /** * @brief Labels for constant parameters in a fused operations parameter pack. * * Used to identify which constant parameter to set or get in a * cudnnFusedOpsConstParamPack_t. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef enum { /* set XDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get XDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_XDESC = 0, /**< Input tensor descriptor (X). */ /* set/get XDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_XDATA_PLACEHOLDER = 1, /**< Placeholder for input data pointer alignment (X). */ /* set/get BN_MODE: pass cudnnBatchNormMode_t* */ CUDNN_PARAM_BN_MODE = 2, /**< Batch normalization mode. */ /* set CUDNN_PARAM_BN_EQSCALEBIAS_DESC: pass previously initialized cudnnTensorDescriptor_t */ /* get CUDNN_PARAM_BN_EQSCALEBIAS_DESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_BN_EQSCALEBIAS_DESC = 3, /**< Equivalent scale/bias descriptor for BN fusion. */ /* set/get BN_EQSCALE_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_EQSCALE_PLACEHOLDER = 4, /**< Placeholder for BN equivalent scale pointer alignment. */ /* set/get BN_EQBIAS_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_EQBIAS_PLACEHOLDER = 5, /**< Placeholder for BN equivalent bias pointer alignment. */ /* set ACTIVATION_DESC: pass previously initialized cudnnActivationDescriptor_t */ /* get ACTIVATION_DESC: pass previously created cudnnActivationDescriptor_t */ CUDNN_PARAM_ACTIVATION_DESC = 6, /**< Activation descriptor. */ /* set CONV_DESC: pass previously initialized cudnnConvolutionDescriptor_t */ /* get CONV_DESC: pass previously created cudnnConvolutionDescriptor_t */ CUDNN_PARAM_CONV_DESC = 7, /**< Convolution descriptor. */ /* set WDESC: pass previously initialized cudnnFilterDescriptor_t */ /* get WDESC: pass previously created cudnnFilterDescriptor_t */ CUDNN_PARAM_WDESC = 8, /**< Filter descriptor (W). */ /* set/get WDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_WDATA_PLACEHOLDER = 9, /**< Placeholder for filter data pointer alignment (W). */ /* set DWDESC: pass previously initialized cudnnFilterDescriptor_t */ /* get DWDESC: pass previously created cudnnFilterDescriptor_t */ CUDNN_PARAM_DWDESC = 10, /**< Filter gradient descriptor (dW). */ /* set/get DWDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_DWDATA_PLACEHOLDER = 11, /**< Placeholder for filter gradient data pointer alignment (dW). */ /* set YDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get YDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_YDESC = 12, /**< Output tensor descriptor (Y). */ /* set/get YDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_YDATA_PLACEHOLDER = 13, /**< Placeholder for output data pointer alignment (Y). */ /* set DYDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get DYDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_DYDESC = 14, /**< Output gradient tensor descriptor (dY). */ /* set/get DYDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_DYDATA_PLACEHOLDER = 15, /**< Placeholder for output gradient data pointer alignment (dY). */ /* set YSTATS_DESC: pass previously initialized cudnnTensorDescriptor_t */ /* get YSTATS_DESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_YSTATS_DESC = 16, /**< Output statistics tensor descriptor. */ /* set/get YSUM_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_YSUM_PLACEHOLDER = 17, /**< Placeholder for Y sum pointer alignment. */ /* set/get YSQSUM_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_YSQSUM_PLACEHOLDER = 18, /**< Placeholder for Y squared sum pointer alignment. */ /* set CUDNN_PARAM_BN_SCALEBIAS_MEANVAR_DESC: pass previously initialized cudnnTensorDescriptor_t */ /* get CUDNN_PARAM_BN_SCALEBIAS_MEANVAR_DESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_BN_SCALEBIAS_MEANVAR_DESC = 19, /**< BN scale/bias/mean/variance tensor descriptor. */ /* set/get CUDNN_PARAM_BN_SCALE_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_SCALE_PLACEHOLDER = 20, /**< Placeholder for BN scale pointer alignment. */ /* set/get CUDNN_PARAM_BN_BIAS_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_BIAS_PLACEHOLDER = 21, /**< Placeholder for BN bias pointer alignment. */ /* set/get CUDNN_PARAM_BN_SAVED_MEAN_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_SAVED_MEAN_PLACEHOLDER = 22, /**< Placeholder for BN saved mean pointer alignment. */ /* set/get CUDNN_PARAM_BN_SAVED_INVSTD_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_SAVED_INVSTD_PLACEHOLDER = 23, /**< Placeholder for BN saved inverse standard deviation pointer