/* * 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_adv.h * @brief cuDNN Advanced library - RNN, CTC loss, multi-head attention, and related operations. * @since cuDNN 9.0.0 */ #if !defined(CUDNN_ADV_H_) #define CUDNN_ADV_H_ #include #include "cudnn_version.h" #include "cudnn_ops.h" /* These version numbers are autogenerated, do not edit manually. */ #define CUDNN_ADV_MAJOR 9 #define CUDNN_ADV_MINOR 24 #define CUDNN_ADV_PATCH 0 #if (CUDNN_ADV_MAJOR != CUDNN_MAJOR) || (CUDNN_ADV_MINOR != CUDNN_MINOR) || (CUDNN_ADV_PATCH != CUDNN_PATCHLEVEL) #error Version mismatch in cuDNN ADV INFER!!! #endif #if defined(__cplusplus) extern "C" { #endif /* BASIC RNN API */ /** * @brief RNN computation algorithm selection. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_RNN_ALGO_STANDARD = 0, /**< Standard cuBLASLt-based algorithm. @since cuDNN 9.0.0 */ CUDNN_RNN_ALGO_PERSIST_STATIC = 1, /**< Persistent kernel with static compilation. @since cuDNN 9.0.0 */ CUDNN_RNN_ALGO_PERSIST_DYNAMIC = 2, /**< Runtime-compiled persistent kernels via NVRTC. @since cuDNN 9.0.0 */ CUDNN_RNN_ALGO_PERSIST_STATIC_SMALL_H = 3, /**< Register-based approach for smaller hidden states. @since cuDNN 9.0.0 */ CUDNN_RNN_ALGO_COUNT = 4, /**< Number of RNN algorithms. @since cuDNN 9.0.0 */ } cudnnRNNAlgo_t; /** * @brief Specifies inference or training mode for RNN forward pass. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_FWD_MODE_INFERENCE = 0, /**< Inference mode. @since cuDNN 9.0.0 */ CUDNN_FWD_MODE_TRAINING = 1, /**< Training mode (reserves space for backward pass). @since cuDNN 9.0.0 */ } cudnnForwardMode_t; /** * @brief RNN cell type selection. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_RNN_RELU = 0, /**< Single-gate RNN cell with ReLU activation. @since cuDNN 9.0.0 */ CUDNN_RNN_TANH = 1, /**< Single-gate RNN cell with tanh activation. @since cuDNN 9.0.0 */ CUDNN_LSTM = 2, /**< Four-gate LSTM with optional recurrent projection and clipping. @since cuDNN 9.0.0 */ CUDNN_GRU = 3, /**< Three-gate GRU network. @since cuDNN 9.0.0 */ } cudnnRNNMode_t; /** * @brief Number of bias vectors used in RNN cell formulas. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_RNN_NO_BIAS = 0, /**< No biases used. @since cuDNN 9.0.0 */ CUDNN_RNN_SINGLE_INP_BIAS = 1, /**< One input bias in input GEMM. @since cuDNN 9.0.0 */ CUDNN_RNN_DOUBLE_BIAS = 2, /**< Two bias vectors (default). @since cuDNN 9.0.0 */ CUDNN_RNN_SINGLE_REC_BIAS = 3 /**< One recurrent bias in recurrent GEMM. @since cuDNN 9.0.0 */ } cudnnRNNBiasMode_t; /** * @brief RNN recurrence direction mode. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_UNIDIRECTIONAL = 0, /**< Single direction, first input to last. @since cuDNN 9.0.0 */ CUDNN_BIDIRECTIONAL = 1, /**< Both directions, outputs concatenated at each layer. @since cuDNN 9.0.0 */ } cudnnDirectionMode_t; /** * @brief RNN first layer input behavior. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_LINEAR_INPUT = 0, /**< Biased matrix multiplication at first layer. @since cuDNN 9.0.0 */ CUDNN_SKIP_INPUT = 1, /**< Fixed identity matrix at first layer (no operation). @since cuDNN 9.0.0 */ } cudnnRNNInputMode_t; /** * @brief LSTM cell clipping mode. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_RNN_CLIP_NONE = 0, /**< Disables LSTM cell clipping. @since cuDNN 9.0.0 */ CUDNN_RNN_CLIP_MINMAX = 1, /**< Enables LSTM cell clipping. @since cuDNN 9.0.0 */ } cudnnRNNClipMode_t; /** * @brief RNN data memory layout. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_UNPACKED = 0, /**< Padded, sequence-major layout. @since cuDNN 9.0.0 */ CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_PACKED = 1, /**< Packed, sequence-major layout. @since cuDNN 9.0.0 */ CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED = 2, /**< Padded, batch-major layout. @since cuDNN 9.0.0 */ } cudnnRNNDataLayout_t; /* For auxFlags in cudnnSetRNNDescriptor_v8() */ #define CUDNN_RNN_PADDED_IO_DISABLED 0 #define CUDNN_RNN_PADDED_IO_ENABLED (1U << 0) /** @brief Opaque RNN descriptor. @since cuDNN 9.0.0 */ struct cudnnRNNStruct; typedef struct cudnnRNNStruct *cudnnRNNDescriptor_t; /** @brief Opaque RNN data descriptor. @since cuDNN 9.0.0 */ struct cudnnRNNDataStruct; typedef struct cudnnRNNDataStruct *cudnnRNNDataDescriptor_t; /** * @brief Creates an RNN descriptor. * * @param[out] rnnDesc Pointer to the created RNN descriptor. * * @retval CUDNN_STATUS_SUCCESS Descriptor created