1 //===--- cuda/dynamic_cuda/cuda.h --------------------------------- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // The parts of the cuda api that are presently in use by the openmp cuda plugin 10 // 11 //===----------------------------------------------------------------------===// 12 13 #ifndef DYNAMIC_CUDA_CUDA_H_INCLUDED 14 #define DYNAMIC_CUDA_CUDA_H_INCLUDED 15 16 #include <cstddef> 17 #include <cstdint> 18 19 typedef int CUdevice; 20 typedef uintptr_t CUdeviceptr; 21 typedef struct CUmod_st *CUmodule; 22 typedef struct CUctx_st *CUcontext; 23 typedef struct CUfunc_st *CUfunction; 24 typedef struct CUstream_st *CUstream; 25 typedef struct CUevent_st *CUevent; 26 27 typedef enum cudaError_enum { 28 CUDA_SUCCESS = 0, 29 CUDA_ERROR_INVALID_VALUE = 1, 30 CUDA_ERROR_NO_DEVICE = 100, 31 CUDA_ERROR_INVALID_HANDLE = 400, 32 } CUresult; 33 34 typedef enum CUstream_flags_enum { 35 CU_STREAM_DEFAULT = 0x0, 36 CU_STREAM_NON_BLOCKING = 0x1, 37 } CUstream_flags; 38 39 typedef enum CUlimit_enum { 40 CU_LIMIT_STACK_SIZE = 0x0, 41 CU_LIMIT_PRINTF_FIFO_SIZE = 0x1, 42 CU_LIMIT_MALLOC_HEAP_SIZE = 0x2, 43 CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH = 0x3, 44 CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 0x4, 45 CU_LIMIT_MAX_L2_FETCH_GRANULARITY = 0x5, 46 CU_LIMIT_PERSISTING_L2_CACHE_SIZE = 0x6, 47 CU_LIMIT_MAX 48 } CUlimit; 49 50 typedef enum CUdevice_attribute_enum { 51 CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK = 1, 52 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X = 2, 53 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y = 3, 54 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z = 4, 55 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X = 5, 56 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y = 6, 57 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z = 7, 58 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK = 8, 59 CU_DEVICE_ATTRIBUTE_SHARED_MEMORY_PER_BLOCK = 8, 60 CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY = 9, 61 CU_DEVICE_ATTRIBUTE_WARP_SIZE = 10, 62 CU_DEVICE_ATTRIBUTE_MAX_PITCH = 11, 63 CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK = 12, 64 CU_DEVICE_ATTRIBUTE_REGISTERS_PER_BLOCK = 12, 65 CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13, 66 CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT = 14, 67 CU_DEVICE_ATTRIBUTE_GPU_OVERLAP = 15, 68 CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16, 69 CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT = 17, 70 CU_DEVICE_ATTRIBUTE_INTEGRATED = 18, 71 CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY = 19, 72 CU_DEVICE_ATTRIBUTE_COMPUTE_MODE = 20, 73 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH = 21, 74 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH = 22, 75 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT = 23, 76 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH = 24, 77 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT = 25, 78 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH = 26, 79 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH = 27, 80 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT = 28, 81 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS = 29, 82 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_WIDTH = 27, 83 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_HEIGHT = 28, 84 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_NUMSLICES = 29, 85 CU_DEVICE_ATTRIBUTE_SURFACE_ALIGNMENT = 30, 86 CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS = 31, 87 CU_DEVICE_ATTRIBUTE_ECC_ENABLED = 32, 88 CU_DEVICE_ATTRIBUTE_PCI_BUS_ID = 33, 89 CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID = 34, 90 CU_DEVICE_ATTRIBUTE_TCC_DRIVER = 35, 91 CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36, 92 CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH = 37, 93 CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE = 38, 94 CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39, 95 CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT = 40, 96 CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING = 41, 97 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH = 42, 98 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS = 43, 99 