|
| 1 | +#include "cuda_compat.h" |
| 2 | + |
| 3 | +namespace ModuleBase { |
| 4 | +namespace cuda_compat { |
| 5 | + |
| 6 | +//--------------------------------------------------------------------------- |
| 7 | +// Implementation of printDeprecatedDeviceInfo and printComputeModeInfo |
| 8 | +//--------------------------------------------------------------------------- |
| 9 | +void printDeprecatedDeviceInfo(std::ostream& ofs_device, const cudaDeviceProp& deviceProp) |
| 10 | +{ |
| 11 | +#if defined(CUDA_VERSION) && CUDA_VERSION < 13000 |
| 12 | + char msg[1024]; |
| 13 | + sprintf(msg, |
| 14 | + " GPU Max Clock rate: %.0f MHz (%0.2f " |
| 15 | + "GHz)\n", |
| 16 | + deviceProp.clockRate * 1e-3f, deviceProp.clockRate * 1e-6f); |
| 17 | + ofs_device << msg << std::endl; |
| 18 | + // This is supported in CUDA 5.0 (runtime API device properties) |
| 19 | + sprintf(msg, " Memory Clock rate: %.0f Mhz\n", |
| 20 | + deviceProp.memoryClockRate * 1e-3f); |
| 21 | + ofs_device << msg << std::endl; |
| 22 | + |
| 23 | + sprintf(msg, " Memory Bus Width: %d-bit\n", |
| 24 | + deviceProp.memoryBusWidth); |
| 25 | + ofs_device << msg << std::endl; |
| 26 | + |
| 27 | + sprintf(msg, |
| 28 | + " Concurrent copy and kernel execution: %s with %d copy " |
| 29 | + "engine(s)\n", |
| 30 | + (deviceProp.deviceOverlap ? "Yes" : "No"), |
| 31 | + deviceProp.asyncEngineCount); |
| 32 | + ofs_device << msg << std::endl; |
| 33 | + sprintf(msg, " Run time limit on kernels: %s\n", |
| 34 | + deviceProp.kernelExecTimeoutEnabled ? "Yes" : "No"); |
| 35 | + ofs_device << msg << std::endl; |
| 36 | +#endif |
| 37 | +} |
| 38 | + |
| 39 | +void printComputeModeInfo(std::ostream& ofs_device, const cudaDeviceProp& deviceProp) |
| 40 | +{ |
| 41 | +#if defined(CUDA_VERSION) && CUDA_VERSION < 13000 |
| 42 | + char msg[1024]; |
| 43 | + sprintf(msg, " Supports MultiDevice Co-op Kernel Launch: %s\n", |
| 44 | + deviceProp.cooperativeMultiDeviceLaunch ? "Yes" : "No"); |
| 45 | + ofs_device << msg << std::endl; |
| 46 | + |
| 47 | + const char *sComputeMode[] = { |
| 48 | + "Default (multiple host threads can use ::cudaSetDevice() with device " |
| 49 | + "simultaneously)", |
| 50 | + "Exclusive (only one host thread in one process is able to use " |
| 51 | + "::cudaSetDevice() with this device)", |
| 52 | + "Prohibited (no host thread can use ::cudaSetDevice() with this " |
| 53 | + "device)", |
| 54 | + "Exclusive Process (many threads in one process is able to use " |
| 55 | + "::cudaSetDevice() with this device)", |
| 56 | + "Unknown", |
| 57 | + NULL}; |
| 58 | + sprintf(msg, " Compute Mode:\n"); |
| 59 | + ofs_device << msg << std::endl; |
| 60 | + ofs_device << " " << sComputeMode[deviceProp.computeMode] << std::endl |
| 61 | + << std::endl; |
| 62 | +#endif |
| 63 | +} |
| 64 | + |
| 65 | +//------------------------------------------------------------------------------------------------- |
| 66 | +// Implementation of cufftGetErrorStringCompat |
| 67 | +//------------------------------------------------------------------------------------------------- |
| 68 | +const char* cufftGetErrorStringCompat(cufftResult_t error) |
| 69 | +{ |
| 70 | + switch (error) |
| 71 | + { |
| 72 | + case CUFFT_SUCCESS: |
| 73 | + return "CUFFT_SUCCESS"; |
| 74 | + case CUFFT_INVALID_PLAN: |
| 75 | + return "CUFFT_INVALID_PLAN"; |
| 76 | + case CUFFT_ALLOC_FAILED: |
| 77 | + return "CUFFT_ALLOC_FAILED"; |
| 78 | + case CUFFT_INVALID_TYPE: |
| 79 | + return "CUFFT_INVALID_TYPE"; |
| 80 | + case CUFFT_INVALID_VALUE: |
| 81 | + return "CUFFT_INVALID_VALUE"; |
| 82 | + case CUFFT_INTERNAL_ERROR: |
| 83 | + return "CUFFT_INTERNAL_ERROR"; |
| 84 | + case CUFFT_EXEC_FAILED: |
| 85 | + return "CUFFT_EXEC_FAILED"; |
| 86 | + case CUFFT_SETUP_FAILED: |
| 87 | + return "CUFFT_SETUP_FAILED"; |
| 88 | + case CUFFT_INVALID_SIZE: |
| 89 | + return "CUFFT_INVALID_SIZE"; |
| 90 | + case CUFFT_UNALIGNED_DATA: |
| 91 | + return "CUFFT_UNALIGNED_DATA"; |
| 92 | + case CUFFT_INVALID_DEVICE: |
| 93 | + return "CUFFT_INVALID_DEVICE"; |
| 94 | + case CUFFT_NO_WORKSPACE: |
| 95 | + return "CUFFT_NO_WORKSPACE"; |
| 96 | + case CUFFT_NOT_IMPLEMENTED: |
| 97 | + return "CUFFT_NOT_IMPLEMENTED"; |
| 98 | + case CUFFT_NOT_SUPPORTED: |
| 99 | + return "CUFFT_NOT_SUPPORTED"; |
| 100 | + |
| 101 | +#if defined(CUDA_VERSION) && CUDA_VERSION < 13000 |
| 102 | + case CUFFT_INCOMPLETE_PARAMETER_LIST: |
| 103 | + return "CUFFT_INCOMPLETE_PARAMETER_LIST"; |
| 104 | + case CUFFT_PARSE_ERROR: |
| 105 | + return "CUFFT_PARSE_ERROR"; |
| 106 | + case CUFFT_LICENSE_ERROR: |
| 107 | + return "CUFFT_LICENSE_ERROR"; |
| 108 | +#endif |
| 109 | + |
| 110 | + default: |
| 111 | + return "<unknown>"; |
| 112 | + } |
| 113 | +} |
| 114 | + |
| 115 | +} // namespace cuda_compat |
| 116 | +} // namespace ModuleBase |
0 commit comments