successfully. * * @since cuDNN 9.0.0 * @see cudnnDestroyRNNDescriptor */ cudnnStatus_t CUDNNWINAPI cudnnCreateRNNDescriptor(cudnnRNNDescriptor_t *rnnDesc); /** * @brief Destroys an RNN descriptor. * * @param[in] rnnDesc RNN descriptor to destroy. * * @retval CUDNN_STATUS_SUCCESS Descriptor destroyed successfully. * * @since cuDNN 9.0.0 * @see cudnnCreateRNNDescriptor */ cudnnStatus_t CUDNNWINAPI cudnnDestroyRNNDescriptor(cudnnRNNDescriptor_t rnnDesc); /* * mathPrec in cudnnSetRNNDescriptor_v8() specifies compute precision. * Compute precision is further modified by mathType that sets the * preferred option for using NVIDIA Tensor Cores. dataType specify * input/output data type and weight/bias type. */ /** * @brief Configures an RNN descriptor with network parameters. * * @param[in,out] rnnDesc RNN descriptor to configure. * @param[in] algo RNN computation algorithm. * @param[in] cellMode RNN cell type (RELU, TANH, LSTM, GRU). * @param[in] biasMode Bias configuration. * @param[in] dirMode Unidirectional or bidirectional. * @param[in] inputMode First layer input behavior. * @param[in] dataType Input/output and weight data type. * @param[in] mathPrec Compute precision. * @param[in] mathType Tensor Core usage preference. * @param[in] inputSize Input vector size. * @param[in] hiddenSize Hidden state size. * @param[in] projSize Recurrent projection size (0 to disable). * @param[in] numLayers Number of stacked RNN layers. * @param[in] dropoutDesc Dropout descriptor for inter-layer dropout. * @param[in] auxFlags Auxiliary flags (e.g., CUDNN_RNN_PADDED_IO_ENABLED). * * @retval CUDNN_STATUS_SUCCESS Descriptor configured successfully. * @retval CUDNN_STATUS_BAD_PARAM Invalid parameter. * @retval CUDNN_STATUS_NOT_SUPPORTED Unsupported configuration. * * @since cuDNN 9.0.0 * @see cudnnGetRNNDescriptor_v8 */ cudnnStatus_t CUDNNWINAPI cudnnSetRNNDescriptor_v8(cudnnRNNDescriptor_t rnnDesc, cudnnRNNAlgo_t algo, cudnnRNNMode_t cellMode, cudnnRNNBiasMode_t biasMode, cudnnDirectionMode_t dirMode, cudnnRNNInputMode_t inputMode, cudnnDataType_t dataType, cudnnDataType_t mathPrec, cudnnMathType_t mathType, int32_t inputSize, int32_t hiddenSize, int32_t projSize, int32_t numLayers, cudnnDropoutDescriptor_t dropoutDesc, uint32_t auxFlags); /** * @brief Retrieves RNN descriptor parameters. * * @param[in] rnnDesc RNN descriptor to query. * @param[out] algo RNN algorithm. * @param[out] cellMode Cell type. * @param[out] biasMode Bias configuration. * @param[out] dirMode Direction mode. * @param[out] inputMode Input mode. * @param[out] dataType Data type. * @param[out] mathPrec Math precision. * @param[out] mathType Math type. * @param[out] inputSize Input size. * @param[out] hiddenSize Hidden size. * @param[out] projSize Projection size. * @param[out] numLayers Number of layers. * @param[out] dropoutDesc Dropout descriptor. * @param[out] auxFlags Auxiliary flags. * * @retval CUDNN_STATUS_SUCCESS Query succeeded. * * @since cuDNN 9.0.0 * @see cudnnSetRNNDescriptor_v8 */ cudnnStatus_t CUDNNWINAPI cudnnGetRNNDescriptor_v8(cudnnRNNDescriptor_t rnnDesc, cudnnRNNAlgo_t *algo, cudnnRNNMode_t *cellMode, cudnnRNNBiasMode_t *biasMode, cudnnDirectionMode_t *dirMode, cudnnRNNInputMode_t *inputMode, cudnnDataType_t *dataType, cudnnDataType_t *mathPrec, cudnnMathType_t *mathType, int32_t *inputSize, int32_t *hiddenSize, int32_t *projSize, int32_t *numLayers, cudnnDropoutDescriptor_t *dropoutDesc, uint32_t *auxFlags); /** * @brief Configures LSTM cell clipping parameters. * * @deprecated Since cuDNN 9.0.0. Use cudnnRNNSetClip_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnRNNSetClip_v8(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t clipMode, cudnnNanPropagation_t clipNanOpt, double lclip, double rclip); /** * @brief Configures LSTM cell clipping parameters. * * @param[in,out] rnnDesc RNN descriptor. * @param[in] clipMode Clipping mode (NONE or MINMAX). * @param[in] lclip Left (minimum) clipping value. * @param[in] rclip Right (maximum) clipping value. * * @retval CUDNN_STATUS_SUCCESS Clipping configured. * * @since cuDNN 9.0.0 * @see cudnnRNNGetClip_v9 */ cudnnStatus_t CUDNNWINAPI cudnnRNNSetClip_v9(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t clipMode, double lclip, double rclip); /** * @brief Retrieves LSTM cell clipping settings. * * @deprecated Since cuDNN 9.0.0. Use cudnnRNNGetClip_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnRNNGetClip_v8(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t *clipMode, cudnnNanPropagation_t *clipNanOpt, double *lclip, double *rclip); /** * @brief Retrieves LSTM cell clipping settings. * * @param[in] rnnDesc RNN descriptor. * @param[out] clipMode Clipping mode. * @param[out] lclip Left clipping value. * @param[out] rclip Right clipping value. * * @retval CUDNN_STATUS_SUCCESS Query succeeded. * * @since cuDNN 9.0.0 * @see cudnnRNNSetClip_v9 */ cudnnStatus_t CUDNNWINAPI cudnnRNNGetClip_v9(cudnnRNNDescriptor_t rnnDesc, cudnnRNNClipMode_t *clipMode, double *lclip, double *rclip); /** * @brief Compiles persistent RNN code using NVRTC for dynamic algorithm. * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor (must use PERSIST_DYNAMIC algorithm). * @param[in] miniBatch Exact mini-batch size for compilation. * * @retval CUDNN_STATUS_SUCCESS Compilation succeeded. * @retval CUDNN_STATUS_NOT_SUPPORTED Unsupported configuration. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnBuildRNNDynamic(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, int miniBatch); /** * @brief Computes workspace and reserve space buffer sizes for RNN. * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[in] fwdMode Inference or training mode. * @param[in] xDesc Input data descriptor. * @param[out] workSpaceSize Required workspace size in bytes. * @param[out] reserveSpaceSize Required reserve space size in bytes (training only). * * @retval CUDNN_STATUS_SUCCESS Sizes computed successfully. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnGetRNNTempSpaceSizes(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, cudnnForwardMode_t fwdMode, cudnnRNNDataDescriptor_t xDesc, size_t *workSpaceSize, size_t *reserveSpaceSize); /** * @brief Reports required GPU memory for all RNN weight parameters. * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[out] weightSpaceSize Required weight space size in bytes. * * @retval CUDNN_STATUS_SUCCESS Size computed. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnGetRNNWeightSpaceSize(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, size_t *weightSpaceSize); /** * @brief Obtains start address and shape of RNN weight matrices and bias vectors. * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[in] pseudoLayer Pseudo-layer index (physical layer and direction). * @param[in] weightSpaceSize Total weight space size. * @param[in] weightSpace Pointer to weight space. * @param[in] linLayerID Linear layer ID within the RNN cell. * @param[out] mDesc Tensor descriptor for the weight matrix. * @param[out] mAddr Start address of the weight matrix. * @param[out] bDesc Tensor descriptor for the bias vector. * @param[out] bAddr Start address of the bias vector. * * @retval CUDNN_STATUS_SUCCESS Parameters retrieved. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnGetRNNWeightParams(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, int32_t pseudoLayer, size_t weightSpaceSize, const void *weightSpace, int32_t linLayerID, cudnnTensorDescriptor_t mDesc, void **mAddr, cudnnTensorDescriptor_t bDesc, void **bAddr); /** * @brief Creates an RNN data descriptor. * @param[out] rnnDataDesc Pointer to created descriptor. * @retval CUDNN_STATUS_SUCCESS Descriptor created. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnCreateRNNDataDescriptor(cudnnRNNDataDescriptor_t *rnnDataDesc); /** * @brief Destroys an RNN data descriptor. * @param[in] rnnDataDesc Descriptor to destroy. * @retval CUDNN_STATUS_SUCCESS Descriptor destroyed. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnDestroyRNNDataDescriptor(cudnnRNNDataDescriptor_t rnnDataDesc); /** * @brief Configures an RNN data descriptor with layout and sequence information. * * @param[in,out] rnnDataDesc RNN data descriptor. * @param[in] dataType Data type. * @param[in] layout Data layout (sequence-major or batch-major). * @param[in] maxSeqLength Maximum sequence length. * @param[in] batchSize Batch size. * @param[in] vectorSize Input vector size. * @param[in] seqLengthArray Length of each sequence in the batch. * @param[in,out] paddingFill Symbol for filling padding positions. * * @retval CUDNN_STATUS_SUCCESS Descriptor configured. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnSetRNNDataDescriptor(cudnnRNNDataDescriptor_t rnnDataDesc, cudnnDataType_t dataType, cudnnRNNDataLayout_t layout, int maxSeqLength, int batchSize, int vectorSize, const int seqLengthArray[], /* length of each sequence in the batch */ void *paddingFill); /* symbol for filling padding position in output */ /** * @brief Retrieves RNN data descriptor parameters. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnGetRNNDataDescriptor(cudnnRNNDataDescriptor_t rnnDataDesc, cudnnDataType_t *dataType, cudnnRNNDataLayout_t *layout, int *maxSeqLength, int *batchSize, int *vectorSize, int arrayLengthRequested, int seqLengthArray[], void *paddingFill); /** * @brief Computes the forward pass of an RNN network. * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[in] fwdMode Inference or training mode. * @param[in] devSeqLengths Per-batch sequence lengths (device memory). * @param[in] xDesc Input data descriptor. * @param[in] x Input data pointer. * @param[in] yDesc Output data descriptor. * @param[out] y Output data pointer. * @param[in] hDesc Hidden state descriptor. * @param[in] hx Initial hidden state (NULL for zero). * @param[out] hy Final hidden state (NULL to discard). * @param[in] cDesc Cell state descriptor (LSTM only). * @param[in] cx Initial cell state (NULL for zero). * @param[out] cy Final cell state (NULL to discard). * @param[in] weightSpaceSize Weight space size in bytes. * @param[in] weightSpace Weight space pointer. * @param[in] workSpaceSize Workspace size in bytes. * @param[in,out] workSpace Workspace pointer. * @param[in] reserveSpaceSize Reserve space size (training only). * @param[in,out] reserveSpace Reserve space pointer (training only). * * @retval CUDNN_STATUS_SUCCESS Forward pass completed. * @retval CUDNN_STATUS_BAD_PARAM Invalid parameter. * @retval CUDNN_STATUS_EXECUTION_FAILED Execution failed. * * @since cuDNN 9.0.0 * @see cudnnRNNBackwardData_v8, cudnnRNNBackwardWeights_v8 */ cudnnStatus_t CUDNNWINAPI cudnnRNNForward(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, cudnnForwardMode_t fwdMode, const int32_t devSeqLengths[], cudnnRNNDataDescriptor_t xDesc, const void *x, cudnnRNNDataDescriptor_t yDesc, void *y, cudnnTensorDescriptor_t hDesc, const void *hx, void *hy, cudnnTensorDescriptor_t cDesc, const void *cx, void *cy, size_t weightSpaceSize, const void *weightSpace, size_t workSpaceSize, void *workSpace, size_t reserveSpaceSize, void *reserveSpace); /* Sequence data descriptor */ /** * @brief Sequence data dimension indices. * @deprecated Since cuDNN 9.0.0. Use RNN data descriptors instead. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_SEQDATA_TIME_DIM = 0, /**< Time/sequence length dimension. @since cuDNN 9.0.0 */ CUDNN_SEQDATA_BATCH_DIM = 1, /**< Batch dimension. @since cuDNN 9.0.0 */ CUDNN_SEQDATA_BEAM_DIM = 2, /**< Beam dimension. @since cuDNN 9.0.0 */ CUDNN_SEQDATA_VECT_DIM = 3 /**< Vector dimension. @since cuDNN 9.0.0 */ } cudnnSeqDataAxis_t; /** @brief Opaque sequence data descriptor. @deprecated Since cuDNN 9.0.0. @since cuDNN 9.0.0 */ struct cudnnSeqDataStruct; typedef struct cudnnSeqDataStruct *cudnnSeqDataDescriptor_t CUDNN_DEPRECATED; #define CUDNN_SEQDATA_DIM_COUNT 4 /* dimension count */ /** * @brief Creates a sequence data descriptor. * @deprecated Since cuDNN 9.0.0. Use RNN data descriptors instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateSeqDataDescriptor(cudnnSeqDataDescriptor_t *seqDataDesc); /** * @brief Destroys a sequence data descriptor. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroySeqDataDescriptor(cudnnSeqDataDescriptor_t seqDataDesc); /** * @brief Configures a sequence data descriptor. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetSeqDataDescriptor(cudnnSeqDataDescriptor_t seqDataDesc, cudnnDataType_t dataType, int nbDims, const int dimA[], const cudnnSeqDataAxis_t axes[], size_t seqLengthArraySize, const int seqLengthArray[], void *paddingFill); /** * @brief Retrieves sequence data descriptor parameters. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetSeqDataDescriptor(const cudnnSeqDataDescriptor_t seqDataDesc, cudnnDataType_t *dataType, int *nbDims, int nbDimsRequested, int dimA[], cudnnSeqDataAxis_t axes[], size_t *seqLengthArraySize, size_t seqLengthSizeRequested, int seqLengthArray[], void *paddingFill); /* Multihead Attention */ /* * Multi-head attention options passed via 'attnMode' in cudnnSetAttnDescriptor(). * Use the bitwise OR operator to combine several settings listed below. Additional * minor options can be added here w/o changing or introducing new API functions. */ #define CUDNN_ATTN_QUERYMAP_ALL_TO_ONE 0 /* multiple Q-s map to a single (K,V) set when beam size > 1 */ #define CUDNN_ATTN_QUERYMAP_ONE_TO_ONE (1U << 0) /* multiple Q-s map to multiple (K,V) sets when beam size > 1 */ #define CUDNN_ATTN_DISABLE_PROJ_BIASES 0 /* no biases in attention input and output projections */ #define CUDNN_ATTN_ENABLE_PROJ_BIASES (1U << 1) /* use biases in attention input and output projections */ /** @brief Opaque multi-head attention descriptor. @deprecated Since cuDNN 9.0.0. @since cuDNN 9.0.0 */ struct cudnnAttnStruct; typedef struct cudnnAttnStruct *cudnnAttnDescriptor_t CUDNN_DEPRECATED; /** * @brief Creates a multi-head attention descriptor. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnCreateAttnDescriptor(cudnnAttnDescriptor_t *attnDesc); /** * @brief Destroys a multi-head attention descriptor. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnDestroyAttnDescriptor(cudnnAttnDescriptor_t attnDesc); /** * @brief Configures a multi-head attention descriptor. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetAttnDescriptor(cudnnAttnDescriptor_t attnDesc, unsigned attnMode, int nHeads, double smScaler, cudnnDataType_t dataType, cudnnDataType_t computePrec, cudnnMathType_t mathType, cudnnDropoutDescriptor_t attnDropoutDesc, cudnnDropoutDescriptor_t postDropoutDesc, int qSize, int kSize, int vSize, int qProjSize, int kProjSize, int vProjSize, int oProjSize, int qoMaxSeqLength, int kvMaxSeqLength, int maxBatchSize, int maxBeamSize); /** * @brief Retrieves multi-head attention descriptor parameters. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetAttnDescriptor(cudnnAttnDescriptor_t attnDesc, unsigned *attnMode, int *nHeads, double *smScaler, cudnnDataType_t *dataType, cudnnDataType_t *computePrec, cudnnMathType_t *mathType, cudnnDropoutDescriptor_t *attnDropoutDesc, cudnnDropoutDescriptor_t *postDropoutDesc, int *qSize, int *kSize, int *vSize, int *qProjSize, int *kProjSize, int *vProjSize, int *oProjSize, int *qoMaxSeqLength, int *kvMaxSeqLength, int *maxBatchSize, int *maxBeamSize); /** * @brief Computes weight, workspace, and reserve space sizes for multi-head attention. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetMultiHeadAttnBuffers(cudnnHandle_t handle, const cudnnAttnDescriptor_t attnDesc, size_t *weightSizeInBytes, size_t *workSpaceSizeInBytes, size_t *reserveSpaceSizeInBytes); /** * @brief Specifies weight/bias groups in multi-head attention layers. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_MH_ATTN_Q_WEIGHTS = 0, /**< Query projection weights. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_K_WEIGHTS = 1, /**< Key projection weights. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_V_WEIGHTS = 2, /**< Value projection weights. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_O_WEIGHTS = 3, /**< Output projection weights. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_Q_BIASES = 4, /**< Query projection biases. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_K_BIASES = 5, /**< Key projection biases. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_V_BIASES = 6, /**< Value projection biases. @since cuDNN 9.0.0 */ CUDNN_MH_ATTN_O_BIASES = 7, /**< Output projection biases. @since cuDNN 9.0.0 */ } cudnnMultiHeadAttnWeightKind_t; #define CUDNN_ATTN_WKIND_COUNT 8 /* Number of attention weight/bias tensors */ /** * @brief Obtains shape and start address of attention weight/bias tensors. * @deprecated Since cuDNN 9.0.0. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetMultiHeadAttnWeights(cudnnHandle_t handle, const cudnnAttnDescriptor_t attnDesc, cudnnMultiHeadAttnWeightKind_t wKind, size_t weightSizeInBytes, const void *weights, cudnnTensorDescriptor_t wDesc, void **wAddr); /** * @brief Computes multi-head attention forward pass. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnMultiHeadAttnForward(cudnnHandle_t handle, const cudnnAttnDescriptor_t attnDesc, int currIdx, const int loWinIdx[], const int hiWinIdx[], const int devSeqLengthsQO[], const int devSeqLengthsKV[], const cudnnSeqDataDescriptor_t qDesc, const void *queries, const void *residuals, const cudnnSeqDataDescriptor_t kDesc, const void *keys, const cudnnSeqDataDescriptor_t vDesc, const void *values, const cudnnSeqDataDescriptor_t oDesc, void *out, size_t weightSizeInBytes, const void *weights, size_t workSpaceSizeInBytes, void *workSpace, size_t reserveSpaceSizeInBytes, void *reserveSpace); /* * \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 cudnnAdvVersionCheck(void); /** * @brief Weight gradient accumulation mode. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_WGRAD_MODE_ADD = 0, /**< Add partial gradients to existing buffer. @since cuDNN 9.0.0 */ CUDNN_WGRAD_MODE_SET = 1, /**< Overwrite buffer with partial gradients. @since cuDNN 9.0.0 */ } cudnnWgradMode_t; /** * @brief Computes RNN data gradients (backward pass with respect to inputs). * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[in] devSeqLengths Per-batch sequence lengths (device memory). * @param[in] yDesc Output data descriptor. * @param[in] y Forward output data. * @param[in] dy Output gradient data. * @param[in] xDesc Input data descriptor. * @param[out] dx Computed input gradient. * @param[in] hDesc Hidden state descriptor. * @param[in] hx Initial hidden state from forward pass. * @param[in] dhy Hidden state gradient (from upstream). * @param[out] dhx Computed initial hidden state gradient. * @param[in] cDesc Cell state descriptor (LSTM only). * @param[in] cx Initial cell state from forward pass. * @param[in] dcy Cell state gradient (from upstream). * @param[out] dcx Computed initial cell state gradient. * @param[in] weightSpaceSize Weight space size. * @param[in] weightSpace Weight space pointer. * @param[in] workSpaceSize Workspace size. * @param[in,out] workSpace Workspace pointer. * @param[in] reserveSpaceSize Reserve space size. * @param[in,out] reserveSpace Reserve space (from forward training pass). * * @retval CUDNN_STATUS_SUCCESS Backward data pass completed. * * @since cuDNN 9.0.0 * @see cudnnRNNForward, cudnnRNNBackwardWeights_v8 */ cudnnStatus_t CUDNNWINAPI cudnnRNNBackwardData_v8(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, const int32_t devSeqLengths[], cudnnRNNDataDescriptor_t yDesc, const void *y, const void *dy, cudnnRNNDataDescriptor_t xDesc, void *dx, cudnnTensorDescriptor_t hDesc, const void *hx, const void *dhy, void *dhx, cudnnTensorDescriptor_t cDesc, const void *cx, const void *dcy, void *dcx, size_t weightSpaceSize, const void *weightSpace, size_t workSpaceSize, void *workSpace, size_t reserveSpaceSize, void *reserveSpace); /** * @brief Computes RNN weight gradients (backward pass with respect to parameters). * * @param[in] handle cuDNN handle. * @param[in] rnnDesc RNN descriptor. * @param[in] addGrad Accumulate (ADD) or overwrite (SET) gradients. * @param[in] devSeqLengths Per-batch sequence lengths (device memory). * @param[in] xDesc Input data descriptor. * @param[in] x Input data. * @param[in] hDesc Hidden state descriptor. * @param[in] hx Initial hidden state. * @param[in] yDesc Output data descriptor. * @param[in] y Forward output data. * @param[in] weightSpaceSize Weight space size. * @param[in,out] dweightSpace Computed weight gradients. * @param[in] workSpaceSize Workspace size. * @param[in,out] workSpace Workspace pointer. * @param[in] reserveSpaceSize Reserve space size. * @param[in,out] reserveSpace Reserve space (from forward training pass). * * @retval CUDNN_STATUS_SUCCESS Weight gradients computed. * * @since cuDNN 9.0.0 * @see cudnnRNNForward, cudnnRNNBackwardData_v8 */ cudnnStatus_t CUDNNWINAPI cudnnRNNBackwardWeights_v8(cudnnHandle_t handle, cudnnRNNDescriptor_t rnnDesc, cudnnWgradMode_t addGrad, const int32_t devSeqLengths[], cudnnRNNDataDescriptor_t xDesc, const void *x, cudnnTensorDescriptor_t hDesc, const void *hx, cudnnRNNDataDescriptor_t yDesc, const void *y, size_t weightSpaceSize, void *dweightSpace, size_t workSpaceSize, void *workSpace, size_t reserveSpaceSize, void *reserveSpace); /** * @brief Computes multi-head attention data gradients. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnMultiHeadAttnBackwardData(cudnnHandle_t handle, const cudnnAttnDescriptor_t attnDesc, const int loWinIdx[], const int hiWinIdx[], const int devSeqLengthsDQDO[], const int devSeqLengthsDKDV[], const cudnnSeqDataDescriptor_t doDesc, const void *dout, const cudnnSeqDataDescriptor_t dqDesc, void *dqueries, const void *queries, const cudnnSeqDataDescriptor_t dkDesc, void *dkeys, const void *keys, const cudnnSeqDataDescriptor_t dvDesc, void *dvalues, const void *values, size_t weightSizeInBytes, const void *weights, size_t workSpaceSizeInBytes, void *workSpace, size_t reserveSpaceSizeInBytes, void *reserveSpace); /** * @brief Computes multi-head attention weight gradients. * @deprecated Since cuDNN 9.0.0. Use graph API SDPA operations instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnMultiHeadAttnBackwardWeights(cudnnHandle_t handle, const cudnnAttnDescriptor_t attnDesc, cudnnWgradMode_t addGrad, const cudnnSeqDataDescriptor_t qDesc, const void *queries, const cudnnSeqDataDescriptor_t kDesc, const void *keys, const cudnnSeqDataDescriptor_t vDesc, const void *values, const cudnnSeqDataDescriptor_t doDesc, const void *dout, size_t weightSizeInBytes, const void *weights, void *dweights, size_t workSpaceSizeInBytes, void *workSpace, size_t reserveSpaceSizeInBytes, void *reserveSpace); /* * CTC (Connectionist Temporal Classification) loss descriptor create/destory/set/get functions */ /** * @brief Input normalization mode for loss functions. * @since cuDNN 9.0.0 */ typedef enum { CUDNN_LOSS_NORMALIZATION_NONE = 0, /**< Input treated as normalized probability. @since cuDNN 9.0.0 */ CUDNN_LOSS_NORMALIZATION_SOFTMAX = 1, /**< Input treated as unnormalized activation (softmax applied). @since cuDNN 9.0.0 */ } cudnnLossNormalizationMode_t; /** * @brief Creates a CTC loss descriptor. * @param[out] ctcLossDesc Pointer to created descriptor. * @retval CUDNN_STATUS_SUCCESS Descriptor created. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnCreateCTCLossDescriptor(cudnnCTCLossDescriptor_t *ctcLossDesc); /** * @brief Configures a CTC loss descriptor with compute type. * @deprecated Since cuDNN 9.0.0. Use cudnnSetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t compType); /** * @brief Configures CTC loss with normalization mode. * @deprecated Since cuDNN 9.0.0. Use cudnnSetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetCTCLossDescriptorEx(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t compType, cudnnLossNormalizationMode_t normMode, cudnnNanPropagation_t gradMode); /** * @brief Configures CTC loss with normalization, gradient mode, and max label length. * @deprecated Since cuDNN 9.0.0. Use cudnnSetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnSetCTCLossDescriptor_v8(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t compType, cudnnLossNormalizationMode_t normMode, cudnnNanPropagation_t gradMode, int maxLabelLength); /** * @brief Configures CTC loss with normalization, CTC gradient mode, and max label length. * * @param[in,out] ctcLossDesc CTC loss descriptor. * @param[in] compType Compute data type. * @param[in] normMode Loss normalization mode. * @param[in] ctcGradMode Gradient mode for out-of-bounds samples. * @param[in] maxLabelLength Maximum label length. * * @retval CUDNN_STATUS_SUCCESS Descriptor configured. * * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnSetCTCLossDescriptor_v9(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t compType, cudnnLossNormalizationMode_t normMode, cudnnCTCGradMode_t ctcGradMode, int maxLabelLength); /** * @brief Retrieves CTC loss compute type. * @deprecated Since cuDNN 9.0.0. Use cudnnGetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t *compType); /** * @brief Retrieves CTC loss extended parameters. * @deprecated Since cuDNN 9.0.0. Use cudnnGetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossDescriptorEx(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t *compType, cudnnLossNormalizationMode_t *normMode, cudnnNanPropagation_t *gradMode); /** * @brief Retrieves CTC loss v8 parameters. * @deprecated Since cuDNN 9.0.0. Use cudnnGetCTCLossDescriptor_v9 instead. * @since cuDNN 9.0.0 */ CUDNN_DEPRECATED cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossDescriptor_v8(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t *compType, cudnnLossNormalizationMode_t *normMode, cudnnNanPropagation_t *gradMode, int *maxLabelLength); /** * @brief Retrieves CTC loss v9 parameters. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossDescriptor_v9(cudnnCTCLossDescriptor_t ctcLossDesc, cudnnDataType_t *compType, cudnnLossNormalizationMode_t *normMode, cudnnCTCGradMode_t *ctcGradMode, int *maxLabelLength); /** * @brief Destroys a CTC loss descriptor. * @param[in] ctcLossDesc Descriptor to destroy. * @retval CUDNN_STATUS_SUCCESS Descriptor destroyed. * @since cuDNN 9.0.0 */ cudnnStatus_t CUDNNWINAPI cudnnDestroyCTCLossDescriptor(cudnnCTCLossDescriptor_t ctcLossDesc); /** * @brief Computes CTC loss and gradients given probabilities and labels. * * Labels and sequence lengths are in CPU memory. For GPU-memory variant, use cudnnCTCLoss_v8. * * @param[in] handle cuDNN handle. * @param[in] probsDesc Tensor descriptor for probabilities (T x N x A). * @param[in] probs Probabilities after softmax (GPU memory). * @param[in] hostLabels Labels (CPU memory). * @param[in] hostLabelLengths Length of each label (CPU memory). * @param[in] hostInputLengths Timing step lengths per batch (CPU memory). * @param[out] costs CTC costs (GPU memory). * @param[in] gradientsDesc Tensor descriptor for gradients (T x N x A). * @param[out] gradients CTC gradients (GPU memory, NULL for costs only). * @param[in] algo CTC loss algorithm. * @param[in] ctcLossDesc CTC loss descriptor. * @param[in] workspace Workspace (GPU memory). * @param[in] workSpaceSizeInBytes Workspace size. * * @retval CUDNN_STATUS_SUCCESS CTC loss computed. * * @since cuDNN 9.0.0 * @see cudnnCTCLoss_v8, cudnnGetCTCLossWorkspaceSize */ cudnnStatus_t CUDNNWINAPI cudnnCTCLoss( cudnnHandle_t handle, const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the mini batch size, A is the alphabet size) */ const void *probs, /* probabilities after softmax, in GPU memory */ const int hostLabels[], /* labels, in CPU memory */ const int hostLabelLengths[], /* the length of each label, in CPU memory */ const int hostInputLengths[], /* the lengths of timing steps in each batch, in CPU memory */ void *costs, /* the returned costs of CTC, in GPU memory */ const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A */ void *gradients, /* the returned CTC gradients, in GPU memory, to compute costs only, set it to NULL */ cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */ cudnnCTCLossDescriptor_t ctcLossDesc, void *workspace, /* pointer to the workspace, in GPU memory */ size_t workSpaceSizeInBytes); /* size of the workspace */ /** * @brief Computes CTC loss and gradients (v8, supports CUDA graphs with GPU memory labels). * * Labels and sequence lengths are in GPU memory (unlike cudnnCTCLoss which uses CPU memory). * * @since cuDNN 9.0.0 * @see cudnnCTCLoss, cudnnGetCTCLossWorkspaceSize_v8 */ cudnnStatus_t CUDNNWINAPI cudnnCTCLoss_v8( cudnnHandle_t handle, cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */ cudnnCTCLossDescriptor_t ctcLossDesc, const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the mini batch size, A is the alphabet size) */ const void *probs, /* probabilities after softmax, in GPU memory */ const int labels[], /* labels, in GPU memory */ const int labelLengths[], /* the length of each label, in GPU memory */ const int inputLengths[], /* the lengths of timing steps in each batch, in GPU memory */ void *costs, /* the returned costs of CTC, in GPU memory */ const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A */ void *gradients, /* the returned CTC gradients, in GPU memory, to compute costs only, set it to NULL */ size_t workSpaceSizeInBytes, /* size of the workspace */ void *workspace); /* pointer to the workspace, in GPU memory */ /** * @brief Returns the GPU workspace size required for CTC loss computation. * @since cuDNN 9.0.0 * @see cudnnCTCLoss */ cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossWorkspaceSize( cudnnHandle_t handle, const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the mini batch size, A is the alphabet size) */ const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A. To compute costs only, set it to NULL */ const int *labels, /* labels, in CPU memory */ const int *labelLengths, /* the length of each label, in CPU memory */ const int *inputLengths, /* the lengths of timing steps in each batch, in CPU memory */ cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */ cudnnCTCLossDescriptor_t ctcLossDesc, size_t *sizeInBytes); /* pointer to the returned workspace size */ /** * @brief Returns the GPU workspace size required for CTC loss v8 computation. * @since cuDNN 9.0.0 * @see cudnnCTCLoss_v8 */ cudnnStatus_t CUDNNWINAPI cudnnGetCTCLossWorkspaceSize_v8( cudnnHandle_t handle, cudnnCTCLossAlgo_t algo, /* algorithm selected, supported now 0 and 1 */ cudnnCTCLossDescriptor_t ctcLossDesc, const cudnnTensorDescriptor_t probsDesc, /* Tensor descriptor for probabilities, the dimensions are T,N,A (T is the timing steps, N is the mini batch size, A is the alphabet size) */ const cudnnTensorDescriptor_t gradientsDesc, /* Tensor descriptor for gradients, the dimensions are T,N,A. To compute costs only, set it to NULL */ size_t *sizeInBytes); /* pointer to the returned workspace size */ #if defined(__cplusplus) } #endif #endif /* CUDNN_ADV_H_ */