openmp/libomptarget/plugins/cuda/dynamic_cuda/cuda.h

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