CU_DEVICE_ATTRIBUTE_CAN_TEX2D_GATHER = 44, 100 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_WIDTH = 45, 101 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_HEIGHT = 46, 102 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE = 47, 103 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE = 48, 104 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE = 49, 105 CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID = 50, 106 CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT = 51, 107 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_WIDTH = 52, 108 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH = 53, 109 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS = 54, 110 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_WIDTH = 55, 111 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_WIDTH = 56, 112 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_HEIGHT = 57, 113 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_WIDTH = 58, 114 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_HEIGHT = 59, 115 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_DEPTH = 60, 116 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_WIDTH = 61, 117 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_LAYERS = 62, 118 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_WIDTH = 63, 119 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_HEIGHT = 64, 120 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_LAYERS = 65, 121 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_WIDTH = 66, 122 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_WIDTH = 67, 123 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_LAYERS = 68, 124 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH = 69, 125 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH = 70, 126 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT = 71, 127 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH = 72, 128 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH = 73, 129 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT = 74, 130 CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75, 131 CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76, 132 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH = 77, 133 CU_DEVICE_ATTRIBUTE_STREAM_PRIORITIES_SUPPORTED = 78, 134 CU_DEVICE_ATTRIBUTE_GLOBAL_L1_CACHE_SUPPORTED = 79, 135 CU_DEVICE_ATTRIBUTE_LOCAL_L1_CACHE_SUPPORTED = 80, 136 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR = 81, 137 CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR = 82, 138 CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY = 83, 139 CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD = 84, 140 CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD_GROUP_ID = 85, 141 CU_DEVICE_ATTRIBUTE_HOST_NATIVE_ATOMIC_SUPPORTED = 86, 142 CU_DEVICE_ATTRIBUTE_SINGLE_TO_DOUBLE_PRECISION_PERF_RATIO = 87, 143 CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS = 88, 144 CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS = 89, 145 CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED = 90, 146 CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM = 91, 147 CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS = 92, 148 CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS = 93, 149 CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR = 94, 150 CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH = 95, 151 CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH = 96, 152 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN = 97, 153 CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES = 98, 154 CU_DEVICE_ATTRIBUTE_HOST_REGISTER_SUPPORTED = 99, 155 CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES = 100, 156 CU_DEVICE_ATTRIBUTE_DIRECT_MANAGED_MEM_ACCESS_FROM_HOST = 101, 157 CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED = 102, 158 CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED = 102, 159 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR_SUPPORTED = 103, 160 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_HANDLE_SUPPORTED = 104, 161 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_KMT_HANDLE_SUPPORTED = 105, 162 CU_DEVICE_ATTRIBUTE_MAX_BLOCKS_PER_MULTIPROCESSOR = 