AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #423 from emankov/cuDNN

Browse Source 
[HIPIFY] Initial cuDNN support

[ROCm/hip commit: 49024a5a55]
This commit is contained in:
Evgeny Mankov
2018-05-04 10:01:47 +03:00
committed by GitHub
parent 1b652107f7 057b13fc4a
commit 7bbca7fa08
7 changed files with 91 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/hip/hipify-clang/src/CUDA2HipMap.cpp
+80
View File
@@ -404,6 +404,9 @@ const std::map <llvm::StringRef, hipCounter> CUDA_INCLUDE_MAP{
{"curand_precalc.h", {"hiprand_kernel.h", CONV_INCLUDE, API_RAND}},
{"curand_uniform.h", {"hiprand_kernel.h", CONV_INCLUDE, API_RAND}},
// CUDNN includes
{"cudnn.h", {"hipDNN.h", CONV_INCLUDE_CUDA_MAIN_H, API_RAND}},
// HIP includes
// TODO: uncomment this when hip/cudacommon.h will be renamed to hip/hipcommon.h
// {"cudacommon.h", {"hipcommon.h", CONV_INCLUDE, API_RUNTIME}},
@@ -2881,6 +2884,83 @@ const std::map<llvm::StringRef, hipCounter> CUDA_IDENTIFIER_MAP{
{"curand_poisson4", {"hiprand_poisson4", CONV_DEVICE_FUNC, API_RAND}},
{"curand_Philox4x32_10", {"hiprand_Philox4x32_10", CONV_DEVICE_FUNC, API_RAND, HIP_UNSUPPORTED}},
// unchanged function names: skipahead, skipahead_sequence, skipahead_subsequence
///////////////////////////// cuDNN /////////////////////////////
{"cudnnContext", {"hipdnnContext", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
{"cudnnHandle_t", {"hipdnnHandle_t", CONV_TYPE, API_DNN}},
{"cudnnStatus_t", {"hipdnnStatus_t", CONV_TYPE, API_DNN}},
{"CUDNN_STATUS_SUCCESS", {"HIPDNN_STATUS_SUCCESS", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
{"CUDNN_STATUS_NOT_INITIALIZED", {"HIPDNN_STATUS_NOT_INITIALIZED", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
{"CUDNN_STATUS_ALLOC_FAILED", {"HIPDNN_STATUS_ALLOC_FAILED", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
{"CUDNN_STATUS_BAD_PARAM", {"HIPDNN_STATUS_BAD_PARAM", CONV_NUMERIC_LITERAL, API_DNN}}, // 3
{"CUDNN_STATUS_INTERNAL_ERROR", {"HIPDNN_STATUS_INTERNAL_ERROR", CONV_NUMERIC_LITERAL, API_DNN}}, // 4
{"CUDNN_STATUS_INVALID_VALUE", {"HIPDNN_STATUS_INVALID_VALUE", CONV_NUMERIC_LITERAL, API_DNN}}, // 5
{"CUDNN_STATUS_ARCH_MISMATCH", {"HIPDNN_STATUS_ARCH_MISMATCH", CONV_NUMERIC_LITERAL, API_DNN}}, // 6
{"CUDNN_STATUS_MAPPING_ERROR", {"HIPDNN_STATUS_MAPPING_ERROR", CONV_NUMERIC_LITERAL, API_DNN}}, // 7
{"CUDNN_STATUS_EXECUTION_FAILED", {"HIPDNN_STATUS_EXECUTION_FAILED", CONV_NUMERIC_LITERAL, API_DNN}}, // 8
{"CUDNN_STATUS_NOT_SUPPORTED", {"HIPDNN_STATUS_NOT_SUPPORTED", CONV_NUMERIC_LITERAL, API_DNN}}, // 9
{"CUDNN_STATUS_LICENSE_ERROR", {"HIPDNN_STATUS_LICENSE_ERROR", CONV_NUMERIC_LITERAL, API_DNN}}, // 10
{"CUDNN_STATUS_RUNTIME_PREREQUISITE_MISSING", {"HIPDNN_STATUS_RUNTIME_PREREQUISITE_MISSING", CONV_NUMERIC_LITERAL, API_DNN}}, // 11
{"CUDNN_STATUS_RUNTIME_IN_PROGRESS", {"HIPDNN_STATUS_RUNTIME_IN_PROGRESS", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 12
{"CUDNN_STATUS_RUNTIME_FP_OVERFLOW", {"HIPDNN_STATUS_RUNTIME_FP_OVERFLOW", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 13
{"cudnnRuntimeTag_t", {"hipdnnRuntimeTag_t", CONV_TYPE, API_DNN, HIP_UNSUPPORTED}},
{"cudnnTensorDescriptor_t", {"hipdnnTensorDescriptor_t", CONV_TYPE, API_DNN}},
{"cudnnConvolutionDescriptor_t", {"hipdnnConvolutionDescriptor_t", CONV_TYPE, API_DNN}},
{"cudnnConvolutionMode_t", {"hipdnnConvolutionMode_t", CONV_TYPE, API_DNN}},
