Merge 'master' into 'amd-master'
Change-Id: Id78d37905036fb495b522b0a741301ee91fcf325
Этот коммит содержится в:
Jenkins
@@ -193,6 +193,18 @@ if(USE_PROF_API EQUAL 1)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
################
|
||||
# Detect Code Object Manager
|
||||
################
|
||||
|
||||
if(HIP_PLATFORM STREQUAL "hcc")
|
||||
find_package(amd_comgr REQUIRED CONFIG
|
||||
PATHS
|
||||
/opt/rocm/lib/cmake/amd_comgr
|
||||
)
|
||||
MESSAGE(STATUS "Code Object Manager found at ${amd_comgr_DIR}.")
|
||||
endif()
|
||||
|
||||
#############################
|
||||
# Build steps
|
||||
#############################
|
||||
@@ -264,6 +276,9 @@ if(HIP_PLATFORM STREQUAL "hcc")
|
||||
target_link_libraries(hip_hcc_static PRIVATE hc_am)
|
||||
endif()
|
||||
|
||||
target_link_libraries(hip_hcc PRIVATE amd_comgr)
|
||||
target_link_libraries(hip_hcc_static PRIVATE amd_comgr)
|
||||
|
||||
string(REPLACE " " ";" HCC_CXX_FLAGS_LIST ${HCC_CXX_FLAGS})
|
||||
foreach(TARGET hip_hcc hip_hcc_static)
|
||||
target_include_directories(${TARGET} SYSTEM INTERFACE $<INSTALL_INTERFACE:$<INSTALL_PREFIX>/include>;${HSA_PATH}/include)
|
||||
|
||||
поставляемый
+4
-4
@@ -296,13 +296,13 @@ def docker_upload_dockerhub( String local_org, String image_name, String remote_
|
||||
String build_config = 'Release'
|
||||
String job_name = env.JOB_NAME.toLowerCase( )
|
||||
|
||||
// The following launches 3 builds in parallel: rocm-head, rocm-2.1.x and cuda-10.x
|
||||
parallel rocm_2_1:
|
||||
// The following launches 3 builds in parallel: rocm-head, rocm-2.3.x and cuda-10.x
|
||||
parallel rocm_2_3:
|
||||
{
|
||||
node('hip-rocm')
|
||||
{
|
||||
String hcc_ver = 'rocm-2.1.x'
|
||||
String from_image = 'ci_test_nodes/rocm-2.1.x/ubuntu-16.04:latest'
|
||||
String hcc_ver = 'rocm-2.3.x'
|
||||
String from_image = 'ci_test_nodes/rocm-2.3.x/ubuntu-16.04:latest'
|
||||
String inside_args = '--device=/dev/kfd --device=/dev/dri --group-add=video'
|
||||
|
||||
// Checkout source code, dependencies and version files
|
||||
|
||||
@@ -265,9 +265,11 @@ if ($HIP_PLATFORM eq "clang") {
|
||||
|
||||
$HIPCXXFLAGS .= " -isystem $HIP_PATH/include/hip/hcc_detail/cuda";
|
||||
$HIPCXXFLAGS .= " -isystem $HSA_PATH/include";
|
||||
$HIPCXXFLAGS .= " -isystem /opt/rocm/include";
|
||||
$HIPCXXFLAGS .= " -Wno-deprecated-register";
|
||||
|
||||
$HIPLDFLAGS .= " -L$HSA_PATH/lib -L$ROCM_PATH/lib -lhsa-runtime64 -lhc_am ";
|
||||
$HIPLDFLAGS .= " -L/opt/rocm/lib -Wl,--rpath=/opt/rocm/lib -lamd_comgr ";
|
||||
# $HIPLDFLAGS .= " -L$HCC_HOME/compiler/lib -lLLVMAMDGPUDesc -lLLVMAMDGPUUtils -lLLVMMC -lLLVMCore -lLLVMSupport ";
|
||||
|
||||
# Add trace marker library:
|
||||
|
||||
@@ -20,9 +20,12 @@
|
||||
| 0 |*`CUBLAS_OP_N`* |*`HIPBLAS_OP_N`* | 111 |
|
||||
| 1 |*`CUBLAS_OP_T`* |*`HIPBLAS_OP_T`* | 112 |
|
||||
| 2 |*`CUBLAS_OP_C`* |*`HIPBLAS_OP_C`* | 113 |
|
||||
| 2 |*`CUBLAS_OP_HERMITAN`* |*`HIPBLAS_OP_C`* | 113 |
|
||||
| 3 |*`CUBLAS_OP_CONJG`* | |
|
||||
| enum |***`cublasFillMode_t`*** |***`hipblasFillMode_t`*** |
|
||||
| 0 |*`CUBLAS_FILL_MODE_LOWER`* |*`HIPBLAS_FILL_MODE_LOWER`* | 121 |
|
||||
| 1 |*`CUBLAS_FILL_MODE_UPPER`* |*`HIPBLAS_FILL_MODE_UPPER`* | 122 |
|
||||
| 1 |*`CUBLAS_FILL_MODE_FULL`* |*`HIPBLAS_FILL_MODE_FULL`* | 123 |
|
||||
| enum |***`cublasDiagType_t`*** |***`hipblasDiagType_t`*** |
|
||||
| 0 |*`CUBLAS_DIAG_NON_UNIT`* |*`HIPBLAS_DIAG_NON_UNIT`* | 131 |
|
||||
| 1 |*`CUBLAS_DIAG_UNIT`* |*`HIPBLAS_DIAG_UNIT`* | 132 |
|
||||
@@ -98,6 +101,7 @@
|
||||
|`cublasGetVersion` | |
|
||||
|`cublasGetVersion_v2` | |
|
||||
|`cublasGetProperty` | |
|
||||
|`cublasGetCudartVersion` | |
|
||||
|`cublasGetStream` |`hipblasGetStream` |
|
||||
|`cublasGetStream_v2` |`hipblasGetStream` |
|
||||
|`cublasSetStream` |`hipblasSetStream` |
|
||||
@@ -173,6 +177,7 @@
|
||||
|`cublasDcopy` |`hipblasDcopy` |
|
||||
|`cublasDcopy_v2` |`hipblasDcopy` |
|
||||
|`cublasCcopy` | |
|
||||
|`cublasCopyEx` | |
|
||||
|`cublasCcopy_v2` | |
|
||||
|`cublasZcopy` | |
|
||||
|`cublasZcopy_v2` | |
|
||||
@@ -184,6 +189,7 @@
|
||||
|`cublasCswap_v2` | |
|
||||
|`cublasZswap` | |
|
||||
|`cublasZswap_v2` | |
|
||||
|`cublasIamaxEx` | |
|
||||
|`cublasIsamax` |`hipblasIsamax` |
|
||||
|`cublasIsamax_v2` |`hipblasIsamax` |
|
||||
|`cublasIdamax` |`hipblasIdamax` |
|
||||
@@ -192,6 +198,7 @@
|
||||
|`cublasIcamax_v2` | |
|
||||
|`cublasIzamax` | |
|
||||
|`cublasIzamax_v2` | |
|
||||
|`cublasIaminEx` | |
|
||||
|`cublasIsamin` | |
|
||||
|`cublasIsamin_v2` | |
|
||||
|`cublasIdamin` | |
|
||||
@@ -200,6 +207,7 @@
|
||||
|`cublasIcamin_v2` | |
|
||||
|`cublasIzamin` | |
|
||||
|`cublasIzamin_v2` | |
|
||||
|`cublasAsumEx` | |
|
||||
|`cublasSasum` |`hipblasSasum` |
|
||||
|`cublasSasum_v2` |`hipblasSasum` |
|
||||
|`cublasDasum` |`hipblasDasum` |
|
||||
@@ -208,6 +216,7 @@
|
||||
|`cublasScasum_v2` | |
|
||||
|`cublasDzasum` | |
|
||||
|`cublasDzasum_v2` | |
|
||||
|`cublasRotEx` | |
|
||||
|`cublasSrot` | |
|
||||
|`cublasSrot_v2` | |
|
||||
|`cublasDrot` | |
|
||||
@@ -216,6 +225,7 @@
|
||||
|`cublasCrot_v2` | |
|
||||
|`cublasZrot` | |
|
||||
|`cublasZrot_v2` | |
|
||||
|`cublasRotgEx` | |
|
||||
|`cublasZdrot` | |
|
||||
|`cublasZdrot_v2` | |
|
||||
|`cublasSrotg` | |
|
||||
@@ -226,10 +236,12 @@
|
||||
|`cublasCrotg_v2` | |
|
||||
|`cublasZrotg` | |
|
||||
|`cublasZrotg_v2` | |
|
||||
|`cublasRotmEx` | |
|
||||
|`cublasSrotm` | |
|
||||
|`cublasSrotm_v2` | |
|
||||
|`cublasDrotm` | |
|
||||
|`cublasDrotm_v2` | |
|
||||
|`cublasRotmgEx` | |
|
||||
|`cublasSrotmg` | |
|
||||
|`cublasSrotmg_v2` | |
|
||||
|`cublasDrotmg` | |
|
||||
|
||||
@@ -2,8 +2,8 @@
|
||||
|
||||
## **1. CUDNN Data types**
|
||||
|
||||
| **type** | **CUDA** | **HIP** |
|
||||
|-------------:|---------------------------------------------------------------|------------------------------------------------------------|
|
||||
| **type** | **CUDA** | **HIP** |**HIP value** (if differs) |
|
||||
|-------------:|---------------------------------------------------------------|------------------------------------------------------------|---------------------------|
|
||||
| define |`CUDNN_VERSION` |`HIPDNN_VERSION` |
|
||||
| struct |`cudnnContext` | |
|
||||
| struct* |`cudnnHandle_t` |`hipdnnHandle_t` |
|
||||
@@ -47,6 +47,8 @@
|
||||
| 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`* |
|
||||
@@ -71,6 +73,9 @@
|
||||
| 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`* |
|
||||
@@ -169,6 +174,11 @@
|
||||
| 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`* |
|
||||
@@ -215,6 +225,27 @@
|
||||
| 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`* | |
|
||||
|
||||
## **2. CUDNN API functions**
|
||||
|
||||
@@ -239,6 +270,12 @@
|
||||
|`cudnnGetTensorSizeInBytes` | |
|
||||
|`cudnnDestroyTensorDescriptor` |`hipdnnDestroyTensorDescriptor` |
|
||||
|`cudnnTransformTensor` | |
|
||||
|`cudnnTransformTensorEx` | |
|
||||
|`cudnnInitTransformDest` | |
|
||||
|`cudnnCreateTensorTransformDescriptor` | |
|
||||
|`cudnnSetTensorTransformDescriptor` | |
|
||||
|`cudnnGetTensorTransformDescriptor` | |
|
||||
|`cudnnDestroyTensorTransformDescriptor` | |
|
||||
|`cudnnAddTensor` |`hipdnnAddTensor` |
|
||||
|`cudnnCreateOpTensorDescriptor` |`hipdnnCreateOpTensorDescriptor` |
