[HIPIFY] Cooperative groups support

[ROCm/hip commit: aa285d978e]
This commit is contained in:
Evgeny Mankov
2019-08-12 19:20:13 +03:00
والد f4d869c8eb
کامیت bf883d95e9
6فایلهای تغییر یافته به همراه70 افزوده شده و 69 حذف شده
@@ -522,7 +522,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DRIVER_FUNCTION_MAP{
// cudaOccupancyMaxActiveBlocksPerMultiprocessor
{"cuOccupancyMaxActiveBlocksPerMultiprocessor", {"hipOccupancyMaxActiveBlocksPerMultiprocessor", "", CONV_OCCUPANCY, API_DRIVER}},
// cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags
{"cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", {"hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "", CONV_OCCUPANCY, API_DRIVER, HIP_UNSUPPORTED}},
{"cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", {"hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "", CONV_OCCUPANCY, API_DRIVER}},
// cudaOccupancyMaxPotentialBlockSize
{"cuOccupancyMaxPotentialBlockSize", {"hipOccupancyMaxPotentialBlockSize", "", CONV_OCCUPANCY, API_DRIVER}},
// cudaOccupancyMaxPotentialBlockSizeWithFlags
@@ -486,9 +486,9 @@ const std::map<llvm::StringRef, hipCounter> CUDA_DRIVER_TYPE_NAME_MAP{
// no analogue: cudaDevAttrReserved94
{"CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR", {"hipDeviceAttributeCanUseStreamWaitValueNor", "", CONV_NUMERIC_LITERAL, API_DRIVER, HIP_UNSUPPORTED}}, // 94
// cudaDevAttrCooperativeLaunch
{"CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH", {"hipDeviceAttributeCooperativeLaunch", "", CONV_NUMERIC_LITERAL, API_DRIVER, HIP_UNSUPPORTED}}, // 95
{"CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH", {"hipDeviceAttributeCooperativeLaunch", "", CONV_NUMERIC_LITERAL, API_DRIVER}}, // 95
// cudaDevAttrCooperativeMultiDeviceLaunch
{"CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH", {"hipDeviceAttributeCooperativeMultiDeviceLaunch", "", CONV_NUMERIC_LITERAL, API_DRIVER, HIP_UNSUPPORTED}}, // 96
{"CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH", {"hipDeviceAttributeCooperativeMultiDeviceLaunch", "", CONV_NUMERIC_LITERAL, API_DRIVER}}, // 96
// cudaDevAttrMaxSharedMemoryPerBlockOptin
{"CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN", {"hipDeviceAttributeMaxSharedMemoryPerBlockOptin", "", CONV_NUMERIC_LITERAL, API_DRIVER, HIP_UNSUPPORTED}}, // 97
// cudaDevAttrCanFlushRemoteWrites
@@ -194,10 +194,10 @@ const std::map<llvm::StringRef, hipCounter> CUDA_RUNTIME_FUNCTION_MAP{
{"cudaGetParameterBufferV2", {"hipGetParameterBufferV2", "", CONV_EXECUTION, API_RUNTIME, HIP_UNSUPPORTED}},
// no analogue
// NOTE: Not equal to cuLaunchCooperativeKernel due to different signatures
{"cudaLaunchCooperativeKernel", {"hipLaunchCooperativeKernel", "", CONV_EXECUTION, API_RUNTIME, HIP_UNSUPPORTED}},
{"cudaLaunchCooperativeKernel", {"hipLaunchCooperativeKernel", "", CONV_EXECUTION, API_RUNTIME}},
// no analogue
// NOTE: Not equal to cuLaunchCooperativeKernelMultiDevice due to different signatures
{"cudaLaunchCooperativeKernelMultiDevice", {"hipLaunchCooperativeKernelMultiDevice", "", CONV_EXECUTION, API_RUNTIME, HIP_UNSUPPORTED}},
{"cudaLaunchCooperativeKernelMultiDevice", {"hipLaunchCooperativeKernelMultiDevice", "", CONV_EXECUTION, API_RUNTIME}},
// cuLaunchHostFunc
{"cudaLaunchHostFunc", {"hipLaunchHostFunc", "", CONV_EXECUTION, API_RUNTIME, HIP_UNSUPPORTED}},
// no analogue
@@ -212,7 +212,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_RUNTIME_FUNCTION_MAP{
//
{"cudaOccupancyMaxActiveBlocksPerMultiprocessor", {"hipOccupancyMaxActiveBlocksPerMultiprocessor", "", CONV_OCCUPANCY, API_RUNTIME}},
// cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags
{"cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", {"hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "", CONV_OCCUPANCY, API_RUNTIME, HIP_UNSUPPORTED}},
{"cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", {"hipOccupancyMaxActiveBlocksPerMultiprocessorWithFlags", "", CONV_OCCUPANCY, API_RUNTIME}},
// cuOccupancyMaxPotentialBlockSize
{"cudaOccupancyMaxPotentialBlockSize", {"hipOccupancyMaxPotentialBlockSize", "", CONV_OCCUPANCY, API_RUNTIME}},
// cuOccupancyMaxPotentialBlockSizeWithFlags
@@ -84,7 +84,7 @@ const std::map<llvm::StringRef, hipCounter> CUDA_RUNTIME_TYPE_NAME_MAP {
// no analogue
// CUDA_LAUNCH_PARAMS struct differs
{"cudaLaunchParams", {"hipLaunchParams", "", CONV_TYPE, API_RUNTIME, HIP_UNSUPPORTED}},
{"cudaLaunchParams", {"hipLaunchParams", "", CONV_TYPE, API_RUNTIME}},
// no analogue
// NOTE: HIP struct is bigger and contains cudaMemcpy3DParms only in the beginning
@@ -402,9 +402,9 @@ const std::map<llvm::StringRef, hipCounter> CUDA_RUNTIME_TYPE_NAME_MAP {
// CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR
{"cudaDevAttrReserved94", {"hipDeviceAttributeCanUseStreamWaitValueNor", "", CONV_NUMERIC_LITERAL, API_RUNTIME, HIP_UNSUPPORTED}}, // 94
// CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH
{"cudaDevAttrCooperativeLaunch", {"hipDeviceAttributeCooperativeLaunch", "", CONV_NUMERIC_LITERAL, API_RUNTIME, HIP_UNSUPPORTED}}, // 95
{"cudaDevAttrCooperativeLaunch", {"hipDeviceAttributeCooperativeLaunch", "", CONV_NUMERIC_LITERAL, API_RUNTIME}}, // 95
// CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH
{"cudaDevAttrCooperativeMultiDeviceLaunch", {"hipDeviceAttributeCooperativeMultiDeviceLaunch", "", CONV_NUMERIC_LITERAL, API_RUNTIME, HIP_UNSUPPORTED}}, // 96
{"cudaDevAttrCooperativeMultiDeviceLaunch", {"hipDeviceAttributeCooperativeMultiDeviceLaunch", "", CONV_NUMERIC_LITERAL, API_RUNTIME}}, // 96
// CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN
{"cudaDevAttrMaxSharedMemoryPerBlockOptin", {"hipDeviceAttributeMaxSharedMemoryPerBlockOptin", "", CONV_NUMERIC_LITERAL, API_RUNTIME, HIP_UNSUPPORTED}}, // 97
// CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES