Files

T

Evgeny Mankov b72909055d [HIPIFY][DNN] Support of cuDNN 7.6.0

+ Update docs and README.md accordingly

2019-05-27 21:44:53 +03:00

72 KiB

Исходник Ответственный История

CUDNN API supported by HIP

1. CUDNN Data types

type	CUDA	HIP	HIP value (if differs)
define	`CUDNN_VERSION`	`HIPDNN_VERSION`
struct	`cudnnContext`
struct*	`cudnnHandle_t`	`hipdnnHandle_t`
enum	`cudnnStatus_t`	`hipdnnStatus_t`
0	`CUDNN_STATUS_SUCCESS`	`HIPDNN_STATUS_SUCCESS`
1	`CUDNN_STATUS_NOT_INITIALIZED`	`HIPDNN_STATUS_NOT_INITIALIZED`
2	`CUDNN_STATUS_ALLOC_FAILED`	`HIPDNN_STATUS_ALLOC_FAILED`
3	`CUDNN_STATUS_BAD_PARAM`	`HIPDNN_STATUS_BAD_PARAM`
4	`CUDNN_STATUS_INTERNAL_ERROR`	`HIPDNN_STATUS_INTERNAL_ERROR`
5	`CUDNN_STATUS_INVALID_VALUE`	`HIPDNN_STATUS_INVALID_VALUE`
6	`CUDNN_STATUS_ARCH_MISMATCH`	`HIPDNN_STATUS_ARCH_MISMATCH`
7	`CUDNN_STATUS_MAPPING_ERROR`	`HIPDNN_STATUS_MAPPING_ERROR`
8	`CUDNN_STATUS_EXECUTION_FAILED`	`HIPDNN_STATUS_EXECUTION_FAILED`
9	`CUDNN_STATUS_NOT_SUPPORTED`	`HIPDNN_STATUS_NOT_SUPPORTED`
10	`CUDNN_STATUS_LICENSE_ERROR`	`HIPDNN_STATUS_LICENSE_ERROR`
11	`CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING`	`HIPDNN_STATUS_RUNTIME_PREREQUISITE_MISSING`
12	`CUDNN_STATUS_RUNTIME_IN_PROGRESS`
13	`CUDNN_STATUS_RUNTIME_FP_OVERFLOW`
struct	`cudnnRuntimeTag_t`
enum	`cudnnErrQueryMode_t`
0	`CUDNN_ERRQUERY_RAWCODE`
1	`CUDNN_ERRQUERY_NONBLOCKING`
2	`CUDNN_ERRQUERY_BLOCKING`
struct	`cudnnTensorStruct`
struct*	`cudnnTensorDescriptor_t`	`hipdnnTensorDescriptor_t`
struct	`cudnnConvolutionStruct`
struct*	`cudnnConvolutionDescriptor_t`	`hipdnnConvolutionDescriptor_t`
struct	`cudnnPoolingStruct`
struct*	`cudnnPoolingDescriptor_t`	`hipdnnPoolingDescriptor_t`
struct	`cudnnFilterStruct`
struct*	`cudnnFilterDescriptor_t`	`hipdnnFilterDescriptor_t`
struct	`cudnnLRNStruct`
struct*	`cudnnLRNDescriptor_t`	`hipdnnLRNDescriptor_t`
struct	`cudnnActivationStruct`
struct*	`cudnnActivationDescriptor_t`	`hipdnnActivationDescriptor_t`
struct	`cudnnSpatialTransformerStruct`
struct*	`cudnnSpatialTransformerDescriptor_t`
struct	`cudnnOpTensorStruct`
struct*	`cudnnOpTensorDescriptor_t`	`hipdnnOpTensorDescriptor_t`
struct	`cudnnReduceTensorStruct`
struct*	`cudnnReduceTensorDescriptor_t`	`hipdnnReduceTensorDescriptor_t`
struct	`cudnnCTCLossStruct`
struct*	`cudnnCTCLossDescriptor_t`
struct	`cudnnTensorTransformStruct`
struct*	`cudnnTensorTransformDescriptor_t`
enum	`cudnnDataType_t`	`hipdnnDataType_t`
0	`CUDNN_DATA_FLOAT`	`HIPDNN_DATA_FLOAT`
1	`CUDNN_DATA_DOUBLE`	`HIPDNN_DATA_DOUBLE`
2	`CUDNN_DATA_HALF`	`HIPDNN_DATA_HALF`
3	`CUDNN_DATA_INT8`	`HIPDNN_DATA_INT8`
4	`CUDNN_DATA_INT32`	`HIPDNN_DATA_INT32`
5	`CUDNN_DATA_INT8x4`	`HIPDNN_DATA_INT8x4`
6	`CUDNN_DATA_UINT8`
7	`CUDNN_DATA_UINT8x4`
8	`CUDNN_DATA_INT8x32`
enum	`cudnnMathType_t`	`hipdnnMathType_t`
0	`CUDNN_DEFAULT_MATH`	`HIPDNN_DEFAULT_MATH`
1	`CUDNN_TENSOR_OP_MATH`	`HIPDNN_TENSOR_OP_MATH`
enum	`cudnnNanPropagation_t`	`hipdnnNanPropagation_t`
0	`CUDNN_NOT_PROPAGATE_NAN`	`HIPDNN_NOT_PROPAGATE_NAN`
1	`CUDNN_PROPAGATE_NAN`	`HIPDNN_PROPAGATE_NAN`
enum	`cudnnDeterminism_t`
0	`CUDNN_NON_DETERMINISTIC`
1	`CUDNN_DETERMINISTIC`
define	`CUDNN_DIM_MAX`
enum	`cudnnTensorFormat_t`	`hipdnnTensorFormat_t`
0	`CUDNN_TENSOR_NCHW`	`HIPDNN_TENSOR_NCHW`
1	`CUDNN_TENSOR_NHWC`	`HIPDNN_TENSOR_NHWC`
2	`CUDNN_TENSOR_NCHW_VECT_C`	`HIPDNN_TENSOR_NCHW_VECT_C`
enum	`cudnnFoldingDirection_t`
0	`CUDNN_TRANSFORM_FOLD`
1	`CUDNN_TRANSFORM_UNFOLD`
enum	`cudnnOpTensorOp_t`	`hipdnnOpTensorOp_t`
0	`CUDNN_OP_TENSOR_ADD`	`HIPDNN_OP_TENSOR_ADD`
1	`CUDNN_OP_TENSOR_MUL`	`HIPDNN_OP_TENSOR_MUL`
2	`CUDNN_OP_TENSOR_MIN`	`HIPDNN_OP_TENSOR_MIN`
3	`CUDNN_OP_TENSOR_MAX`	`HIPDNN_OP_TENSOR_MAX`
4	`CUDNN_OP_TENSOR_SQRT`	`HIPDNN_OP_TENSOR_SQRT`
5	`CUDNN_OP_TENSOR_NOT`	`HIPDNN_OP_TENSOR_NOT`
enum	`cudnnReduceTensorOp_t`	`hipdnnReduceTensorOp_t`
0	`CUDNN_REDUCE_TENSOR_ADD`	`HIPDNN_REDUCE_TENSOR_ADD`
1	`CUDNN_REDUCE_TENSOR_MUL`	`HIPDNN_REDUCE_TENSOR_MUL`
2	`CUDNN_REDUCE_TENSOR_MIN`	`HIPDNN_REDUCE_TENSOR_MIN`
3	`CUDNN_REDUCE_TENSOR_MAX`	`HIPDNN_REDUCE_TENSOR_MAX`
4	`CUDNN_REDUCE_TENSOR_AMAX`	`HIPDNN_REDUCE_TENSOR_AMAX`
5	`CUDNN_REDUCE_TENSOR_AVG`	`HIPDNN_REDUCE_TENSOR_AVG`
6	`CUDNN_REDUCE_TENSOR_NORM1`	`HIPDNN_REDUCE_TENSOR_NORM1`
7	`CUDNN_REDUCE_TENSOR_NORM2`	`HIPDNN_REDUCE_TENSOR_NORM2`
8	`CUDNN_REDUCE_TENSOR_MUL_NO_ZEROS`	`HIPDNN_REDUCE_TENSOR_MUL_NO_ZEROS`
enum	`cudnnReduceTensorIndices_t`	`hipdnnReduceTensorIndices_t`
0	`CUDNN_REDUCE_TENSOR_NO_INDICES`	`HIPDNN_REDUCE_TENSOR_NO_INDICES`
1	`CUDNN_REDUCE_TENSOR_FLATTENED_INDICES`	`HIPDNN_REDUCE_TENSOR_FLATTENED_INDICES`
enum	`cudnnIndicesType_t`	`hipdnnIndicesType_t`
0	`CUDNN_32BIT_INDICES`	`HIPDNN_32BIT_INDICES`
1	`CUDNN_64BIT_INDICES`	`HIPDNN_64BIT_INDICES`
2	`CUDNN_16BIT_INDICES`	`HIPDNN_16BIT_INDICES`
3	`CUDNN_8BIT_INDICES`	`HIPDNN_8BIT_INDICES`
enum	`cudnnConvolutionMode_t`	`hipdnnConvolutionMode_t`
0	`CUDNN_CONVOLUTION`	`HIPDNN_CONVOLUTION`
1	`CUDNN_CROSS_CORRELATION`	`HIPDNN_CROSS_CORRELATION`
enum	`cudnnConvolutionFwdPreference_t`	`hipdnnConvolutionFwdPreference_t`
0	`CUDNN_CONVOLUTION_FWD_NO_WORKSPACE`	`HIPDNN_CONVOLUTION_FWD_NO_WORKSPACE`
1	`CUDNN_CONVOLUTION_FWD_PREFER_FASTEST`	`HIPDNN_CONVOLUTION_FWD_PREFER_FASTEST`
2	`CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT`	`HIPDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT`
enum	`cudnnConvolutionFwdAlgo_t`	`hipdnnConvolutionFwdAlgo_t`
0	`CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM`	`HIPDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM`
1	`CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM`	`HIPDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM`
2	`CUDNN_CONVOLUTION_FWD_ALGO_GEMM`	`HIPDNN_CONVOLUTION_FWD_ALGO_GEMM`
3	`CUDNN_CONVOLUTION_FWD_ALGO_DIRECT`	`HIPDNN_CONVOLUTION_FWD_ALGO_DIRECT`
4	`CUDNN_CONVOLUTION_FWD_ALGO_FFT`	`HIPDNN_CONVOLUTION_FWD_ALGO_FFT`
5	`CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING`	`HIPDNN_CONVOLUTION_FWD_ALGO_FFT_TILING`
6	`CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD`	`HIPDNN_CONVOLUTION_FWD_ALGO_WINOGRAD`
7	`CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED`	`HIPDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED`
8	`CUDNN_CONVOLUTION_FWD_ALGO_COUNT`	`HIPDNN_CONVOLUTION_FWD_ALGO_COUNT`
struct	`cudnnConvolutionFwdAlgoPerf_t`	`hipdnnConvolutionFwdAlgoPerf_t`
enum	`cudnnConvolutionBwdFilterPreference_t`	`hipdnnConvolutionBwdFilterPreference_t`
0	`CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE`	`HIPDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE`
1	`CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST`	`HIPDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST`
2	`CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT`	`HIPDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT`
enum	`cudnnConvolutionBwdFilterAlgo_t`	`hipdnnConvolutionBwdFilterAlgo_t`
0	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_0`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_0`
1	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_1`
2	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT`
3	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_3`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_3`
4	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD`
5	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED`
6	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT_TILING`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_FFT_TILING`
7	`CUDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT`	`HIPDNN_CONVOLUTION_BWD_FILTER_ALGO_COUNT`
struct	`cudnnConvolutionBwdDataAlgoPerf_t`	`hipdnnConvolutionBwdDataAlgoPerf_t`
enum	`cudnnSoftmaxAlgorithm_t`	`hipdnnSoftmaxAlgorithm_t`
0	`CUDNN_SOFTMAX_FAST`	`HIPDNN_SOFTMAX_FAST`
1	`CUDNN_SOFTMAX_ACCURATE`	`HIPDNN_SOFTMAX_ACCURATE`
2	`CUDNN_SOFTMAX_LOG`	`HIPDNN_SOFTMAX_LOG`
enum	`cudnnSoftmaxMode_t`	`hipdnnSoftmaxMode_t`
0	`CUDNN_SOFTMAX_MODE_INSTANCE`	`HIPDNN_SOFTMAX_MODE_INSTANCE`
1	`CUDNN_SOFTMAX_MODE_CHANNEL`	`HIPDNN_SOFTMAX_MODE_CHANNEL`
enum	`cudnnPoolingMode_t`	`hipdnnPoolingMode_t`
0	`CUDNN_POOLING_MAX`	`HIPDNN_POOLING_MAX`
1	`CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING`	`HIPDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING`
2	`CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING`	`HIPDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING`
3	`CUDNN_POOLING_MAX_DETERMINISTIC`	`HIPDNN_POOLING_MAX_DETERMINISTIC`
enum	`cudnnActivationMode_t`	`hipdnnActivationMode_t`
0	`CUDNN_ACTIVATION_SIGMOID`	`HIPDNN_ACTIVATION_SIGMOID`
1	`CUDNN_ACTIVATION_RELU`	`HIPDNN_ACTIVATION_RELU`
2	`CUDNN_ACTIVATION_TANH`	`HIPDNN_ACTIVATION_TANH`
3	`CUDNN_ACTIVATION_CLIPPED_RELU`	`HIPDNN_ACTIVATION_CLIPPED_RELU`
4	`CUDNN_ACTIVATION_ELU`	`HIPDNN_ACTIVATION_ELU`
5	`CUDNN_ACTIVATION_IDENTITY`	`HIPDNN_ACTIVATION_PATHTRU`
define	`CUDNN_LRN_MIN_N`
define	`CUDNN_LRN_MAX_N`
define	`CUDNN_LRN_MIN_K`
define	`CUDNN_LRN_MIN_BETA`
enum	`cudnnLRNMode_t`	`hipdnnLRNMode_t`
0	`CUDNN_LRN_CROSS_CHANNEL_DIM1`	`HIPDNN_LRN_CROSS_CHANNEL`
enum	`cudnnDivNormMode_t`
0	`CUDNN_DIVNORM_PRECOMPUTED_MEANS`
enum	`cudnnBatchNormMode_t`	`hipdnnBatchNormMode_t`
0	`CUDNN_BATCHNORM_PER_ACTIVATION`	`HIPDNN_BATCHNORM_PER_ACTIVATION`
1	`CUDNN_BATCHNORM_SPATIAL`	`HIPDNN_BATCHNORM_SPATIAL`
2	`CUDNN_BATCHNORM_SPATIAL_PERSISTENT`	`HIPDNN_BATCHNORM_SPATIAL_PERSISTENT`
define	`CUDNN_BN_MIN_EPSILON`	`HIPDNN_BN_MIN_EPSILON`
enum	`cudnnSamplerType_t`
0	`CUDNN_SAMPLER_BILINEAR`
struct	`cudnnDropoutStruct`
struct*	`cudnnDropoutDescriptor_t`	`hipdnnDropoutDescriptor_t`
enum	`cudnnRNNMode_t`	`hipdnnRNNMode_t`
0	`CUDNN_RNN_RELU`	`HIPDNN_RNN_RELU`
1	`CUDNN_RNN_TANH`	`HIPDNN_RNN_TANH`
2	`CUDNN_LSTM`	`HIPDNN_LSTM`
3	`CUDNN_GRU`	`HIPDNN_GRU`
enum	`cudnnRNNBiasMode_t`	`hipdnnRNNBiasMode_t`
0	`CUDNN_RNN_NO_BIAS`	`HIPDNN_RNN_NO_BIAS`
1	`CUDNN_RNN_SINGLE_INP_BIAS`	`HIPDNN_RNN_WITH_BIAS`
2	`CUDNN_RNN_DOUBLE_BIAS`	`HIPDNN_RNN_WITH_BIAS`	1
3	`CUDNN_RNN_SINGLE_REC_BIAS`	`HIPDNN_RNN_WITH_BIAS`	1
enum	`cudnnDirectionMode_t`	`hipdnnDirectionMode_t`
0	`CUDNN_UNIDIRECTIONAL`	`HIPDNN_UNIDIRECTIONAL`
1	`CUDNN_BIDIRECTIONAL`	`HIPDNN_BIDIRECTIONAL`
enum	`cudnnRNNAlgo_t`	`hipdnnRNNAlgo_t`
0	`CUDNN_RNN_ALGO_STANDARD`	`HIPDNN_RNN_ALGO_STANDARD`
1	`CUDNN_RNN_ALGO_PERSIST_STATIC`	`HIPDNN_RNN_ALGO_PERSIST_STATIC`
2	`CUDNN_RNN_ALGO_PERSIST_DYNAMIC`	`HIPDNN_RNN_ALGO_PERSIST_DYNAMIC`
3	`CUDNN_RNN_ALGO_COUNT`
struct	`cudnnAlgorithmStruct`
struct*	`cudnnAlgorithmDescriptor_t`
struct	`cudnnAlgorithmPerformanceStruct`
struct*	`cudnnAlgorithmPerformance_t`
struct	`cudnnRNNStruct`
struct*	`cudnnRNNDescriptor_t`	`hipdnnRNNDescriptor_t`
struct	`cudnnPersistentRNNPlan`
struct*	`cudnnPersistentRNNPlan_t`	`hipdnnPersistentRNNPlan_t`
enum	`cudnnCTCLossAlgo_t`
0	`CUDNN_CTC_LOSS_ALGO_DETERMINISTIC`
1	`CUDNN_CTC_LOSS_ALGO_NON_DETERMINISTIC`
struct	`cudnnAlgorithm_t`
enum	`cudnnSeverity_t`
0	`CUDNN_SEV_FATAL`
1	`CUDNN_SEV_ERROR`
2	`CUDNN_SEV_WARNING`
3	`CUDNN_SEV_INFO`
define	`CUDNN_SEV_ERROR_EN`
define	`CUDNN_SEV_WARNING_EN`
define	`CUDNN_SEV_INFO_EN`
struct	`cudnnDebug_t`
struct	`cudnnCallback_t`
enum	`cudnnBatchNormOps_t`
0	`CUDNN_BATCHNORM_OPS_BN`
1	`CUDNN_BATCHNORM_OPS_BN_ACTIVATION`
2	`CUDNN_BATCHNORM_OPS_BN_ADD_ACTIVATION`
enum	`cudnnRNNClipMode_t`
0	`CUDNN_RNN_CLIP_NONE`
1	`CUDNN_RNN_CLIP_MINMAX`
struct	`cudnnRNNDataStruct`
struct*	`cudnnRNNDataDescriptor_t`
enum	`cudnnRNNDataLayout_t`
0	`CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_UNPACKED`
1	`CUDNN_RNN_DATA_LAYOUT_SEQ_MAJOR_PACKED`
2	`CUDNN_RNN_DATA_LAYOUT_BATCH_MAJOR_UNPACKED`
enum	`cudnnRNNPaddingMode_t`
0	`CUDNN_RNN_PADDED_IO_DISABLED`
1	`CUDNN_RNN_PADDED_IO_ENABLED`
enum	`cudnnSeqDataAxis_t`
0	`CUDNN_SEQDATA_TIME_DIM`
1	`CUDNN_SEQDATA_BATCH_DIM`
2	`CUDNN_SEQDATA_BEAM_DIM`
3	`CUDNN_SEQDATA_VECT_DIM`
define	`CUDNN_SEQDATA_DIM_COUNT`
struct	`cudnnSeqDataStruct`
struct*	`cudnnSeqDataDescriptor_t`
enum	`cudnnAttnQueryMap_t`
0	`CUDNN_ATTN_QUERYMAP_ALL_TO_ONE`
1	`CUDNN_ATTN_QUERYMAP_ONE_TO_ONE`
struct	`cudnnAttnStruct`
struct*	`cudnnAttnDescriptor_t`
enum	`cudnnMultiHeadAttnWeightKind_t`
0	`CUDNN_MH_ATTN_Q_WEIGHTS`
1	`CUDNN_MH_ATTN_K_WEIGHTS`
2	`CUDNN_MH_ATTN_V_WEIGHTS`
3	`CUDNN_MH_ATTN_O_WEIGHTS`
enum	`cudnnWgradMode_t`
0	`CUDNN_WGRAD_MODE_ADD`
1	`CUDNN_WGRAD_MODE_SET`
enum	`cudnnReorderType_t`