|
||||
|`cudnnSetOpTensorDescriptor` |`hipdnnSetOpTensorDescriptor` |
|
||||
@@ -383,6 +420,8 @@
|
||||
|`cudnnDestroyRNNDataDescriptor` | |
|
||||
|`cudnnSetRNNDataDescriptor` | |
|
||||
|`cudnnGetRNNDataDescriptor` | |
|
||||
|`cudnnSetRNNBiasMode` | |
|
||||
|`cudnnGetRNNBiasMode` | |
|
||||
|`cudnnCreateCTCLossDescriptor` | |
|
||||
|`cudnnSetCTCLossDescriptor` | |
|
||||
|`cudnnGetCTCLossDescriptor` | |
|
||||
@@ -410,3 +449,16 @@
|
||||
|`cudnnGetBatchNormalizationTrainingExReserveSpaceSize` | |
|
||||
|`cudnnRNNSetClip` | |
|
||||
|`cudnnRNNGetClip` | |
|
||||
|`cudnnCreateSeqDataDescriptor` | |
|
||||
|`cudnnDestroySeqDataDescriptor` | |
|
||||
|`cudnnSetSeqDataDescriptor` | |
|
||||
|`cudnnGetSeqDataDescriptor` | |
|
||||
|`cudnnCreateAttnDescriptor` | |
|
||||
|`cudnnDestroyAttnDescriptor` | |
|
||||
|`cudnnSetAttnDescriptor` | |
|
||||
|`cudnnGetAttnDescriptor` | |
|
||||
|`cudnnGetMultiHeadAttnBuffers` | |
|
||||
|`cudnnGetMultiHeadAttnWeights` | |
|
||||
|`cudnnMultiHeadAttnForward` | |
|
||||
|`cudnnMultiHeadAttnBackwardData` | |
|
||||
|`cudnnMultiHeadAttnBackwardWeights` | |
|
||||
|
||||
@@ -82,6 +82,27 @@
|
||||
| typedef |`cusparseColorInfo_t` | |
|
||||
| struct |`pruneInfo` | |
|
||||
| typedef |`pruneInfo_t` | |
|
||||
| enum |***`cusparseCsr2CscAlg_t`*** | |
|
||||
| 1 |*`CUSPARSE_CSR2CSC_ALG1`* | |
|
||||
| 2 |*`CUSPARSE_CSR2CSC_ALG2`* | |
|
||||
| enum |***`cusparseFormat_t`*** | |
|
||||
| 1 |*`CUSPARSE_FORMAT_CSR`* | |
|
||||
| 2 |*`CUSPARSE_FORMAT_CSC`* | |
|
||||
| 3 |*`CUSPARSE_FORMAT_COO`* | |
|
||||
| enum |***`cusparseOrder_t`*** | |
|
||||
| 1 |*`CUSPARSE_ORDER_COL`* | |
|
||||
| 2 |*`CUSPARSE_ORDER_ROW`* | |
|
||||
| enum |***`cusparseSpMMAlg_t`*** | |
|
||||
| 1 |*`CUSPARSE_COOMM_ALG1`* | |
|
||||
| 2 |*`CUSPARSE_COOMM_ALG2`* | |
|
||||
| 3 |*`CUSPARSE_COOMM_ALG3`* | |
|
||||
| enum |***`cusparseIndexType_t`*** | |
|
||||
| 1 |*`CUSPARSE_INDEX_16U`* | |
|
||||
| 2 |*`CUSPARSE_INDEX_32I`* | |
|
||||
| struct |`cusparseSpMatDescr` | |
|
||||
| typedef |`cusparseSpMatDescr_t` | |
|
||||
| struct |`cusparseDnMatDescr` | |
|
||||
| typedef |`cusparseDnMatDescr_t` | |
|
||||
|
||||
## **2. cuSPARSE Helper Function Reference**
|
||||
|
||||
@@ -530,6 +551,8 @@
|
||||
|`cusparseCcsr2csc` | |
|
||||
|`cusparseZcsr2csc` | |
|
||||
|`cusparseCsr2cscEx` | |
|
||||
|`cusparseCsr2cscEx2` | |
|
||||
|`cusparseCsr2cscEx2_bufferSize` | |
|
||||
|`cusparseScsr2dense` | |
|
||||
|`cusparseDcsr2dense` | |
|
||||
|`cusparseCcsr2dense` | |
|
||||
@@ -616,3 +639,34 @@
|
||||
|`cusparseDnnz_compress` | |
|
||||
|`cusparseCnnz_compress` | |
|
||||
|`cusparseZnnz_compress` | |
|
||||
|
||||
## **10. cuSPARSE Generic API Reference**
|
||||
|
||||
## ***10.1. Generic Sparse API helper functions***
|
||||
|
||||
| **CUDA** | **HIP** |
|
||||
|-----------------------------------------------------------|-------------------------------------------------|
|
||||
|`cusparseCreateCoo` | |
|
||||
|`cusparseDestroySpMat` | |
|
||||
|`cusparseCooGet` | |
|
||||
|`cusparseSpMatGetFormat` | |
|
||||
|`cusparseSpMatGetIndexBase` | |
|
||||
|`cusparseSpMatSetNumBatches` | |
|
||||
|`cusparseSpMatGetNumBatches` | |
|
||||
|
||||
## ***10.2. Generic Dense API helper functions***
|
||||
|
||||
| **CUDA** | **HIP** |
|
||||
|-----------------------------------------------------------|-------------------------------------------------|
|
||||
|`cusparseCreateDnMat` | |
|
||||
|`cusparseDestroyDnMat` | |
|
||||
|`cusparseDnMatGet` | |
|
||||
|`cusparseDnMatSetStridedBatch` | |
|
||||
|`cusparseDnMatGetStridedBatch` | |
|
||||
|
||||
## ***10.3. Generic SpMM API functions***
|
||||
|
||||
| **CUDA** | **HIP** |
|
||||
|-----------------------------------------------------------|-------------------------------------------------|
|
||||
|`cusparseSpMM` | |
|
||||
|`cusparseSpMM_bufferSize` | |
|
||||
|
||||
@@ -421,6 +421,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP{
|
||||
{"cublasGetStream_v2", {"hipblasGetStream", "rocblas_get_stream", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasGetPointerMode_v2", {"hipblasGetPointerMode", "rocblas_set_pointer_mode", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasSetPointerMode_v2", {"hipblasSetPointerMode", "rocblas_get_pointer_mode", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasGetCudartVersion", {"hipblasGetCudartVersion", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// GEMV
|
||||
{"cublasSgemv_v2", {"hipblasSgemv", "rocblas_sgemv", CONV_LIB_FUNC, API_BLAS}},
|
||||
@@ -631,36 +632,42 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP{
|
||||
{"cublasZaxpy_v2", {"hipblasZaxpy", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// COPY
|
||||
{"cublasCopyEx", {"hipblasCopyEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasScopy_v2", {"hipblasScopy", "rocblas_scopy", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasDcopy_v2", {"hipblasDcopy", "rocblas_dcopy", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasCcopy_v2", {"hipblasCcopy", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasZcopy_v2", {"hipblasZcopy", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// SWAP
|
||||
{"cublasSwapEx", {"hipblasSwapEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSswap_v2", {"hipblasSswap", "rocblas_sswap", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasDswap_v2", {"hipblasDswap", "rocblas_dswap", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasCswap_v2", {"hipblasCswap", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasZswap_v2", {"hipblasZswap", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// AMAX
|
||||
{"cublasIamaxEx", {"hipblasIamaxEx", "", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasIsamax_v2", {"hipblasIsamax", "rocblas_isamax", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasIdamax_v2", {"hipblasIdamax", "rocblas_idamax", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasIcamax_v2", {"hipblasIcamax", "", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasIzamax_v2", {"hipblasIzamax", "", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
|
||||
// AMIN
|
||||
{"cublasIaminEx", {"hipblasIaminEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasIsamin_v2", {"hipblasIsamin", "rocblas_isamin", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasIdamin_v2", {"hipblasIdamin", "rocblas_idamin", CONV_LIB_FUNC, API_BLAS, HIP_UNSUPPORTED}},
|
||||
{"cublasIcamin_v2", {"hipblasIcamin", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasIzamin_v2", {"hipblasIzamin", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// ASUM
|
||||
{"cublasAsumEx", {"hipblasAsumEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSasum_v2", {"hipblasSasum", "rocblas_sasum", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasDasum_v2", {"hipblasDasum", "rocblas_dasum", CONV_LIB_FUNC, API_BLAS}},
|
||||
{"cublasScasum_v2", {"hipblasScasum", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasDzasum_v2", {"hipblasDzasum", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// ROT
|
||||
{"cublasRotEx", {"hipblasRotEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSrot_v2", {"hipblasSrot", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasDrot_v2", {"hipblasDrot", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasCrot_v2", {"hipblasCrot", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
@@ -669,16 +676,19 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP{
|
||||
{"cublasZdrot_v2", {"hipblasZdrot", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// ROTG
|
||||
{"cublasRotgEx", {"hipblasRotgEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSrotg_v2", {"hipblasSrotg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasDrotg_v2", {"hipblasDrotg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasCrotg_v2", {"hipblasCrotg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasZrotg_v2", {"hipblasZrotg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// ROTM
|
||||
{"cublasRotmEx", {"hipblasRotmEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSrotm_v2", {"hipblasSrotm", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasDrotm_v2", {"hipblasDrotm", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// ROTMG
|
||||
{"cublasRotmgEx", {"hipblasRotmgEx", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasSrotmg_v2", {"hipblasSrotmg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
{"cublasDrotmg_v2", {"hipblasDrotmg", "", CONV_LIB_FUNC, API_BLAS, UNSUPPORTED}},
|
||||
};
|
||||
|
||||
@@ -29,6 +29,8 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_TYPE_NAME_MAP{
|
||||
{"CUBLAS_OP_N", {"HIPBLAS_OP_N", "rocblas_operation_none", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_OP_T", {"HIPBLAS_OP_T", "rocblas_operation_transpose", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_OP_C", {"HIPBLAS_OP_C", "rocblas_operation_conjugate_transpose", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_OP_HERMITAN", {"HIPBLAS_OP_C", "rocblas_operation_conjugate_transpose", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_OP_CONJG", {"HIPBLAS_OP_CONJG", "rocblas_operation_conjugate", CONV_NUMERIC_LITERAL, API_BLAS, UNSUPPORTED}},
|
||||
|
||||
// Blas statuses
|
||||
{"cublasStatus", {"hipblasStatus_t", "rocblas_status", CONV_TYPE, API_BLAS}},
|
||||
@@ -48,6 +50,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_TYPE_NAME_MAP{
|
||||
{"cublasFillMode_t", {"hipblasFillMode_t", "rocblas_fill", CONV_TYPE, API_BLAS}},
|
||||
{"CUBLAS_FILL_MODE_LOWER", {"HIPBLAS_FILL_MODE_LOWER", "rocblas_fill_lower", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_FILL_MODE_UPPER", {"HIPBLAS_FILL_MODE_UPPER", "rocblas_fill_upper", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
{"CUBLAS_FILL_MODE_FULL", {"HIPBLAS_FILL_MODE_FULL", "rocblas_fill_full", CONV_NUMERIC_LITERAL, API_BLAS}},
|
||||
|
||||
// Blas Diag Types
|
||||
{"cublasDiagType_t", {"hipblasDiagType_t", "rocblas_diagonal", CONV_TYPE, API_BLAS}},
|
||||
|
||||
@@ -49,6 +49,12 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_FUNCTION_MAP{
|
||||
{"cudnnGetTensorSizeInBytes", {"hipdnnGetTensorSizeInBytes", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnDestroyTensorDescriptor", {"hipdnnDestroyTensorDescriptor", "", CONV_LIB_FUNC, API_DNN}},
|
||||
{"cudnnTransformTensor", {"hipdnnTransformTensor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnTransformTensorEx", {"hipdnnTransformTensorEx", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnInitTransformDest", {"hipdnnInitTransformDest", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnCreateTensorTransformDescriptor", {"hipdnnCreateTensorTransformDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSetTensorTransformDescriptor", {"hipdnnSetTensorTransformDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetTensorTransformDescriptor", {"hipdnnGetTensorTransformDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnDestroyTensorTransformDescriptor", {"hipdnnDestroyTensorTransformDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnAddTensor", {"hipdnnAddTensor", "", CONV_LIB_FUNC, API_DNN}},
|
||||
{"cudnnCreateOpTensorDescriptor", {"hipdnnCreateOpTensorDescriptor", "", CONV_LIB_FUNC, API_DNN}},
|
||||
{"cudnnSetOpTensorDescriptor", {"hipdnnSetOpTensorDescriptor", "", CONV_LIB_FUNC, API_DNN}},
|
||||
@@ -223,6 +229,8 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_FUNCTION_MAP{
|
||||
{"cudnnDestroyRNNDataDescriptor", {"hipdnnDestroyRNNDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSetRNNDataDescriptor", {"hipdnnSetRNNDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetRNNDataDescriptor", {"hipdnnGetRNNDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSetRNNBiasMode", {"hipdnnSetRNNBiasMode", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetRNNBiasMode", {"hipdnnGetRNNBiasMode", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
|
||||
// cuDNN Connectionist Temporal Classification loss functions
|
||||
{"cudnnCreateCTCLossDescriptor", {"hipdnnCreateCTCLossDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
@@ -250,4 +258,20 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_FUNCTION_MAP{
|
||||
{"cudnnRNNSetClip", {"hipdnnRNNSetClip", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnRNNGetClip", {"hipdnnRNNGetClip", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
|
||||
// cuDNN Sequence functions
|
||||
{"cudnnCreateSeqDataDescriptor", {"hipdnnCreateSeqDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnDestroySeqDataDescriptor", {"hipdnnDestroySeqDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSetSeqDataDescriptor", {"hipdnnSetSeqDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetSeqDataDescriptor", {"hipdnnGetSeqDataDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
|
||||
// cuDNN Multihead Attention functions
|
||||
{"cudnnCreateAttnDescriptor", {"hipdnnCreateAttnDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnDestroyAttnDescriptor", {"hipdnnDestroyAttnDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSetAttnDescriptor", {"hipdnnSetAttnDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetAttnDescriptor", {"hipdnnGetAttnDescriptor", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetMultiHeadAttnBuffers", {"hipdnnGetMultiHeadAttnBuffers", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnGetMultiHeadAttnWeights", {"hipdnnGetMultiHeadAttnWeights", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnMultiHeadAttnForward", {"hipdnnMultiHeadAttnForward", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnMultiHeadAttnBackwardData", {"hipdnnMultiHeadAttnBackwardData", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnMultiHeadAttnBackwardWeights", {"hipdnnMultiHeadAttnBackwardWeights", "", CONV_LIB_FUNC, API_DNN, HIP_UNSUPPORTED}},
|
||||
};
|
||||
|
||||
@@ -35,6 +35,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"CUDNN_SEV_ERROR_EN", {"HIPDNN_SEV_ERROR_EN", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_SEV_WARNING_EN", {"HIPDNN_SEV_WARNING_EN", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_SEV_INFO_EN", {"HIPDNN_SEV_INFO_EN", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_SEQDATA_DIM_COUNT", {"HIPDNN_SEQDATA_DIM_COUNT", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 4
|
||||
|
||||
// cuDNN enums
|
||||
{"cudnnStatus_t", {"hipdnnStatus_t", "", CONV_TYPE, API_DNN}},
|
||||
@@ -60,6 +61,9 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"CUDNN_TENSOR_NCHW", {"HIPDNN_TENSOR_NCHW", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
|
||||
{"CUDNN_TENSOR_NHWC", {"HIPDNN_TENSOR_NHWC", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
{"CUDNN_TENSOR_NCHW_VECT_C", {"HIPDNN_TENSOR_NCHW_VECT_C", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
|
||||
{"cudnnFoldingDirection_t", {"hipdnnFoldingDirection_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_TRANSFORM_FOLD", {"HIPDNN_TRANSFORM_FOLD", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0U
|
||||
{"CUDNN_TRANSFORM_UNFOLD", {"HIPDNN_TRANSFORM_UNFOLD", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1U
|
||||
{"cudnnDataType_t", {"hipdnnDataType_t", "", CONV_TYPE, API_DNN}},
|
||||
{"CUDNN_DATA_FLOAT", {"HIPDNN_DATA_FLOAT", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
|
||||
{"CUDNN_DATA_DOUBLE", {"HIPDNN_DATA_DOUBLE", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
@@ -147,6 +151,11 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"CUDNN_RNN_TANH", {"HIPDNN_RNN_TANH", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
{"CUDNN_LSTM", {"HIPDNN_LSTM", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
|
||||
{"CUDNN_GRU", {"HIPDNN_GRU", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 3
|
||||
{"cudnnRNNBiasMode_t", {"hipdnnRNNBiasMode_t", "", CONV_TYPE, API_DNN}},
|
||||
{"CUDNN_RNN_NO_BIAS", {"HIPDNN_RNN_NO_BIAS", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
|
||||
{"CUDNN_RNN_SINGLE_INP_BIAS", {"HIPDNN_RNN_WITH_BIAS", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
{"CUDNN_RNN_DOUBLE_BIAS", {"HIPDNN_RNN_WITH_BIAS", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
{"CUDNN_RNN_SINGLE_REC_BIAS", {"HIPDNN_RNN_WITH_BIAS", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
{"cudnnOpTensorOp_t", {"hipdnnOpTensorOp_t", "", CONV_TYPE, API_DNN}},
|
||||
{"CUDNN_OP_TENSOR_ADD", {"HIPDNN_OP_TENSOR_ADD", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
|
||||
{"CUDNN_OP_TENSOR_MUL", {"HIPDNN_OP_TENSOR_MUL", "", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
|
||||
@@ -215,6 +224,22 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"cudnnRNNPaddingMode_t", {"hipdnnRNNPaddingMode_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_RNN_PADDED_IO_DISABLED", {"HIPDNN_RNN_PADDED_IO_DISABLED", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0
|
||||
{"CUDNN_RNN_PADDED_IO_ENABLED", {"HIPDNN_RNN_PADDED_IO_ENABLED", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1
|
||||
{"cudnnSeqDataAxis_t", {"hipdnnSeqDataAxis_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_SEQDATA_TIME_DIM", {"HIPDNN_SEQDATA_TIME_DIM", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0
|
||||
{"CUDNN_SEQDATA_BATCH_DIM", {"HIPDNN_SEQDATA_BATCH_DIM", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1
|
||||
{"CUDNN_SEQDATA_BEAM_DIM", {"HIPDNN_SEQDATA_BEAM_DIM", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 2
|
||||
{"CUDNN_SEQDATA_VECT_DIM", {"HIPDNN_SEQDATA_VECT_DIM", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 3
|
||||
{"cudnnAttnQueryMap_t", {"hipdnnAttnQueryMap_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_ATTN_QUERYMAP_ALL_TO_ONE", {"HIPDNN_ATTN_QUERYMAP_ALL_TO_ONE", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0
|
||||
{"CUDNN_ATTN_QUERYMAP_ONE_TO_ONE", {"HIPDNN_ATTN_QUERYMAP_ONE_TO_ONE", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1
|
||||
{"cudnnMultiHeadAttnWeightKind_t", {"hipdnnMultiHeadAttnWeightKind_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_MH_ATTN_Q_WEIGHTS", {"HIPDNN_MH_ATTN_Q_WEIGHTS", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0
|
||||
{"CUDNN_MH_ATTN_K_WEIGHTS", {"HIPDNN_MH_ATTN_K_WEIGHTS", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1
|
||||
{"CUDNN_MH_ATTN_V_WEIGHTS", {"HIPDNN_MH_ATTN_V_WEIGHTS", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 2
|
||||
{"CUDNN_MH_ATTN_O_WEIGHTS", {"HIPDNN_MH_ATTN_O_WEIGHTS", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 3
|
||||
{"cudnnWgradMode_t", {"hipdnnWgradMode_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"CUDNN_WGRAD_MODE_ADD", {"HIPDNN_WGRAD_MODE_ADD", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 0
|
||||
{"CUDNN_WGRAD_MODE_SET", {"HIPDNN_WGRAD_MODE_SET", "", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 1
|
||||
|
||||
// cuDNN types
|
||||
{"cudnnContext", {"hipdnnContext", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
@@ -239,6 +264,8 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"cudnnReduceTensorDescriptor_t", {"hipdnnReduceTensorDescriptor_t", "", CONV_TYPE, API_DNN}},
|
||||
{"cudnnCTCLossStruct", {"hipdnnCTCLossStruct", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnCTCLossDescriptor_t", {"hipdnnCTCLossDescriptor_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnTensorTransformStruct", {"hipdnnTensorTransformStruct", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnTensorTransformDescriptor_t", {"hipdnnTensorTransformDescriptor_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnConvolutionFwdAlgoPerf_t", {"hipdnnConvolutionFwdAlgoPerf_t", "", CONV_TYPE, API_DNN}},
|
||||
{"cudnnConvolutionBwdFilterAlgoPerf_t", {"hipdnnConvolutionBwdFilterAlgoPerf_t", "", CONV_TYPE, API_DNN}},
|
||||
{"cudnnConvolutionBwdDataAlgoPerf_t", {"hipdnnConvolutionBwdDataAlgoPerf_t", "", CONV_TYPE, API_DNN}},
|
||||
@@ -257,4 +284,9 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DNN_TYPE_NAME_MAP{
|
||||
{"cudnnCallback_t", {"hipdnnCallback_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnRNNDataStruct", {"hipdnnRNNDataStruct", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnRNNDataDescriptor_t", {"hipdnnRNNDataDescriptor_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSeqDataStruct", {"hipdnnSeqDataStruct", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnSeqDataDescriptor_t", {"hipdnnSeqDataDescriptor_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnAttnStruct", {"hipdnnAttnStruct", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
{"cudnnAttnDescriptor_t", {"hipdnnAttnDescriptor_t", "", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
|
||||
|
||||
};
|
||||
|
||||
@@ -516,6 +516,8 @@ const std::map<llvm::StringRef, hipCounter> CUDA_SPARSE_FUNCTION_MAP{
|
||||
{"cusparseZcsr2csc", {"hipsparseZcsr2csc", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseCsr2cscEx", {"hipsparseCsr2cscEx", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseCsr2cscEx2", {"hipsparseCsr2cscEx2", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseCsr2cscEx2_bufferSize", {"hipsparseCsr2cscEx2_bufferSize", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseScsr2dense", {"hipsparseScsr2dense", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDcsr2dense", {"hipsparseDcsr2dense", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
@@ -616,9 +618,9 @@ const std::map<llvm::StringRef, hipCounter> CUDA_SPARSE_FUNCTION_MAP{
|
||||