{"CUDNN_CONVOLUTION", {"HIPDNN_CONVOLUTION", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
{"CUDNN_CROSS_CORRELATION", {"HIPDNN_CROSS_CORRELATION", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
{"cudnnTensorFormat_t", {"hipdnnTensorFormat_t", CONV_TYPE, API_DNN}},
{"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
{"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
{"CUDNN_DATA_HALF", {"HIPDNN_DATA_HALF", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
{"CUDNN_DATA_INT8", {"HIPDNN_DATA_INT8", CONV_NUMERIC_LITERAL, API_DNN}}, // 3
{"CUDNN_DATA_INT32", {"HIPDNN_DATA_INT32", CONV_NUMERIC_LITERAL, API_DNN}}, // 4
{"CUDNN_DATA_INT8x4", {"HIPDNN_DATA_INT8x4", CONV_NUMERIC_LITERAL, API_DNN}}, // 5
{"CUDNN_DATA_UINT8", {"HIPDNN_DATA_UINT8", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 6
{"CUDNN_DATA_UINT8x4", {"HIPDNN_DATA_UINT8x4", CONV_NUMERIC_LITERAL, API_DNN, HIP_UNSUPPORTED}}, // 7
{"cudnnConvolutionFwdAlgo_t", {"hipdnnConvolutionFwdAlgo_t", CONV_TYPE, API_DNN}},
{"CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM", {"HIPDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_GEMM", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
{"CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM", {"HIPDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
{"CUDNN_CONVOLUTION_FWD_ALGO_GEMM", {"HIPDNN_CONVOLUTION_FWD_ALGO_GEMM", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
{"CUDNN_CONVOLUTION_FWD_ALGO_DIRECT", {"HIPDNN_CONVOLUTION_FWD_ALGO_DIRECT", CONV_NUMERIC_LITERAL, API_DNN}}, // 3
{"CUDNN_CONVOLUTION_FWD_ALGO_FFT", {"HIPDNN_CONVOLUTION_FWD_ALGO_FFT", CONV_NUMERIC_LITERAL, API_DNN}}, // 4
{"CUDNN_CONVOLUTION_FWD_ALGO_FFT_TILING", {"HIPDNN_CONVOLUTION_FWD_ALGO_FFT_TILING", CONV_NUMERIC_LITERAL, API_DNN}}, // 5
{"CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD", {"HIPDNN_CONVOLUTION_FWD_ALGO_WINOGRAD", CONV_NUMERIC_LITERAL, API_DNN}}, // 6
{"CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED", {"HIPDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED", CONV_NUMERIC_LITERAL, API_DNN}}, // 7
{"CUDNN_CONVOLUTION_FWD_ALGO_COUNT", {"HIPDNN_CONVOLUTION_FWD_ALGO_COUNT", CONV_NUMERIC_LITERAL, API_DNN}}, // 8
{"cudnnConvolutionFwdPreference_t", {"hipdnnConvolutionFwdPreference_t", CONV_TYPE, API_DNN}},
{"CUDNN_CONVOLUTION_FWD_NO_WORKSPACE", {"HIPDNN_CONVOLUTION_FWD_NO_WORKSPACE", CONV_NUMERIC_LITERAL, API_DNN}}, // 0
{"CUDNN_CONVOLUTION_FWD_PREFER_FASTEST", {"HIPDNN_CONVOLUTION_FWD_PREFER_FASTEST", CONV_NUMERIC_LITERAL, API_DNN}}, // 1
{"CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT", {"HIPDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT", CONV_NUMERIC_LITERAL, API_DNN}}, // 2
{"cudnnFilterDescriptor_t", {"hipdnnFilterDescriptor_t", CONV_TYPE, API_DNN}},
{"cudnnGetVersion", {"hipdnnGetVersion", CONV_VERSION, API_DNN}},
{"cudnnGetCudartVersion", {"hipdnnGetCudartVersion", CONV_VERSION, API_DNN, HIP_UNSUPPORTED}},
{"cudnnGetErrorString", {"hipdnnGetErrorString", CONV_ERROR, API_DNN}},
{"cudnnCreate", {"hipdnnCreate", CONV_MATH_FUNC, API_DNN}},
{"cudnnCreateTensorDescriptor", {"hipdnnCreateTensorDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnSetTensor4dDescriptor", {"hipdnnSetTensor4dDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnSetConvolution2dDescriptor", {"hipdnnSetConvolution2dDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnGetConvolution2dForwardOutputDim", {"hipdnnGetConvolution2dForwardOutputDim", CONV_MATH_FUNC, API_DNN}},