106, 163 CU_DEVICE_ATTRIBUTE_GENERIC_COMPRESSION_SUPPORTED = 107, 164 CU_DEVICE_ATTRIBUTE_MAX_PERSISTING_L2_CACHE_SIZE = 108, 165 CU_DEVICE_ATTRIBUTE_MAX_ACCESS_POLICY_WINDOW_SIZE = 109, 166 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED = 110, 167 CU_DEVICE_ATTRIBUTE_RESERVED_SHARED_MEMORY_PER_BLOCK = 111, 168 CU_DEVICE_ATTRIBUTE_SPARSE_CUDA_ARRAY_SUPPORTED = 112, 169 CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED = 113, 170 CU_DEVICE_ATTRIBUTE_TIMELINE_SEMAPHORE_INTEROP_SUPPORTED = 114, 171 CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED = 115, 172 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED = 116, 173 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS = 117, 174 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING = 118, 175 CU_DEVICE_ATTRIBUTE_MEMPOOL_SUPPORTED_HANDLE_TYPES = 119, 176 CU_DEVICE_ATTRIBUTE_MAX, 177 } CUdevice_attribute; 178 179 typedef enum CUfunction_attribute_enum { 180 CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK = 0, 181 } CUfunction_attribute; 182 183 typedef enum CUctx_flags_enum { 184 CU_CTX_SCHED_BLOCKING_SYNC = 0x04, 185 CU_CTX_SCHED_MASK = 0x07, 186 } CUctx_flags; 187 188 typedef enum CUmemAttach_flags_enum { 189 CU_MEM_ATTACH_GLOBAL = 0x1, 190 CU_MEM_ATTACH_HOST = 0x2, 191 CU_MEM_ATTACH_SINGLE = 0x4, 192 } CUmemAttach_flags; 193 194 typedef enum CUcomputeMode_enum { 195 CU_COMPUTEMODE_DEFAULT = 0, 196 CU_COMPUTEMODE_PROHIBITED = 2, 197 CU_COMPUTEMODE_EXCLUSIVE_PROCESS = 3, 198 } CUcompute_mode; 199 200 typedef enum CUevent_flags_enum { 201 CU_EVENT_DEFAULT = 0x0, 202 CU_EVENT_BLOCKING_SYNC = 0x1, 203 CU_EVENT_DISABLE_TIMING = 0x2, 204 CU_EVENT_INTERPROCESS = 0x4 205 } CUevent_flags; 206 207 CUresult cuCtxGetDevice(CUdevice *); 208 CUresult cuDeviceGet(CUdevice *, int); 209 CUresult cuDeviceGetAttribute(int *, CUdevice_attribute, CUdevice); 210 CUresult cuDeviceGetCount(int *); 211 CUresult cuFuncGetAttribute(int *, CUfunction_attribute, CUfunction); 212 213 // Device info 214 CUresult cuDeviceGetName(char *, int, CUdevice); 215 CUresult cuDeviceTotalMem(size_t *, CUdevice); 216 CUresult cuDriverGetVersion(int *); 217 218 CUresult cuGetErrorString(CUresult, const char **); 219 CUresult cuInit(unsigned); 220 CUresult cuLaunchKernel(CUfunction, unsigned, unsigned, unsigned, unsigned, 221 unsigned, unsigned, unsigned, CUstream, void **, 222 void **); 223 224 CUresult cuMemAlloc(CUdeviceptr *, size_t); 225 CUresult cuMemAllocHost(void **, size_t); 226 CUresult cuMemAllocManaged(CUdeviceptr *, size_t, unsigned int); 227 228 CUresult cuMemcpyDtoDAsync(CUdeviceptr, CUdeviceptr, size_t, CUstream); 229 CUresult cuMemcpyDtoH(void *, CUdeviceptr, size_t); 230 CUresult cuMemcpyDtoHAsync(void *, CUdeviceptr, size_t, CUstream); 231 CUresult cuMemcpyHtoD(CUdeviceptr, const void *, size_t); 232 CUresult cuMemcpyHtoDAsync(CUdeviceptr, const void *, size_t, CUstream); 233 234 CUresult cuMemFree(CUdeviceptr); 235 CUresult cuMemFreeHost(void *); 236 237 CUresult cuModuleGetFunction(CUfunction *, CUmodule, const char *); 238 CUresult cuModuleGetGlobal(CUdeviceptr *, size_t *, CUmodule, const char *); 239 240 CUresult cuModuleUnload(CUmodule); 241 CUresult cuStreamCreate(CUstream *, unsigned); 242 CUresult cuStreamDestroy(CUstream); 243 CUresult cuStreamSynchronize(CUstream); 244 CUresult cuCtxSetCurrent(CUcontext); 245 CUresult cuDevicePrimaryCtxRelease(CUdevice); 246 CUresult cuDevicePrimaryCtxGetState(CUdevice, unsigned *, int *); 247 CUresult cuDevicePrimaryCtxSetFlags(CUdevice, unsigned); 248 CUresult cuDevicePrimaryCtxRetain(CUcontext *, CUdevice); 249 CUresult cuModuleLoadDataEx(CUmodule *, const void *, unsigned, void *, 250 void **); 251 252 CUresult cuDeviceCanAccessPeer(int *, CUdevice, CUdevice); 253 CUresult cuCtxEnablePeerAccess(CUcontext, unsigned); 254 CUresult cuMemcpyPeerAsync(CUdeviceptr, CUcontext, CUdeviceptr, CUcontext, 255 size_t, CUstream); 256 257 CUresult cuCtxGetLimit(size_t *, CUlimit); 258 CUresult cuCtxSetLimit(CUlimit, size_t); 259 260 CUresult cuEventCreate(CUevent *, unsigned int); 261 CUresult cuEventRecord(CUevent, CUstream); 262 CUresult cuStreamWaitEvent(CUstream, CUevent, unsigned int); 263 CUresult cuEventSynchronize(CUevent); 264 CUresult cuEventDestroy(CUevent); 265 266 #endif 267