{"cusparseSpruneDense2csrNnzByPercentage", {"hipsparseSpruneDense2csrNnzByPercentage", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDpruneDense2csrNnzByPercentage", {"hipsparseDpruneDense2csrNnzByPercentage", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseHpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseHpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseSpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseDpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseHpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseHpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseSpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDpruneCsr2csrByPercentage_bufferSizeExt", {"hipsparseDpruneCsr2csrByPercentage_bufferSizeExt", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseHpruneCsr2csrNnzByPercentage", {"hipsparseHpruneCsr2csrNnzByPercentage", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpruneCsr2csrNnzByPercentage", {"hipsparseSpruneCsr2csrNnzByPercentage", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
@@ -628,4 +630,25 @@ const std::map<llvm::StringRef, hipCounter> CUDA_SPARSE_FUNCTION_MAP{
|
||||
{"cusparseDnnz_compress", {"hipsparseDnnz_compress", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseCnnz_compress", {"hipsparseCnnz_compress", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseZnnz_compress", {"hipsparseZnnz_compress", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
// 13. cuSPARSE Generic API Reference
|
||||
// Generic Sparse API helper functions
|
||||
{"cusparseCreateCoo", {"hipsparseCreateCoo", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDestroySpMat", {"hipsparseDestroySpMat", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseCooGet", {"hipsparseCooGet", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMatGetFormat", {"hipsparseSpMatGetFormat", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMatGetIndexBase", {"hipsparseSpMatGetIndexBase", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMatSetNumBatches", {"hipsparseSpMatSetNumBatches", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMatGetNumBatches", {"hipsparseSpMatGetNumBatches", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
// Generic Dense API helper functions
|
||||
{"cusparseCreateDnMat", {"hipsparseCreateDnMat", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDestroyDnMat", {"hipsparseDestroyDnMat", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDnMatGet", {"hipsparseDnMatGet", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDnMatSetStridedBatch", {"hipsparseDnMatSetStridedBatch", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDnMatGetStridedBatch", {"hipsparseDnMatGetStridedBatch", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
// Generic SpMM API functions
|
||||
{"cusparseSpMM", {"hipsparseSpMM", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMM_bufferSize", {"hipsparseSpMM_bufferSize", "", CONV_LIB_FUNC, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
};
|
||||
|
||||
@@ -68,6 +68,12 @@ const std::map<llvm::StringRef, hipCounter> CUDA_SPARSE_TYPE_NAME_MAP{
|
||||
{"pruneInfo", {"pruneInfo", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"pruneInfo_t", {"pruneInfo_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseSpMatDescr", {"hipsparseSpMatDescr", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseSpMatDescr_t", {"hipsparseSpMatDescr_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseDnMatDescr", {"hipsparseDnMatDescr", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"cusparseDnMatDescr_t", {"hipsparseDnMatDescr_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
// 2. Enums
|
||||
{"cusparseAction_t", {"hipsparseAction_t", "", CONV_TYPE, API_SPARSE}},
|
||||
{"CUSPARSE_ACTION_SYMBOLIC", {"HIPSPARSE_ACTION_SYMBOLIC", "", CONV_NUMERIC_LITERAL, API_SPARSE}},
|
||||
@@ -130,4 +136,27 @@ const std::map<llvm::StringRef, hipCounter> CUDA_SPARSE_TYPE_NAME_MAP{
|
||||
{"CUSPARSE_STATUS_INTERNAL_ERROR", {"HIPSPARSE_STATUS_INTERNAL_ERROR", "", CONV_NUMERIC_LITERAL, API_SPARSE}},
|
||||
{"CUSPARSE_STATUS_MATRIX_TYPE_NOT_SUPPORTED", {"HIPSPARSE_STATUS_MATRIX_TYPE_NOT_SUPPORTED", "", CONV_NUMERIC_LITERAL, API_SPARSE}},
|
||||
{"CUSPARSE_STATUS_ZERO_PIVOT", {"HIPSPARSE_STATUS_ZERO_PIVOT", "", CONV_NUMERIC_LITERAL, API_SPARSE}},
|
||||
|
||||
{"cusparseCsr2CscAlg_t", {"hipsparseCsr2CscAlg_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_CSR2CSC_ALG1", {"HIPSPARSE_CSR2CSC_ALG1", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_CSR2CSC_ALG2", {"HIPSPARSE_CSR2CSC_ALG2", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseFormat_t", {"hipsparseFormat_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_FORMAT_CSR", {"HIPSPARSE_FORMAT_CSR", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_FORMAT_CSC", {"HIPSPARSE_FORMAT_CSC", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_FORMAT_COO", {"HIPSPARSE_FORMAT_COO", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseOrder_t", {"hipsparseOrder_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_ORDER_COL", {"HIPSPARSE_ORDER_COL", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_ORDER_ROW", {"HIPSPARSE_ORDER_ROW", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseSpMMAlg_t", {"hipsparseSpMMAlg_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_COOMM_ALG1", {"HIPSPARSE_COOMM_ALG1", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_COOMM_ALG2", {"HIPSPARSE_COOMM_ALG2", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_COOMM_ALG3", {"HIPSPARSE_COOMM_ALG3", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
{"cusparseIndexType_t", {"hipsparseIndexType_t", "", CONV_TYPE, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_INDEX_16U", {"HIPSPARSE_INDEX_16U", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
{"CUSPARSE_INDEX_32I", {"HIPSPARSE_INDEX_32I", "", CONV_NUMERIC_LITERAL, API_SPARSE, HIP_UNSUPPORTED}},
|
||||
|
||||
};
|
||||
|
||||
@@ -89,9 +89,10 @@ struct Bundled_code {
|
||||
std::vector<char> blob;
|
||||
};
|
||||
|
||||
#define magic_string_ "__CLANG_OFFLOAD_BUNDLE__"
|
||||
|
||||
class Bundled_code_header {
|
||||
// DATA - STATICS
|
||||
static constexpr const char magic_string_[] = "__CLANG_OFFLOAD_BUNDLE__";
|
||||
static constexpr auto magic_string_sz_ = sizeof(magic_string_) - 1;
|
||||
|
||||
// DATA
|
||||
@@ -143,12 +144,12 @@ class Bundled_code_header {
|
||||
std::vector<char>{std::istreambuf_iterator<char>{is}, std::istreambuf_iterator<char>{}},
|
||||
x);
|
||||
}
|
||||
|
||||
// FRIENDS - ACCESSORS
|
||||
friend inline bool valid(const Bundled_code_header& x) {
|
||||
return std::equal(magic_string_, magic_string_ + magic_string_sz_,
|
||||
x.header_.bundler_magic_string_);
|
||||
const std::string ms = {magic_string_};
|
||||
return std::equal(ms.begin(), ms.end(), x.header_.bundler_magic_string_);
|
||||
}
|
||||
|
||||
friend inline const std::vector<Bundled_code>& bundles(const Bundled_code_header& x) {
|
||||
return x.bundles_;
|
||||
}
|
||||
|
||||
@@ -2549,7 +2549,7 @@ hipError_t hipModuleGetFunction(hipFunction_t* function, hipModule_t module, con
|
||||
* @returns hipSuccess, hipErrorInvalidDeviceFunction
|
||||
*/
|
||||
|
||||
hipError_t hipFuncGetAttributes(hipFuncAttributes* attr, const void* func);
|
||||
hipError_t hipFuncGetAttributes(struct hipFuncAttributes* attr, const void* func);
|
||||
|
||||
struct Agent_global {
|
||||
|
||||
|
||||
@@ -39,6 +39,7 @@ THE SOFTWARE.