0	`CUDNN_DEFAULT_REORDER`
1	`CUDNN_NO_REORDER`
enum	`cudnnLossNormalizationMode_t`
0	`CUDNN_LOSS_NORMALIZATION_NONE`
1	`CUDNN_LOSS_NORMALIZATION_SOFTMAX`
struct	`cudnnFusedOpsConstParamStruct`
struct*	`cudnnFusedOpsConstParamPack_t`
struct	`cudnnFusedOpsVariantParamStruct`
struct*	`cudnnFusedOpsVariantParamPack_t`
struct	`cudnnFusedOpsPlanStruct`
struct*	`cudnnFusedOpsPlan_t`
enum	`cudnnFusedOps_t`
0	`CUDNN_FUSED_SCALE_BIAS_ACTIVATION_CONV_BNSTATS`
1	`CUDNN_FUSED_SCALE_BIAS_ACTIVATION_WGRAD`
2	`CUDNN_FUSED_BN_FINALIZE_STATISTICS_TRAINING`
3	`CUDNN_FUSED_BN_FINALIZE_STATISTICS_INFERENCE`
4	`CUDNN_FUSED_CONV_SCALE_BIAS_ADD_ACTIVATION`
5	`CUDNN_FUSED_SCALE_BIAS_ADD_ACTIVATION_GEN_BITMASK`
6	`CUDNN_FUSED_DACTIVATION_FORK_DBATCHNORM`
enum	`cudnnFusedOpsConstParamLabel_t`
0	`CUDNN_PARAM_XDESC`
1	`CUDNN_PARAM_XDATA_PLACEHOLDER`
2	`CUDNN_PARAM_BN_MODE`
3	`CUDNN_PARAM_BN_EQSCALEBIAS_DESC`
4	`CUDNN_PARAM_BN_EQSCALE_PLACEHOLDER`
5	`CUDNN_PARAM_BN_EQBIAS_PLACEHOLDER`
6	`CUDNN_PARAM_ACTIVATION_DESC`
7	`CUDNN_PARAM_CONV_DESC`
8	`CUDNN_PARAM_WDESC`
9	`CUDNN_PARAM_WDATA_PLACEHOLDER`
10	`CUDNN_PARAM_DWDESC`
11	`CUDNN_PARAM_DWDATA_PLACEHOLDER`
12	`CUDNN_PARAM_YDESC`
13	`CUDNN_PARAM_YDATA_PLACEHOLDER`
14	`CUDNN_PARAM_DYDESC`
15	`CUDNN_PARAM_DYDATA_PLACEHOLDER`
16	`CUDNN_PARAM_YSTATS_DESC`
17	`CUDNN_PARAM_YSUM_PLACEHOLDER`
18	`CUDNN_PARAM_YSQSUM_PLACEHOLDER`
19	`CUDNN_PARAM_BN_SCALEBIAS_MEANVAR_DESC`
20	`CUDNN_PARAM_BN_SCALE_PLACEHOLDER`
21	`CUDNN_PARAM_BN_BIAS_PLACEHOLDER`
22	`CUDNN_PARAM_BN_SAVED_MEAN_PLACEHOLDER`
23	`CUDNN_PARAM_BN_SAVED_INVSTD_PLACEHOLDER`
24	`CUDNN_PARAM_BN_RUNNING_MEAN_PLACEHOLDER`
25	`CUDNN_PARAM_BN_RUNNING_VAR_PLACEHOLDER`
26	`CUDNN_PARAM_ZDESC`
27	`CUDNN_PARAM_ZDATA_PLACEHOLDER`
28	`CUDNN_PARAM_BN_Z_EQSCALEBIAS_DESC`
29	`CUDNN_PARAM_BN_Z_EQSCALE_PLACEHOLDER`
30	`CUDNN_PARAM_BN_Z_EQBIAS_PLACEHOLDER`
31	`CUDNN_PARAM_ACTIVATION_BITMASK_DESC`
32	`CUDNN_PARAM_ACTIVATION_BITMASK_PLACEHOLDER`
33	`CUDNN_PARAM_DXDESC`
34	`CUDNN_PARAM_DXDATA_PLACEHOLDER`
35	`CUDNN_PARAM_DZDESC`
36	`CUDNN_PARAM_DZDATA_PLACEHOLDER`
37	`CUDNN_PARAM_BN_DSCALE_PLACEHOLDER`
38	`CUDNN_PARAM_BN_DBIAS_PLACEHOLDER`
enum	`cudnnFusedOpsPointerPlaceHolder_t`
0	`CUDNN_PTR_NULL`
1	`CUDNN_PTR_ELEM_ALIGNED`
2	`CUDNN_PTR_16B_ALIGNED`
enum	`cudnnFusedOpsVariantParamLabel_t`
0	`CUDNN_PTR_XDATA`
1	`CUDNN_PTR_BN_EQSCALE`
2	`CUDNN_PTR_BN_EQBIAS`
3	`CUDNN_PTR_WDATA`
4	`CUDNN_PTR_DWDATA`
5	`CUDNN_PTR_YDATA`
6	`CUDNN_PTR_DYDATA`
7	`CUDNN_PTR_YSUM`
8	`CUDNN_PTR_YSQSUM`
9	`CUDNN_PTR_WORKSPACE`
10	`CUDNN_PTR_BN_SCALE`
11	`CUDNN_PTR_BN_BIAS`
12	`CUDNN_PTR_BN_SAVED_MEAN`
13	`CUDNN_PTR_BN_SAVED_INVSTD`
14	`CUDNN_PTR_BN_RUNNING_MEAN`
15	`CUDNN_PTR_BN_RUNNING_VAR`
16	`CUDNN_PTR_ZDATA`
17	`CUDNN_PTR_BN_Z_EQSCALE`
18	`CUDNN_PTR_BN_Z_EQBIAS`
19	`CUDNN_PTR_ACTIVATION_BITMASK`
20	`CUDNN_PTR_DXDATA`
21	`CUDNN_PTR_DZDATA`
22	`CUDNN_PTR_BN_DSCALE`
23	`CUDNN_PTR_BN_DBIAS`
100	`CUDNN_SCALAR_SIZE_T_WORKSPACE_SIZE_IN_BYTES`
101	`CUDNN_SCALAR_INT64_T_BN_ACCUMULATION_COUNT`
102	`CUDNN_SCALAR_DOUBLE_BN_EXP_AVG_FACTOR`
103	`CUDNN_SCALAR_DOUBLE_BN_EPSILON`

