1 #pragma once 2 3 #include <petscdevice.h> 4 #include <petscpkg_version.h> 5 6 #if defined(__NVCC__) || defined(__CUDACC__) 7 #define PETSC_USING_NVCC 1 8 #endif 9 10 #if PetscDefined(HAVE_CUDA) 11 #include <cuda.h> 12 #include <cuda_runtime.h> 13 #include <cublas_v2.h> 14 #include <cusolverDn.h> 15 #include <cusolverSp.h> 16 #include <cufft.h> 17 18 /* cuBLAS does not have cublasGetErrorName(). We create one on our own. */ 19 PETSC_EXTERN const char *PetscCUBLASGetErrorName(cublasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRCUBLAS macro */ 20 PETSC_EXTERN const char *PetscCUSolverGetErrorName(cusolverStatus_t); 21 PETSC_EXTERN const char *PetscCUFFTGetErrorName(cufftResult); 22 23 /* REMOVE ME */ 24 #define WaitForCUDA() cudaDeviceSynchronize() 25 26 /* CUDART_VERSION = 1000 x major + 10 x minor version */ 27 28 /* Could not find exactly which CUDART_VERSION introduced cudaGetErrorName. At least it was in CUDA 8.0 (Sep. 2016) */ 29 #if PETSC_PKG_CUDA_VERSION_GE(8, 0, 0) 30 #define PetscCallCUDAVoid(...) \ 31 do { \ 32 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 33 PetscCheckAbort(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d (%s) : %s", (PetscErrorCode)_p_cuda_err__, cudaGetErrorName(_p_cuda_err__), cudaGetErrorString(_p_cuda_err__)); \ 34 } while (0) 35 36 #define PetscCallCUDA(...) \ 37 do { \ 38 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 39 PetscCheck(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d (%s) : %s", (PetscErrorCode)_p_cuda_err__, cudaGetErrorName(_p_cuda_err__), cudaGetErrorString(_p_cuda_err__)); \ 40 } while (0) 41 #else /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 42 #define PetscCallCUDA(...) \ 43 do { \ 44 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 45 PetscCheck(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d", (PetscErrorCode)_p_cuda_err__); \ 46 } while (0) 47 48 #define PetscCallCUDAVoid(...) \ 49 do { \ 50 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 51 PetscCheckAbort(_p_cuda_err__ == cudaSuccess, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuda error %d", (PetscErrorCode)_p_cuda_err__); \ 52 } while (0) 53 #endif /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 54 #define CHKERRCUDA(...) PetscCallCUDA(__VA_ARGS__) 55 56 #define PetscCUDACheckLaunch \ 57 do { \ 58 /* Check synchronous errors, i.e. pre-launch */ \ 59 PetscCallCUDA(cudaGetLastError()); \ 60 /* Check asynchronous errors, i.e. kernel failed (ULF) */ \ 61 PetscCallCUDA(cudaDeviceSynchronize()); \ 62 } while (0) 63 64 #define PetscCallCUBLAS(...) \ 65 do { \ 66 const cublasStatus_t _p_cublas_stat__ = __VA_ARGS__; \ 67 if (PetscUnlikely(_p_cublas_stat__ != CUBLAS_STATUS_SUCCESS)) { \ 68 const char *name = PetscCUBLASGetErrorName(_p_cublas_stat__); \ 69 if (((_p_cublas_stat__ == CUBLAS_STATUS_NOT_INITIALIZED) || (_p_cublas_stat__ == CUBLAS_STATUS_ALLOC_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 70 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 71 "cuBLAS error %d (%s). " \ 72 "Reports not initialized or alloc failed; " \ 73 "this indicates the GPU may have run out resources", \ 74 (PetscErrorCode)_p_cublas_stat__, name); \ 75 } else { \ 76 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuBLAS error %d (%s)", (PetscErrorCode)_p_cublas_stat__, name); \ 77 } \ 78 } \ 79 } while (0) 80 #define CHKERRCUBLAS(...) PetscCallCUBLAS(__VA_ARGS__) 81 82 #if (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) /* According to cuda/10.1.168 on OLCF Summit */ 83 #define PetscCallCUSPARSE(...) \ 84 do { \ 85 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \ 86 if (PetscUnlikely(_p_cusparse_stat__)) { \ 87 const char *name = cusparseGetErrorName(_p_cusparse_stat__); \ 88 const char *descr = cusparseGetErrorString(_p_cusparse_stat__); \ 89 PetscCheck((_p_cusparse_stat__ != CUSPARSE_STATUS_NOT_INITIALIZED) && (_p_cusparse_stat__ != CUSPARSE_STATUS_ALLOC_FAILED), PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 90 "cuSPARSE errorcode %d (%s) : %s.; " \ 91 "this indicates the GPU has run out resources", \ 92 (int)_p_cusparse_stat__, name, descr); \ 93 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSPARSE errorcode %d (%s) : %s", (int)_p_cusparse_stat__, name, descr); \ 94 } \ 95 } while (0) 96 #else /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 97 #define PetscCallCUSPARSE(...) \ 98 do { \ 99 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \ 100 PetscCheck(_p_cusparse_stat__ == CUSPARSE_STATUS_SUCCESS, PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSPARSE errorcode %d", (PetscErrorCode)_p_cusparse_stat__); \ 101 } while (0) 102 #endif /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 103 #define CHKERRCUSPARSE(...) PetscCallCUSPARSE(__VA_ARGS__) 104 105 #define PetscCallCUSOLVER(...) \ 106 do { \ 107 const cusolverStatus_t _p_cusolver_stat__ = __VA_ARGS__; \ 108 if (PetscUnlikely(_p_cusolver_stat__ != CUSOLVER_STATUS_SUCCESS)) { \ 109 const char *name = PetscCUSolverGetErrorName(_p_cusolver_stat__); \ 110 if (((_p_cusolver_stat__ == CUSOLVER_STATUS_NOT_INITIALIZED) || (_p_cusolver_stat__ == CUSOLVER_STATUS_ALLOC_FAILED) || (_p_cusolver_stat__ == CUSOLVER_STATUS_INTERNAL_ERROR)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 111 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 112 "cuSolver error %d (%s). " \ 113 "This indicates the GPU may have run out resources", \ 114 (PetscErrorCode)_p_cusolver_stat__, name); \ 115 } else { \ 116 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuSolver error %d (%s)", (PetscErrorCode)_p_cusolver_stat__, name); \ 117 } \ 118 } \ 119 } while (0) 120 #define CHKERRCUSOLVER(...) PetscCallCUSOLVER(__VA_ARGS__) 121 122 #define PetscCallCUFFT(...) \ 123 do { \ 124 const cufftResult_t _p_cufft_stat__ = __VA_ARGS__; \ 125 if (PetscUnlikely(_p_cufft_stat__ != CUFFT_SUCCESS)) { \ 126 const char *name = PetscCUFFTGetErrorName(_p_cufft_stat__); \ 127 if (((_p_cufft_stat__ == CUFFT_SETUP_FAILED) || (_p_cufft_stat__ == CUFFT_ALLOC_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 128 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 129 "cuFFT error %d (%s). " \ 130 "Reports not initialized or alloc failed; " \ 131 "this indicates the GPU has run out resources", \ 132 (PetscErrorCode)_p_cufft_stat__, name); \ 133 } else { \ 134 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuFFT error %d (%s)", (PetscErrorCode)_p_cufft_stat__, name); \ 135 } \ 136 } \ 137 } while (0) 138 #define CHKERRCUFFT(...) PetscCallCUFFT(__VA_ARGS__) 139 140 #define PetscCallCURAND(...) \ 141 do { \ 142 const curandStatus_t _p_curand_stat__ = __VA_ARGS__; \ 143 if (PetscUnlikely(_p_curand_stat__ != CURAND_STATUS_SUCCESS)) { \ 144 if (((_p_curand_stat__ == CURAND_STATUS_INITIALIZATION_FAILED) || (_p_curand_stat__ == CURAND_STATUS_ALLOCATION_FAILED)) && PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 145 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU_RESOURCE, \ 146 "cuRAND error %d. " \ 147 "Reports not initialized or alloc failed; " \ 148 "this indicates the GPU has run out resources", \ 149 (PetscErrorCode)_p_curand_stat__); \ 150 } else { \ 151 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_GPU, "cuRand error %d", (PetscErrorCode)_p_curand_stat__); \ 152 } \ 153 } \ 154 } while (0) 155 #define CHKERRCURAND(...) PetscCallCURAND(__VA_ARGS__) 156 157 PETSC_EXTERN cudaStream_t PetscDefaultCudaStream; // The default stream used by PETSc 158 PETSC_EXTERN PetscErrorCode PetscCUBLASGetHandle(cublasHandle_t *); 159 PETSC_EXTERN PetscErrorCode PetscCUSOLVERDnGetHandle(cusolverDnHandle_t *); 160 PETSC_EXTERN PetscErrorCode PetscGetCurrentCUDAStream(cudaStream_t *); 161 162 #endif // PETSC_HAVE_CUDA 163 164 // these can also be defined in petscdevice_hip.h so we undef and define them *only* if the 165 // current compiler is NVCC. In this case if petscdevice_hip.h is included first, the macros 166 // would already be defined, but they would be empty since we cannot be using HCC at the same 167 // time. 168 #if PetscDefined(USING_NVCC) 169 #undef PETSC_HOST_DECL 170 #undef PETSC_DEVICE_DECL 171 #undef PETSC_KERNEL_DECL 172 #undef PETSC_SHAREDMEM_DECL 173 #undef PETSC_FORCEINLINE 174 #undef PETSC_CONSTMEM_DECL 175 176 #define PETSC_HOST_DECL __host__ 177 #define PETSC_DEVICE_DECL __device__ 178 #define PETSC_KERNEL_DECL __global__ 179 #define PETSC_SHAREDMEM_DECL __shared__ 180 #define PETSC_FORCEINLINE __forceinline__ 181 #define PETSC_CONSTMEM_DECL __constant__ 182 #endif 183 184 #ifndef PETSC_HOST_DECL // use HOST_DECL as canary 185 #define PETSC_HOST_DECL 186 #define PETSC_DEVICE_DECL 187 #define PETSC_KERNEL_DECL 188 #define PETSC_SHAREDMEM_DECL 189 #define PETSC_FORCEINLINE inline 190 #define PETSC_CONSTMEM_DECL 191 #endif 192 193 #ifndef PETSC_DEVICE_DEFINED_DECLS_PRIVATE 194 #define PETSC_DEVICE_DEFINED_DECLS_PRIVATE 195 #define PETSC_HOSTDEVICE_DECL PETSC_HOST_DECL PETSC_DEVICE_DECL 196 #define PETSC_DEVICE_INLINE_DECL PETSC_DEVICE_DECL PETSC_FORCEINLINE 197 #define PETSC_HOSTDEVICE_INLINE_DECL PETSC_HOSTDEVICE_DECL PETSC_FORCEINLINE 198 #endif 199