|
||||
#include <hsa/hsa.h>
|
||||
#include <hsa/hsa_ext_amd.h>
|
||||
#include <hsa/hsa_ven_amd_loader.h>
|
||||
#include <amd_comgr.h>
|
||||
|
||||
#include <link.h>
|
||||
|
||||
@@ -55,6 +56,8 @@ THE SOFTWARE.
|
||||
#include <unordered_map>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
#include <iostream>
|
||||
#include <fstream>
|
||||
|
||||
struct ihipModuleSymbol_t;
|
||||
using hipFunction_t = ihipModuleSymbol_t*;
|
||||
@@ -558,92 +561,185 @@ const std::unordered_map<
|
||||
}
|
||||
|
||||
inline
|
||||
std::size_t parse_args(
|
||||
const std::string& metadata,
|
||||
std::size_t f,
|
||||
std::size_t l,
|
||||
std::vector<std::pair<std::size_t, std::size_t>>& size_align) {
|
||||
if (f == l) return f;
|
||||
if (!size_align.empty()) return l;
|
||||
void checkError(
|
||||
amd_comgr_status_t status,
|
||||
char const *str) {
|
||||
if (status != AMD_COMGR_STATUS_SUCCESS) {
|
||||
const char *status_str;
|
||||
status = amd_comgr_status_string(status, &status_str);
|
||||
if (status == AMD_COMGR_STATUS_SUCCESS)
|
||||
std::cerr << "FAILED: " << str << "\n REASON: " << status_str << std::endl;
|
||||
hip_throw(std::runtime_error{"Metadata parsing failed."});
|
||||
}
|
||||
}
|
||||
|
||||
do {
|
||||
static constexpr size_t size_sz{5};
|
||||
f = metadata.find("Size:", f) + size_sz;
|
||||
class comgr_metadata_node {
|
||||
public:
|
||||
amd_comgr_metadata_node_t node;
|
||||
bool active;
|
||||
comgr_metadata_node() : active(false) {}
|
||||
~comgr_metadata_node() {
|
||||
if(active)
|
||||
checkError(amd_comgr_destroy_metadata(node), "amd_comgr_destroy_metadata");
|
||||
}
|
||||
bool is_active() { return active; }
|
||||
void set_active(bool value) { active = value; }
|
||||
comgr_metadata_node(const comgr_metadata_node&) = delete;
|
||||
comgr_metadata_node(comgr_metadata_node&&) = delete;
|
||||
comgr_metadata_node& operator=(const comgr_metadata_node&) = delete;
|
||||
comgr_metadata_node& operator=(comgr_metadata_node&&) = delete;
|
||||
|
||||
if (l <= f) return f;
|
||||
};
|
||||
|
||||
auto size = std::strtoul(&metadata[f], nullptr, 10);
|
||||
class comgr_data {
|
||||
public:
|
||||
amd_comgr_data_t data;
|
||||
bool active;
|
||||
comgr_data() : active(false) {}
|
||||
~comgr_data() {
|
||||
if(active)
|
||||
checkError(amd_comgr_release_data(data), "amd_comgr_release_data");
|
||||
}
|
||||
bool is_active() { return active; }
|
||||
void set_active(bool value) { active = value; }
|
||||
comgr_data(const comgr_data&) = delete;
|
||||
comgr_data(comgr_data&&) = delete;
|
||||
comgr_data& operator=(const comgr_data&) = delete;
|
||||
comgr_data& operator=(comgr_data&&) = delete;
|
||||
|
||||
static constexpr size_t align_sz{6};
|
||||
f = metadata.find("Align:", f) + align_sz;
|
||||
};
|
||||
|
||||
char* l{};
|
||||
auto align = std::strtoul(&metadata[f], &l, 10);
|
||||
inline
|
||||
std::string lookup_keyword_value(
|
||||
amd_comgr_metadata_node_t& in_node,
|
||||
std::string keyword) {
|
||||
amd_comgr_status_t status;
|
||||
size_t value_size;
|
||||
comgr_metadata_node value_meta;
|
||||
|
||||
f += (l - &metadata[f]) + 1;
|
||||
status = amd_comgr_metadata_lookup(in_node, keyword.c_str(), &value_meta.node);
|
||||
checkError(status, "amd_comgr_metadata_lookup");
|
||||
value_meta.set_active(true);
|
||||
status = amd_comgr_get_metadata_string(value_meta.node, &value_size, NULL);
|
||||
checkError(status, "amd_comgr_get_metadata_string");
|
||||
// Since value_size returns size with null terminator, we don't include for C++ string size
|
||||
value_size--;
|
||||
std::string value(value_size, '\0');
|
||||
status = amd_comgr_get_metadata_string(value_meta.node, &value_size, &value[0]);
|
||||
checkError(status, "amd_comgr_get_metadata_string");
|
||||
|
||||
size_align.emplace_back(size, align);
|
||||
} while (true);
|
||||
return value;
|
||||
}
|
||||
|
||||
inline
|
||||
void read_kernarg_metadata(
|
||||
ELFIO::elfio& reader,
|
||||
void process_kernarg_metadata(
|
||||
amd_comgr_metadata_node_t blob_meta,
|
||||
std::unordered_map<
|
||||
std::string,
|
||||
std::vector<std::pair<std::size_t, std::size_t>>>& kernargs) {
|
||||
// TODO: this is inefficient.
|
||||
auto it = find_section_if(reader, [](const ELFIO::section* x) {
|
||||
return x->get_type() == SHT_NOTE;
|
||||
});
|
||||
amd_comgr_status_t status;
|
||||
amd_comgr_metadata_kind_t mkindLookup;
|
||||
comgr_metadata_node kernelList;
|
||||
std::string kernel_name;
|
||||
size_t num_kernels = 0;
|
||||
size_t num_kern_args = 0;
|
||||
|
||||
if (!it) return;
|
||||
// Kernels is a list of MAPS!!