2. CUDNN API functions

CUDA	HIP
`cudnnGetVersion`	`hipdnnGetVersion`
`cudnnGetCudartVersion`
`cudnnGetErrorString`	`hipdnnGetErrorString`
`cudnnQueryRuntimeError`
`cudnnGetProperty`
`cudnnCreate`	`hipdnnCreate`
`cudnnDestroy`	`hipdnnDestroy`
`cudnnSetStream`	`hipdnnSetStream`
`cudnnSetStream`	`hipdnnGetStream`
`cudnnCreateTensorDescriptor`	`hipdnnCreateTensorDescriptor`
`cudnnSetTensor4dDescriptor`	`hipdnnSetTensor4dDescriptor`
`cudnnSetTensor4dDescriptorEx`	`hipdnnSetTensor4dDescriptorEx`
`cudnnGetTensor4dDescriptor`	`hipdnnGetTensor4dDescriptor`
`cudnnSetTensorNdDescriptor`	`hipdnnSetTensorNdDescriptor`
`cudnnSetTensorNdDescriptorEx`
`cudnnGetTensorNdDescriptor`	`hipdnnGetTensorNdDescriptor`
`cudnnGetTensorSizeInBytes`
`cudnnDestroyTensorDescriptor`	`hipdnnDestroyTensorDescriptor`
`cudnnTransformTensor`
`cudnnTransformTensorEx`
`cudnnInitTransformDest`
`cudnnCreateTensorTransformDescriptor`
`cudnnSetTensorTransformDescriptor`
`cudnnGetTensorTransformDescriptor`
`cudnnDestroyTensorTransformDescriptor`
`cudnnAddTensor`	`hipdnnAddTensor`
`cudnnCreateOpTensorDescriptor`	`hipdnnCreateOpTensorDescriptor`
`cudnnSetOpTensorDescriptor`	`hipdnnSetOpTensorDescriptor`
`cudnnGetOpTensorDescriptor`	`hipdnnGetOpTensorDescriptor`
`cudnnDestroyOpTensorDescriptor`	`hipdnnDestroyOpTensorDescriptor`
`cudnnOpTensor`	`hipdnnOpTensor`
`cudnnGetFoldedConvBackwardDataDescriptors`
`cudnnCreateReduceTensorDescriptor`	`hipdnnCreateReduceTensorDescriptor`
`cudnnSetReduceTensorDescriptor`	`hipdnnSetReduceTensorDescriptor`
`cudnnGetReduceTensorDescriptor`	`hipdnnGetReduceTensorDescriptor`
`cudnnDestroyReduceTensorDescriptor`	`hipdnnDestroyReduceTensorDescriptor`
`cudnnGetReductionIndicesSize`
`cudnnGetReductionWorkspaceSize`	`hipdnnGetReductionWorkspaceSize`
`cudnnReduceTensor`	`hipdnnReduceTensor`
`cudnnSetTensor`	`hipdnnSetTensor`
`cudnnScaleTensor`	`hipdnnScaleTensor`
`cudnnCreateFilterDescriptor`	`hipdnnCreateFilterDescriptor`
`cudnnSetFilter4dDescriptor`	`hipdnnSetFilter4dDescriptor`
`cudnnGetFilter4dDescriptor`	`hipdnnGetFilter4dDescriptor`
`cudnnSetFilterNdDescriptor`	`hipdnnSetFilterNdDescriptor`
`cudnnGetFilterNdDescriptor`	`hipdnnGetFilterNdDescriptor`
`cudnnGetFilterSizeInBytes`
`cudnnTransformFilter`
`cudnnDestroyFilterDescriptor`	`hipdnnDestroyFilterDescriptor`
`cudnnReorderFilterAndBias`
`cudnnCreateConvolutionDescriptor`	`hipdnnCreateConvolutionDescriptor`
`cudnnSetConvolutionMathType`	`hipdnnSetConvolutionMathType`
`cudnnGetConvolutionMathType`
`cudnnSetConvolutionGroupCount`	`hipdnnSetConvolutionGroupCount`
`cudnnGetConvolutionGroupCount`
`cudnnSetConvolutionReorderType`
`cudnnGetConvolutionReorderType`
`cudnnSetConvolution2dDescriptor`	`hipdnnSetConvolution2dDescriptor`
`cudnnGetConvolution2dDescriptor`	`hipdnnGetConvolution2dDescriptor`
`cudnnGetConvolution2dForwardOutputDim`	`hipdnnGetConvolution2dForwardOutputDim`
`cudnnSetConvolutionNdDescriptor`	`hipdnnSetConvolutionNdDescriptor`
`cudnnGetConvolutionNdDescriptor`
`cudnnGetConvolutionNdForwardOutputDim`
`cudnnDestroyConvolutionDescriptor`
`cudnnGetConvolutionForwardAlgorithmMaxCount`
`cudnnFindConvolutionForwardAlgorithm`	`hipdnnFindConvolutionForwardAlgorithm`
`cudnnFindConvolutionForwardAlgorithmEx`	`hipdnnFindConvolutionForwardAlgorithmEx`
`cudnnGetConvolutionForwardAlgorithm`	`hipdnnGetConvolutionForwardAlgorithm`
`cudnnGetConvolutionForwardAlgorithm_v7`
`cudnnGetConvolutionForwardWorkspaceSize`	`hipdnnGetConvolutionForwardWorkspaceSize`
`cudnnConvolutionForward`	`hipdnnConvolutionForward`
`cudnnConvolutionBiasActivationForward`
`cudnnConvolutionBackwardBias`	`hipdnnConvolutionBackwardBias`
`cudnnGetConvolutionBackwardFilterAlgorithmMaxCount`
`cudnnFindConvolutionBackwardFilterAlgorithm`	`hipdnnFindConvolutionBackwardFilterAlgorithm`
`cudnnFindConvolutionBackwardFilterAlgorithmEx`	`hipdnnFindConvolutionBackwardFilterAlgorithmEx`
`cudnnGetConvolutionBackwardFilterAlgorithm`	`hipdnnGetConvolutionBackwardFilterAlgorithm`
`cudnnGetConvolutionBackwardFilterAlgorithm_v7`
`cudnnGetConvolutionBackwardFilterWorkspaceSize`	`hipdnnGetConvolutionBackwardFilterWorkspaceSize`
`cudnnConvolutionBackwardFilter`	`hipdnnConvolutionBackwardFilter`
`cudnnGetConvolutionBackwardDataAlgorithmMaxCount`
`cudnnFindConvolutionBackwardDataAlgorithm`	`hipdnnFindConvolutionBackwardDataAlgorithm`
`cudnnFindConvolutionBackwardDataAlgorithmEx`	`hipdnnFindConvolutionBackwardDataAlgorithmEx`
`cudnnGetConvolutionBackwardDataAlgorithm`	`hipdnnGetConvolutionBackwardDataAlgorithm`
`cudnnGetConvolutionBackwardDataAlgorithm_v7`
`cudnnGetConvolutionBackwardDataWorkspaceSize`	`hipdnnGetConvolutionBackwardDataWorkspaceSize`
`cudnnConvolutionBackwardData`	`hipdnnConvolutionBackwardData`
`cudnnIm2Col`
`cudnnSoftmaxForward`	`hipdnnSoftmaxForward`
`cudnnSoftmaxBackward`	`hipdnnSoftmaxBackward`
`cudnnCreatePoolingDescriptor`	`hipdnnCreatePoolingDescriptor`
`cudnnSetPooling2dDescriptor`	`hipdnnSetPooling2dDescriptor`
`cudnnGetPooling2dDescriptor`	`hipdnnGetPooling2dDescriptor`
`cudnnSetPoolingNdDescriptor`	`hipdnnSetPoolingNdDescriptor`
`cudnnGetPoolingNdDescriptor`
`cudnnGetPoolingNdForwardOutputDim`
`cudnnGetPooling2dForwardOutputDim`	`hipdnnGetPooling2dForwardOutputDim`
`cudnnDestroyPoolingDescriptor`	`hipdnnDestroyPoolingDescriptor`
`cudnnPoolingForward`	`hipdnnPoolingForward`
`cudnnPoolingBackward`	`hipdnnPoolingBackward`
`cudnnCreateActivationDescriptor`	`hipdnnCreateActivationDescriptor`
`cudnnSetActivationDescriptor`	`hipdnnSetActivationDescriptor`
`cudnnGetActivationDescriptor`	`hipdnnGetActivationDescriptor`
`cudnnDestroyActivationDescriptor`	`hipdnnDestroyActivationDescriptor`
`cudnnActivationForward`	`hipdnnActivationForward`
`cudnnActivationBackward`	`hipdnnActivationBackward`
`cudnnCreateLRNDescriptor`	`hipdnnCreateLRNDescriptor`
`cudnnSetLRNDescriptor`	`hipdnnSetLRNDescriptor`
`cudnnGetLRNDescriptor`	`hipdnnGetLRNDescriptor`
`cudnnDestroyLRNDescriptor`	`hipdnnDestroyLRNDescriptor`
`cudnnLRNCrossChannelForward`	`hipdnnLRNCrossChannelForward`
`cudnnLRNCrossChannelBackward`	`hipdnnLRNCrossChannelBackward`
`cudnnDivisiveNormalizationForward`
`cudnnDivisiveNormalizationBackward`
`cudnnDeriveBNTensorDescriptor`	`hipdnnDeriveBNTensorDescriptor`
`cudnnBatchNormalizationForwardTraining`	`hipdnnBatchNormalizationForwardTraining`
`cudnnBatchNormalizationForwardTrainingEx`
`cudnnBatchNormalizationForwardInference`	`hipdnnBatchNormalizationForwardInference`
`cudnnBatchNormalizationBackward`	`hipdnnBatchNormalizationBackward`
`cudnnBatchNormalizationBackwardEx`
`cudnnCreateSpatialTransformerDescriptor`
`cudnnSetSpatialTransformerNdDescriptor`
`cudnnDestroySpatialTransformerDescriptor`
`cudnnSpatialTfGridGeneratorForward`
`cudnnSpatialTfGridGeneratorBackward`
`cudnnSpatialTfSamplerForward`
`cudnnSpatialTfSamplerBackward`
`cudnnCreateDropoutDescriptor`	`hipdnnCreateDropoutDescriptor`
`cudnnDestroyDropoutDescriptor`	`hipdnnDestroyDropoutDescriptor`
`cudnnDropoutGetStatesSize`	`hipdnnDropoutGetStatesSize`
`cudnnDropoutGetReserveSpaceSize`
`cudnnSetDropoutDescriptor`	`hipdnnSetDropoutDescriptor`
`cudnnGetDropoutDescriptor`
`cudnnRestoreDropoutDescriptor`
`cudnnDropoutForward`
`cudnnDropoutBackward`
`cudnnCreateRNNDescriptor`	`hipdnnCreateRNNDescriptor`