{"cudnnCreateFilterDescriptor", {"hipdnnCreateFilterDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnSetFilter4dDescriptor", {"hipdnnSetFilter4dDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnCreateConvolutionDescriptor", {"hipdnnCreateConvolutionDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnGetConvolutionForwardAlgorithm", {"hipdnnGetConvolutionForwardAlgorithm", CONV_MATH_FUNC, API_DNN}},
{"cudnnConvolutionForward", {"hipdnnConvolutionForward", CONV_MATH_FUNC, API_DNN}},
{"cudnnGetConvolutionForwardWorkspaceSize", {"hipdnnGetConvolutionForwardWorkspaceSize", CONV_MATH_FUNC, API_DNN}},
{"cudnnDestroyTensorDescriptor", {"hipdnnDestroyTensorDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnDestroyConvolutionDescriptor", {"hipdnnDestroyConvolutionDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnDestroyFilterDescriptor", {"hipdnnDestroyFilterDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnDestroyFilterDescriptor", {"hipdnnDestroyFilterDescriptor", CONV_MATH_FUNC, API_DNN}},
{"cudnnDestroy", {"hipdnnDestroy", CONV_MATH_FUNC, API_DNN}},
};
const std::map<llvm::StringRef, hipCounter>& CUDA_RENAMES_MAP() {
projects/hip/hipify-clang/src/HipifyAction.cpp
+4
View File
@@ -160,6 +160,10 @@ bool HipifyAction::Exclude(const hipCounter & hipToken) {
insertedRANDHeader = true;
return false;
}
case API_DNN:
if (insertedDNNHeader) { return true; }
insertedDNNHeader = true;
return false;
default:
return false;
}
projects/hip/hipify-clang/src/HipifyAction.h
+1
View File
@@ -27,6 +27,7 @@ private:
bool insertedBLASHeader = false;
bool insertedRANDHeader = false;
bool insertedRAND_kernelHeader = false;
bool insertedDNNHeader = false;
bool firstHeader = false;
bool pragmaOnce = false;
clang::SourceLocation firstHeaderLoc;
projects/hip/hipify-clang/src/Statistics.cpp
+1 -1
View File
@@ -14,7 +14,7 @@ const char *counterNames[NUM_CONV_TYPES] = {
};
const char *apiNames[NUM_API_TYPES] = {
"CUDA Driver API", "CUDA RT API", "CUBLAS API", "CURAND API"
"CUDA Driver API", "CUDA RT API", "CUBLAS API", "CURAND API", "CUDNN API"
};
namespace {
projects/hip/hipify-clang/src/Statistics.h
+1
View File
@@ -56,6 +56,7 @@ enum ApiTypes {
API_RUNTIME,
API_BLAS,
API_RAND,
API_DNN,
API_LAST
};
constexpr int NUM_API_TYPES = (int) ApiTypes::API_LAST;
projects/hip/tests/hipify-clang/lit.cfg
+3 -1
View File
@@ -57,6 +57,8 @@ else:
run_test_ext = ".sh"
clang_args += " -isystem'%s'/samples/common/inc"
config.substitutions.append(("%cuda_args", clang_args % (config.cuda_root, config.cuda_sdk_root)))
clang_args += " -I'%s'/include"
config.substitutions.append(("%cuda_args", clang_args % (config.cuda_root, config.cuda_sdk_root, config.cuda_dnn_root)))
config.substitutions.append(("hipify", '"' + hipify_path + "/hipify-clang" + '"'))
config.substitutions.append(("%run_test", '"' + config.test_source_root + "/run_test" + run_test_ext + '"'))
projects/hip/tests/hipify-clang/lit.site.cfg.in
+1
View File
@@ -4,6 +4,7 @@ import os
config.llvm_tools_dir = "@LLVM_TOOLS_BINARY_DIR@"
config.obj_root = "@CMAKE_CURRENT_BINARY_DIR@"
config.cuda_root = "@CUDA_TOOLKIT_ROOT_DIR@"
config.cuda_dnn_root = "@CUDA_DNN_ROOT_DIR@"
if sys.platform in ['win32']:
config.cuda_sdk_root = "@CUDA_SDK_ROOT_DIR@"
if not config.cuda_sdk_root or config.cuda_sdk_root == "CUDA_SDK_ROOT_DIR-NOTFOUND":