alignment. */ /* set/get CUDNN_PARAM_BN_RUNNING_MEAN_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_RUNNING_MEAN_PLACEHOLDER = 24, /**< Placeholder for BN running mean pointer alignment. */ /* set/get CUDNN_PARAM_BN_RUNNING_VAR_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_RUNNING_VAR_PLACEHOLDER = 25, /**< Placeholder for BN running variance pointer alignment. */ /* set ZDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get ZDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_ZDESC = 26, /**< Residual input tensor descriptor (Z). */ /* set/get ZDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_ZDATA_PLACEHOLDER = 27, /**< Placeholder for residual input data pointer alignment (Z). */ /* set BN_Z_EQSCALEBIAS_DESC: pass previously initialized cudnnTensorDescriptor_t */ /* get BN_Z_EQSCALEBIAS_DESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_BN_Z_EQSCALEBIAS_DESC = 28, /**< BN equivalent scale/bias descriptor for Z branch. */ /* set/get BN_Z_EQSCALE_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_Z_EQSCALE_PLACEHOLDER = 29, /**< Placeholder for BN Z-branch equivalent scale pointer alignment. */ /* set/get BN_Z_EQBIAS_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_Z_EQBIAS_PLACEHOLDER = 30, /**< Placeholder for BN Z-branch equivalent bias pointer alignment. */ /* set ACTIVATION_BITMASK_DESC: pass previously initialized cudnnTensorDescriptor_t */ /* get ACTIVATION_BITMASK_DESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_ACTIVATION_BITMASK_DESC = 31, /**< Activation bitmask tensor descriptor. */ /* set/get ACTIVATION_BITMASK_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_ACTIVATION_BITMASK_PLACEHOLDER = 32, /**< Placeholder for activation bitmask pointer alignment. */ /* set DXDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get DXDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_DXDESC = 33, /**< Input gradient tensor descriptor (dX). */ /* set/get DXDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_DXDATA_PLACEHOLDER = 34, /**< Placeholder for input gradient data pointer alignment (dX). */ /* set DZDESC: pass previously initialized cudnnTensorDescriptor_t */ /* get DZDESC: pass previously created cudnnTensorDescriptor_t */ CUDNN_PARAM_DZDESC = 35, /**< Residual input gradient tensor descriptor (dZ). */ /* set/get DZDATA_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_DZDATA_PLACEHOLDER = 36, /**< Placeholder for residual input gradient data pointer alignment (dZ). */ /* set/get CUDNN_PARAM_BN_DSCALE_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_DSCALE_PLACEHOLDER = 37, /**< Placeholder for BN scale gradient pointer alignment. */ /* set/get CUDNN_PARAM_BN_DBIAS_PLACEHOLDER: pass cudnnFusedOpsPointerPlaceHolder_t* */ CUDNN_PARAM_BN_DBIAS_PLACEHOLDER = 38, /**< Placeholder for BN bias gradient pointer alignment. */ } cudnnFusedOpsConstParamLabel_t CUDNN_DEPRECATED; /** * @brief Pointer alignment modes for fused operations. * * Specifies the alignment guarantee of pointers passed to fused operations, * allowing the runtime to select optimized code paths. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_PTR_NULL = 0, /**< Null pointer (parameter disabled). */ CUDNN_PTR_ELEM_ALIGNED = 1, /**< Pointer is element-aligned. */ CUDNN_PTR_16B_ALIGNED = 2, /**< Pointer is 16-byte aligned. */ } cudnnFusedOpsPointerPlaceHolder_t CUDNN_DEPRECATED; /** * @brief Labels for variant (per-execution) parameters in a fused operations parameter pack. * * Used to identify which variant parameter to set or get in a * cudnnFusedOpsVariantParamPack_t. These include device memory pointers * and scalar host-side values. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef enum { /* set: pass void* pointing to dev memory */ /* get: pass void** pointing to host memory */ CUDNN_PTR_XDATA = 0, /**< Pointer to input data (X) in device memory. */ CUDNN_PTR_BN_EQSCALE = 1, /**< Pointer to BN equivalent scale in device memory. */ CUDNN_PTR_BN_EQBIAS = 2, /**< Pointer to BN equivalent bias in device memory. */ CUDNN_PTR_WDATA = 3, /**< Pointer to filter data (W) in device memory. */ CUDNN_PTR_DWDATA = 4, /**< Pointer to filter gradient data (dW) in device memory. */ CUDNN_PTR_YDATA = 5, /**< Pointer to output data (Y) in device memory. */ CUDNN_PTR_DYDATA = 6, /**< Pointer to output gradient data (dY) in device memory. */ CUDNN_PTR_YSUM = 7, /**< Pointer to Y sum accumulator in device memory. */ CUDNN_PTR_YSQSUM = 8, /**< Pointer to Y squared sum accumulator in device memory. */ CUDNN_PTR_WORKSPACE = 9, /**< Pointer to