`cudnnDestroyRNNDescriptor`	`hipdnnDestroyRNNDescriptor`
`cudnnGetRNNForwardInferenceAlgorithmMaxCount`
`cudnnFindRNNForwardInferenceAlgorithmEx`
`cudnnGetRNNForwardTrainingAlgorithmMaxCount`
`cudnnFindRNNForwardTrainingAlgorithmEx`
`cudnnGetRNNBackwardDataAlgorithmMaxCount`
`cudnnFindRNNBackwardDataAlgorithmEx`
`cudnnGetRNNBackwardWeightsAlgorithmMaxCount`
`cudnnFindRNNBackwardWeightsAlgorithmEx`
`cudnnCreatePersistentRNNPlan`	`hipdnnCreatePersistentRNNPlan`
`cudnnSetPersistentRNNPlan`	`hipdnnSetPersistentRNNPlan`
`cudnnDestroyPersistentRNNPlan`	`hipdnnDestroyPersistentRNNPlan`
`cudnnSetRNNDescriptor`	`hipdnnSetRNNDescriptor`
`cudnnGetRNNDescriptor`	`hipdnnGetRNNDescriptor`
`cudnnSetRNNProjectionLayers`
`cudnnGetRNNProjectionLayers`
`cudnnSetRNNAlgorithmDescriptor`
`cudnnSetRNNMatrixMathType`
`cudnnGetRNNMatrixMathType`
`cudnnGetRNNWorkspaceSize`	`hipdnnGetRNNWorkspaceSize`
`cudnnGetRNNTrainingReserveSize`	`hipdnnGetRNNTrainingReserveSize`
`cudnnGetRNNParamsSize`	`hipdnnGetRNNParamsSize`
`cudnnGetRNNLinLayerMatrixParams`	`hipdnnGetRNNLinLayerMatrixParams`
`cudnnGetRNNLinLayerBiasParams`	`hipdnnGetRNNLinLayerBiasParams`
`cudnnRNNForwardInference`	`hipdnnRNNForwardInference`
`cudnnRNNForwardInferenceEx`
`cudnnRNNForwardTraining`	`hipdnnRNNForwardTraining`
`cudnnRNNForwardTrainingEx`
`cudnnRNNBackwardData`	`hipdnnRNNBackwardData`
`cudnnRNNBackwardDataEx`
`cudnnRNNBackwardWeights`	`hipdnnRNNBackwardWeights`
`cudnnRNNBackwardWeightsEx`
`cudnnSetRNNPaddingMode`
`cudnnGetRNNPaddingMode`
`cudnnCreateRNNDataDescriptor`
`cudnnDestroyRNNDataDescriptor`
`cudnnSetRNNDataDescriptor`
`cudnnGetRNNDataDescriptor`
`cudnnSetRNNBiasMode`
`cudnnGetRNNBiasMode`
`cudnnCreateCTCLossDescriptor`
`cudnnSetCTCLossDescriptor`
`cudnnSetCTCLossDescriptorEx`
`cudnnGetCTCLossDescriptor`
`cudnnGetCTCLossDescriptorEx`
`cudnnDestroyCTCLossDescriptor`
`cudnnCTCLoss`
`cudnnGetCTCLossWorkspaceSize`
`cudnnCreateAlgorithmDescriptor`
`cudnnSetAlgorithmDescriptor`
`cudnnGetAlgorithmDescriptor`
`cudnnCopyAlgorithmDescriptor`
`cudnnDestroyAlgorithmDescriptor`
`cudnnCreateAlgorithmPerformance`
`cudnnSetAlgorithmPerformance`
`cudnnGetAlgorithmPerformance`
`cudnnDestroyAlgorithmPerformance`
`cudnnGetAlgorithmSpaceSize`
`cudnnSaveAlgorithm`
`cudnnRestoreAlgorithm`
`cudnnSetRNNDescriptor_v5`	`hipdnnSetRNNDescriptor_v5`
`cudnnSetRNNDescriptor_v6`	`hipdnnSetRNNDescriptor_v6`
`cudnnSetCallback`
`cudnnGetCallback`
`cudnnGetBatchNormalizationForwardTrainingExWorkspaceSize`
`cudnnGetBatchNormalizationBackwardExWorkspaceSize`
`cudnnGetBatchNormalizationTrainingExReserveSpaceSize`
`cudnnRNNSetClip`
`cudnnRNNGetClip`
`cudnnCreateSeqDataDescriptor`
`cudnnDestroySeqDataDescriptor`
`cudnnSetSeqDataDescriptor`
`cudnnGetSeqDataDescriptor`
`cudnnCreateAttnDescriptor`
`cudnnDestroyAttnDescriptor`
`cudnnSetAttnDescriptor`
`cudnnGetAttnDescriptor`
`cudnnGetMultiHeadAttnBuffers`
`cudnnGetMultiHeadAttnWeights`
`cudnnMultiHeadAttnForward`
`cudnnMultiHeadAttnBackwardData`
`cudnnMultiHeadAttnBackwardWeights`
`cudnnCreateFusedOpsConstParamPack`
`cudnnDestroyFusedOpsConstParamPack`
`cudnnSetFusedOpsConstParamPackAttribute`
`cudnnGetFusedOpsConstParamPackAttribute`
`cudnnCreateFusedOpsVariantParamPack`
`cudnnDestroyFusedOpsVariantParamPack`
`cudnnSetFusedOpsVariantParamPackAttribute`
`cudnnGetFusedOpsVariantParamPackAttribute`
`cudnnCreateFusedOpsPlan`
`cudnnDestroyFusedOpsPlan`
`cudnnMakeFusedOpsPlan`
`cudnnFusedOpsExecute`

72 KiB Исходник Ответственный История

CUDNN API supported by HIP

1. CUDNN Data types

2. CUDNN API functions

72 KiB

Исходник Ответственный История