|
||||
status = amd_comgr_metadata_lookup(blob_meta, "Kernels", &kernelList.node);
|
||||
// FIXME - few hip memset kernels are missing Kernels node
|
||||
if(status != AMD_COMGR_STATUS_SUCCESS)
|
||||
return;
|
||||
kernelList.set_active(true);
|
||||
|
||||
const ELFIO::note_section_accessor acc{reader, it};
|
||||
for (decltype(acc.get_notes_num()) i = 0; i != acc.get_notes_num(); ++i) {
|
||||
ELFIO::Elf_Word type{};
|
||||
std::string name{};
|
||||
void* desc{};
|
||||
ELFIO::Elf_Word desc_size{};
|
||||
|
||||
acc.get_note(i, type, name, desc, desc_size);
|
||||
|
||||
if (name != "AMD") continue; // TODO: switch to using NT_AMD_AMDGPU_HSA_METADATA.
|
||||
|
||||
std::string tmp{
|
||||
static_cast<char*>(desc), static_cast<char*>(desc) + desc_size};
|
||||
|
||||
auto dx = tmp.find("Kernels:");
|
||||
|
||||
if (dx == std::string::npos) continue;
|
||||
|
||||
static constexpr decltype(tmp.size()) kernels_sz{8};
|
||||
dx += kernels_sz;
|
||||
|
||||
do {
|
||||
dx = tmp.find("Name:", dx);
|
||||
|
||||
if (dx == std::string::npos) break;
|
||||
|
||||
static constexpr decltype(tmp.size()) name_sz{5};
|
||||
dx = tmp.find_first_not_of(" '", dx + name_sz);
|
||||
|
||||
auto fn = tmp.substr(dx, tmp.find_first_of("'\n", dx) - dx);
|
||||
dx += fn.size();
|
||||
|
||||
auto dx1 = tmp.find("CodeProps", dx);
|
||||
dx = tmp.find("Args:", dx);
|
||||
|
||||
if (dx1 < dx) {
|
||||
dx = dx1;
|
||||
continue;
|
||||
}
|
||||
if (dx == std::string::npos) break;
|
||||
|
||||
static constexpr decltype(tmp.size()) args_sz{5};
|
||||
dx = parse_args(tmp, dx + args_sz, dx1, kernargs[fn]);
|
||||
} while (true);
|
||||
status = amd_comgr_get_metadata_kind(kernelList.node, &mkindLookup);
|
||||
if (mkindLookup != AMD_COMGR_METADATA_KIND_LIST) {
|
||||
hip_throw(std::runtime_error{"Lookup of Kernels didn't return a list\n"});
|
||||
}
|
||||
|
||||
status = amd_comgr_get_metadata_list_size(kernelList.node, &num_kernels);
|
||||
checkError(status, "amd_comgr_get_metadata_list_size");
|
||||
for (int i = 0; i < num_kernels; i++) {
|
||||
comgr_metadata_node kernelMap;
|
||||
status = amd_comgr_index_list_metadata(kernelList.node, i, &kernelMap.node);
|
||||
checkError(status, "amd_comgr_index_list_metadata");
|
||||
kernelMap.set_active(true);
|
||||
|
||||
kernel_name = std::move(lookup_keyword_value(kernelMap.node, "Name"));
|
||||
|
||||
// Check if this kernel was already processed
|
||||
if(!kernargs[kernel_name].empty()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
comgr_metadata_node kernArgList;
|
||||
status = amd_comgr_metadata_lookup(kernelMap.node, "Args", &kernArgList.node);
|
||||
if (status == AMD_COMGR_STATUS_SUCCESS ) {
|
||||
kernArgList.set_active(true);
|
||||
status = amd_comgr_get_metadata_kind(kernArgList.node, &mkindLookup);
|
||||
if (mkindLookup != AMD_COMGR_METADATA_KIND_LIST) {
|
||||
hip_throw(std::runtime_error{"Lookup of Args didn't return a list\n"});
|
||||
}
|
||||
|
||||
if (status == AMD_COMGR_STATUS_SUCCESS ) {
|
||||
status = amd_comgr_get_metadata_list_size(kernArgList.node, &num_kern_args);
|
||||
checkError(status, "amd_comgr_get_metadata_list_size");
|
||||
for (int k_ar = 0; k_ar < num_kern_args; k_ar++) {
|
||||
comgr_metadata_node kernArgMap;
|
||||
status = amd_comgr_index_list_metadata(kernArgList.node, k_ar, &kernArgMap.node);
|
||||
checkError(status, "amd_comgr_index_list_metadata");
|
||||
kernArgMap.set_active(true);
|
||||
size_t k_arg_size, k_arg_align;
|
||||
|
||||
k_arg_size = std::stoul(lookup_keyword_value(kernArgMap.node, "Size"));
|
||||
k_arg_align = std::stoul(lookup_keyword_value(kernArgMap.node, "Align"));
|
||||
|
||||
// Save it into our kernargs
|
||||
kernargs[kernel_name].emplace_back(k_arg_size, k_arg_align);
|
||||
|
||||
} // end kernArgMap
|
||||
}
|
||||
} // end kernArgList
|
||||
} // end kernelMap
|
||||
} // end kernelList
|
||||
|
||||
inline
|
||||
void read_kernarg_metadata(
|
||||
std::string blob,
|
||||
std::unordered_map<
|
||||
std::string,
|
||||
std::vector<std::pair<std::size_t, std::size_t>>>& kernargs) {
|
||||
|
||||
const char *blob_buf = blob.data();
|
||||
long blob_size = blob.size();
|
||||
|
||||
amd_comgr_status_t status;
|
||||
comgr_data blob_data;
|
||||
status = amd_comgr_create_data(AMD_COMGR_DATA_KIND_RELOCATABLE, &blob_data.data);
|
||||
checkError(status, "amd_comgr_create_data");
|
||||
blob_data.set_active(true);
|
||||
|
||||
status = amd_comgr_set_data(blob_data.data, blob_size, blob_buf);
|
||||
if(status != AMD_COMGR_STATUS_SUCCESS)
|
||||
return;
|
||||
|
||||
// We have a valid code object now
|
||||
status = amd_comgr_set_data_name(blob_data.data, "HIP Code Object");
|
||||
checkError(status, "amd_comgr_set_data_name");
|
||||
|
||||
comgr_metadata_node blob_meta;
|
||||
status = amd_comgr_get_data_metadata(blob_data.data, &blob_meta.node);
|
||||
checkError(status, "amd_comgr_get_data_metadata");
|
||||
blob_meta.set_active(true);
|
||||
|
||||
// Root is a map
|
||||
amd_comgr_metadata_kind_t blob_mkind;
|
||||
status = amd_comgr_get_metadata_kind(blob_meta.node, &blob_mkind);
|
||||
checkError(status, "amd_comgr_get_metadata_kind");
|
||||
if (blob_mkind != AMD_COMGR_METADATA_KIND_MAP) {
|
||||
hip_throw(std::runtime_error{"Root is not map\n"});
|
||||
}
|
||||
|
||||
process_kernarg_metadata(blob_meta.node, kernargs);
|
||||
}
|
||||
|
||||
inline
|
||||
@@ -657,13 +753,7 @@ const std::unordered_map<
|
||||
std::call_once(f, []() {
|
||||
for (auto&& isa_blobs : code_object_blobs()) {
|
||||
for (auto&& blob : isa_blobs.second) {
|
||||
std::stringstream tmp{std::string{blob.cbegin(), blob.cend()}};