workspace in device memory. */ CUDNN_PTR_BN_SCALE = 10, /**< Pointer to BN scale in device memory. */ CUDNN_PTR_BN_BIAS = 11, /**< Pointer to BN bias in device memory. */ CUDNN_PTR_BN_SAVED_MEAN = 12, /**< Pointer to BN saved mean in device memory. */ CUDNN_PTR_BN_SAVED_INVSTD = 13, /**< Pointer to BN saved inverse standard deviation in device memory. */ CUDNN_PTR_BN_RUNNING_MEAN = 14, /**< Pointer to BN running mean in device memory. */ CUDNN_PTR_BN_RUNNING_VAR = 15, /**< Pointer to BN running variance in device memory. */ CUDNN_PTR_ZDATA = 16, /**< Pointer to residual input data (Z) in device memory. */ CUDNN_PTR_BN_Z_EQSCALE = 17, /**< Pointer to BN Z-branch equivalent scale in device memory. */ CUDNN_PTR_BN_Z_EQBIAS = 18, /**< Pointer to BN Z-branch equivalent bias in device memory. */ CUDNN_PTR_ACTIVATION_BITMASK = 19, /**< Pointer to activation bitmask in device memory. */ CUDNN_PTR_DXDATA = 20, /**< Pointer to input gradient data (dX) in device memory. */ CUDNN_PTR_DZDATA = 21, /**< Pointer to residual input gradient data (dZ) in device memory. */ CUDNN_PTR_BN_DSCALE = 22, /**< Pointer to BN scale gradient in device memory. */ CUDNN_PTR_BN_DBIAS = 23, /**< Pointer to BN bias gradient in device memory. */ /* set/get: pass size_t* pointing to host memory */ CUDNN_SCALAR_SIZE_T_WORKSPACE_SIZE_IN_BYTES = 100, /**< Workspace size in bytes (host, size_t). */ /* set/get: pass int64_t* pointing to host memory */ CUDNN_SCALAR_INT64_T_BN_ACCUMULATION_COUNT = 101, /**< BN accumulation count (host, int64_t). */ /* set/get: pass double* pointing to host memory */ CUDNN_SCALAR_DOUBLE_BN_EXP_AVG_FACTOR = 102, /**< BN exponential average factor (host, double). */ /* set/get: pass double* pointing to host memory */ CUDNN_SCALAR_DOUBLE_BN_EPSILON = 103, /**< BN epsilon value (host, double). */ } cudnnFusedOpsVariantParamLabel_t CUDNN_DEPRECATED; /** * @brief Check the version of the cuDNN CNN library. * * Verifies that the CNN sub-library version matches the core cuDNN version. * * @return cudnnStatus_t indicating success or version mismatch. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnCnnVersionCheck(void); /* helper function to provide the convolution backward filter algo that fit best the requirement */ /** * @brief Performance results for backward filter convolution algorithm selection. * * Contains timing, memory usage, and determinism information for a given * backward filter convolution algorithm. Returned by algorithm search functions. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ typedef struct cudnnConvolutionBwdFilterAlgoPerfStruct { cudnnConvolutionBwdFilterAlgo_t algo; /**< The backward filter convolution algorithm. */ cudnnStatus_t status; /**< Status returned when running this algorithm. */ float time; /**< Execution time in milliseconds. */ size_t memory; /**< Workspace memory required in bytes. */ cudnnDeterminism_t determinism; /**< Whether the algorithm is deterministic. */ cudnnMathType_t mathType; /**< Math type used by the algorithm. */ int reserved[3]; /**< Reserved for future use. */ } cudnnConvolutionBwdFilterAlgoPerf_t CUDNN_DEPRECATED; /** * @brief Get the maximum number of backward filter convolution algorithms available. * * @param[in] handle The cuDNN handle. * @param[out] count Pointer to receive the maximum algorithm count. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardFilterAlgorithmMaxCount(cudnnHandle_t handle, int *count); /** * @brief Find the best backward filter convolution algorithm by running benchmarks. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] dwDesc Descriptor for the filter gradient. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionBackwardFilterAlgorithm(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const cudnnTensorDescriptor_t dyDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnFilterDescriptor_t dwDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdFilterAlgoPerf_t *perfResults); /** * @brief Find the best backward filter convolution algorithm with user-provided buffers. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] y Pointer to gradient output data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] dwDesc Descriptor for the filter gradient. * @param[out] dw Pointer to filter gradient data in device memory. * @param[in] requestedAlgoCount Maximum number of algorithms to test. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFindConvolutionBackwardFilterAlgorithmEx(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnTensorDescriptor_t dyDesc, const void *y, const cudnnConvolutionDescriptor_t convDesc, const cudnnFilterDescriptor_t dwDesc, void *dw, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdFilterAlgoPerf_t *perfResults, void *workSpace, size_t workSpaceSizeInBytes); /** * @brief Get backward filter convolution algorithm recommendations without executing them. * * @param[in] handle The cuDNN handle. * @param[in] srcDesc Descriptor for the input tensor. * @param[in] diffDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] gradDesc Descriptor for the filter gradient. * @param[in] requestedAlgoCount Maximum number of algorithms to return. * @param[out] returnedAlgoCount Pointer to receive the actual number returned. * @param[out] perfResults Array to receive the algorithm performance results. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardFilterAlgorithm_v7(cudnnHandle_t handle, const cudnnTensorDescriptor_t srcDesc, const cudnnTensorDescriptor_t diffDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnFilterDescriptor_t gradDesc, const int requestedAlgoCount, int *returnedAlgoCount, cudnnConvolutionBwdFilterAlgoPerf_t *perfResults); /* * convolution algorithm (which requires potentially some workspace) */ /** * @brief Get the minimum workspace size required for a backward filter convolution algorithm. * * @param[in] handle The cuDNN handle. * @param[in] xDesc Descriptor for the input tensor. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] convDesc The convolution descriptor. * @param[in] gradDesc Descriptor for the filter gradient. * @param[in] algo The backward filter convolution algorithm. * @param[out] sizeInBytes Pointer to receive the required workspace size. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Helper function to return the minimum size of the workspace to be passed to the convolution given an algo*/ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnnHandle_t handle, const cudnnTensorDescriptor_t xDesc, const cudnnTensorDescriptor_t dyDesc, const cudnnConvolutionDescriptor_t convDesc, const cudnnFilterDescriptor_t gradDesc, cudnnConvolutionBwdFilterAlgo_t algo, size_t *sizeInBytes); /** * @brief Execute a backward filter convolution (compute gradient with respect to filter weights). * * Computes: dw = alpha * dconv(x, dy) + beta * dw * * @param[in] handle The cuDNN handle. * @param[in] alpha Pointer to scaling factor for the convolution result. * @param[in] xDesc Descriptor for the input tensor. * @param[in] x Pointer to input data in device memory. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] dy Pointer to gradient output data in device memory. * @param[in] convDesc The convolution descriptor. * @param[in] algo The backward filter convolution algorithm. * @param[in] workSpace Pointer to workspace in device memory. * @param[in] workSpaceSizeInBytes Size of the workspace in bytes. * @param[in] beta Pointer to scaling factor for the prior filter gradient. * @param[in] dwDesc Descriptor for the filter gradient. * @param[in,out] dw Pointer to filter gradient data in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnConvolutionBackwardFilter(cudnnHandle_t handle, const void *alpha, const cudnnTensorDescriptor_t xDesc, const void *x, const cudnnTensorDescriptor_t dyDesc, const void *dy, const cudnnConvolutionDescriptor_t convDesc, cudnnConvolutionBwdFilterAlgo_t algo, void *workSpace, size_t workSpaceSizeInBytes, const void *beta, const cudnnFilterDescriptor_t dwDesc, void *dw); /** * @brief Compute the bias gradient for batch convolution. * * Computes: db = alpha * sum(dy) + beta * db, where the sum is over spatial dimensions and batch. * * @param[in] handle The cuDNN handle. * @param[in] alpha Pointer to scaling factor for the bias gradient result. * @param[in] dyDesc Descriptor for the gradient output tensor. * @param[in] dy Pointer to gradient output data in device memory. * @param[in] beta Pointer to scaling factor for the prior bias gradient. * @param[in] dbDesc Descriptor for the bias gradient tensor. * @param[in,out] db Pointer to bias gradient data in device memory. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ /* Function to compute the bias gradient for batch convolution */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnConvolutionBackwardBias(cudnnHandle_t handle, const void *alpha, const cudnnTensorDescriptor_t dyDesc, const void *dy, const void *beta, const cudnnTensorDescriptor_t dbDesc, void *db); /** * @brief Create a fused operations constant parameter pack. * * @param[out] constPack Pointer to receive the newly created constant parameter pack. * @param[in] ops The fused operation type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateFusedOpsConstParamPack(cudnnFusedOpsConstParamPack_t *constPack, cudnnFusedOps_t ops); /** * @brief Destroy a fused operations constant parameter pack. * * @param[in] constPack The constant parameter pack to destroy. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroyFusedOpsConstParamPack(cudnnFusedOpsConstParamPack_t constPack); /** * @brief Set an attribute on a fused operations constant parameter pack. * * @param[in,out] constPack The constant parameter pack. * @param[in] paramLabel The label identifying which parameter to set. * @param[in] param Pointer to the parameter value. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetFusedOpsConstParamPackAttribute(cudnnFusedOpsConstParamPack_t constPack, cudnnFusedOpsConstParamLabel_t paramLabel, const void *param); /** * @brief Get an attribute from a fused operations constant parameter pack. * * @param[in] constPack The constant parameter pack. * @param[in] paramLabel The label identifying which parameter to get. * @param[out] param Pointer to receive the parameter value. * @param[out] isNULL Pointer to receive whether the parameter is NULL. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetFusedOpsConstParamPackAttribute(const cudnnFusedOpsConstParamPack_t constPack, cudnnFusedOpsConstParamLabel_t paramLabel, void *param, int *isNULL); /** * @brief Create a fused operations variant parameter pack. * * @param[out] varPack Pointer to receive the newly created variant parameter pack. * @param[in] ops The fused operation type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateFusedOpsVariantParamPack(cudnnFusedOpsVariantParamPack_t *varPack, cudnnFusedOps_t ops); /** * @brief Destroy a fused operations variant parameter pack. * * @param[in] varPack The variant parameter pack to destroy. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroyFusedOpsVariantParamPack(cudnnFusedOpsVariantParamPack_t varPack); /** * @brief Set an attribute on a fused operations variant parameter pack. * * @param[in,out] varPack The variant parameter pack. * @param[in] paramLabel The label identifying which parameter to set. * @param[in] ptr Pointer to the parameter value. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetFusedOpsVariantParamPackAttribute(cudnnFusedOpsVariantParamPack_t varPack, cudnnFusedOpsVariantParamLabel_t paramLabel, void *ptr); /** * @brief Get an attribute from a fused operations variant parameter pack. * * @param[in] varPack The variant parameter pack. * @param[in] paramLabel The label identifying which parameter to get. * @param[out] ptr Pointer to receive the parameter value. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetFusedOpsVariantParamPackAttribute(const cudnnFusedOpsVariantParamPack_t varPack, cudnnFusedOpsVariantParamLabel_t paramLabel, void *ptr); /** * @brief Create a fused operations execution plan. * * @param[out] plan Pointer to receive the newly created fused operations plan. * @param[in] ops The fused operation type. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateFusedOpsPlan(cudnnFusedOpsPlan_t *plan, cudnnFusedOps_t ops); /** * @brief Destroy a fused operations execution plan. * * @param[in] plan The fused operations plan to destroy. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroyFusedOpsPlan(cudnnFusedOpsPlan_t plan); /** * @brief Build a fused operations execution plan from constant parameters. * * Compiles the plan and returns the required workspace size. * * @param[in] handle The cuDNN handle. * @param[in,out] plan The fused operations plan to build. * @param[in] constPack The constant parameter pack with descriptors. * @param[out] workspaceSizeInBytes Pointer to receive the required workspace size. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnMakeFusedOpsPlan(cudnnHandle_t handle, cudnnFusedOpsPlan_t plan, const cudnnFusedOpsConstParamPack_t constPack, size_t *workspaceSizeInBytes); /** * @brief Execute a fused operations plan. * * @param[in] handle The cuDNN handle. * @param[in] plan The fused operations plan to execute. * @param[in] varPack The variant parameter pack with data pointers and scalar values. * @return cudnnStatus_t indicating success or failure. * * @deprecated Since cuDNN 9.0.0. Use graph API instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnFusedOpsExecute(cudnnHandle_t handle, const cudnnFusedOpsPlan_t plan, cudnnFusedOpsVariantParamPack_t varPack); #if defined(__cplusplus) } #endif #endif /* CUDNN_CNN_H_ */