|
||||
|
||||
ELFIO::elfio reader;
|
||||
|
||||
if (!reader.load(tmp)) continue;
|
||||
|
||||
read_kernarg_metadata(reader, r);
|
||||
read_kernarg_metadata(std::string{blob.cbegin(), blob.cend()}, r);
|
||||
}
|
||||
}
|
||||
});
|
||||
|
||||
@@ -30,9 +30,9 @@ set(CPACK_GENERATOR "TGZ;DEB;RPM")
|
||||
set(CPACK_BINARY_DEB "ON")
|
||||
set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJECT_BINARY_DIR}/postinst;${PROJECT_BINARY_DIR}/prerm")
|
||||
if(@COMPILE_HIP_ATP_MARKER@)
|
||||
set(CPACK_DEBIAN_PACKAGE_DEPENDS "hip_base (= ${CPACK_PACKAGE_VERSION}), ${HCC_PACKAGE_NAME} (= @HCC_PACKAGE_VERSION@), rocm-profiler")
|
||||
set(CPACK_DEBIAN_PACKAGE_DEPENDS "hip_base (= ${CPACK_PACKAGE_VERSION}), ${HCC_PACKAGE_NAME} (= @HCC_PACKAGE_VERSION@), rocm-profiler, comgr (>= 1.1)")
|
||||
else()
|
||||
set(CPACK_DEBIAN_PACKAGE_DEPENDS "hip_base (= ${CPACK_PACKAGE_VERSION}), ${HCC_PACKAGE_NAME} (= @HCC_PACKAGE_VERSION@)")
|
||||
set(CPACK_DEBIAN_PACKAGE_DEPENDS "hip_base (= ${CPACK_PACKAGE_VERSION}), ${HCC_PACKAGE_NAME} (= @HCC_PACKAGE_VERSION@), comgr (>= 1.1)")
|
||||
endif()
|
||||
set(CPACK_BINARY_RPM "ON")
|
||||
set(CPACK_RPM_PACKAGE_ARCHITECTURE "${CMAKE_SYSTEM_PROCESSOR}")
|
||||
@@ -40,9 +40,9 @@ set(CPACK_RPM_POST_INSTALL_SCRIPT_FILE "${PROJECT_BINARY_DIR}/postinst")
|
||||
set(CPACK_RPM_PRE_UNINSTALL_SCRIPT_FILE "${PROJECT_BINARY_DIR}/prerm")
|
||||
set(CPACK_RPM_PACKAGE_AUTOREQPROV " no")
|
||||
if(@COMPILE_HIP_ATP_MARKER@)
|
||||
set(CPACK_RPM_PACKAGE_REQUIRES "hip_base = ${CPACK_PACKAGE_VERSION}, ${HCC_PACKAGE_NAME} = @HCC_PACKAGE_VERSION@, rocm-profiler")
|
||||
set(CPACK_RPM_PACKAGE_REQUIRES "hip_base = ${CPACK_PACKAGE_VERSION}, ${HCC_PACKAGE_NAME} = @HCC_PACKAGE_VERSION@, rocm-profiler, comgr")
|
||||
else()
|
||||
set(CPACK_RPM_PACKAGE_REQUIRES "hip_base = ${CPACK_PACKAGE_VERSION}, ${HCC_PACKAGE_NAME} = @HCC_PACKAGE_VERSION@")
|
||||
set(CPACK_RPM_PACKAGE_REQUIRES "hip_base = ${CPACK_PACKAGE_VERSION}, ${HCC_PACKAGE_NAME} = @HCC_PACKAGE_VERSION@, comgr")
|
||||
endif()
|
||||
set(CPACK_RPM_EXCLUDE_FROM_AUTO_FILELIST_ADDITION "/opt")
|
||||
set(CPACK_SOURCE_GENERATOR "TGZ")
|
||||
|
||||
@@ -10,9 +10,6 @@
|
||||
|
||||
using namespace std;
|
||||
|
||||
// DATA - STATICS
|
||||
constexpr const char hip_impl::Bundled_code_header::magic_string_[];
|
||||
|
||||
// CREATORS
|
||||
hip_impl::Bundled_code_header::Bundled_code_header(const vector<char>& x)
|
||||
: Bundled_code_header{x.cbegin(), x.cend()} {}
|
||||
@@ -22,12 +19,9 @@ hip_impl::Bundled_code_header::Bundled_code_header(
|
||||
// hipLoadModuleData is so poorly specified (for no fault of its own).
|
||||
if (!p) return;
|
||||
|
||||
if (!valid(*static_cast<const Bundled_code_header*>(p))) return;
|
||||
auto ph = static_cast<const Header_*>(p);
|
||||
|
||||
if (!equal(magic_string_, magic_string_ + magic_string_sz_, ph->bundler_magic_string_)) {
|
||||
return;
|
||||
}
|
||||
|
||||
size_t sz = sizeof(Header_) + ph->bundle_cnt_ * sizeof(Bundled_code::Header);
|
||||
auto pb = static_cast<const char*>(p) + sizeof(Header_);
|
||||
auto n = ph->bundle_cnt_;
|
||||
|
||||
@@ -92,6 +92,7 @@ int HIP_EVENT_SYS_RELEASE = 0;
|
||||
int HIP_HOST_COHERENT = 1;
|
||||
|
||||
int HIP_SYNC_HOST_ALLOC = 1;
|
||||
int HIP_SYNC_FREE = 0;
|
||||
|
||||
|
||||
int HIP_INIT_ALLOC = -1;
|
||||
@@ -1279,6 +1280,8 @@ void HipReadEnv() {
|
||||
|
||||
READ_ENV_I(release, HIP_SYNC_STREAM_WAIT, 0, "hipStreamWaitEvent will synchronize to host");
|
||||
|
||||
READ_ENV_I(release, HIP_SYNC_FREE, 0,
|
||||
"Force all calls to hipFree to sync all devices and all streams");
|
||||
|
||||
READ_ENV_I(release, HIP_HOST_COHERENT, 0,
|
||||
"If set, all host memory will be allocated as fine-grained system memory. This "
|
||||
|
||||
@@ -83,6 +83,8 @@ extern int HIP_SYNC_NULL_STREAM;
|
||||
extern int HIP_INIT_ALLOC;
|
||||
extern int HIP_FORCE_NULL_STREAM;
|
||||
|
||||
extern int HIP_SYNC_FREE;
|
||||
|
||||
extern int HIP_DUMP_CODE_OBJECT;
|
||||
|
||||
// TODO - remove when this is standard behavior.
|
||||
|
||||
@@ -1902,10 +1902,6 @@ hipError_t hipFree(void* ptr) {
|
||||
|
||||
hipError_t hipStatus = hipErrorInvalidDevicePointer;
|
||||
|
||||
// Synchronize to ensure all work has finished.
|
||||
ihipGetTlsDefaultCtx()->locked_waitAllStreams(); // ignores non-blocking streams, this waits
|
||||
// for all activity to finish.
|
||||
|
||||
if (ptr) {
|
||||
hc::accelerator acc;
|
||||
#if (__hcc_workweek__ >= 17332)
|
||||
@@ -1916,6 +1912,29 @@ hipError_t hipFree(void* ptr) {
|
||||
am_status_t status = hc::am_memtracker_getinfo(&amPointerInfo, ptr);
|
||||
if (status == AM_SUCCESS) {
|
||||
if (amPointerInfo._hostPointer == NULL) {
|
||||
if (HIP_SYNC_FREE) {
|
||||
// Synchronize all devices, all streams
|
||||
// to ensure all work has finished on all devices.
|
||||
// This is disabled by default.
|
||||
for (unsigned i = 0; i < g_deviceCnt; i++) {
|
||||
ihipGetPrimaryCtx(i)->locked_waitAllStreams();
|
||||
}
|
||||
}
|
||||
else {
|
||||
ihipCtx_t* ctx;
|
||||
if (amPointerInfo._appId != -1) {
|
||||
#if USE_APP_PTR_FOR_CTX
|
||||
ctx = static_cast<ihipCtx_t*>(amPointerInfo._appPtr);
|
||||
#else
|
||||
ctx = ihipGetPrimaryCtx(amPointerInfo._appId);
|
||||
#endif
|
||||
} else {
|
||||
ctx = ihipGetTlsDefaultCtx();
|
||||
}
|
||||
// Synchronize to ensure all work has finished on device owning the memory.
|
||||
ctx->locked_waitAllStreams(); // ignores non-blocking streams, this waits
|
||||
// for all activity to finish.
|
||||
}
|
||||
hc::am_free(ptr);
|
||||
hipStatus = hipSuccess;
|
||||
}
|
||||
|
||||
Ссылка в новой задаче
Block a user