1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
|
#include "caffe2/core/common_gpu.h"
#include <atomic>
#include <cstdlib>
#include <iostream>
#include <sstream>
#include <c10/cuda/CUDAFunctions.h>
#include "caffe2/core/common.h"
#include "caffe2/core/init.h"
#include "caffe2/core/logging.h"
namespace caffe2 {
int NumCudaDevices() {
if (getenv("CAFFE2_DEBUG_CUDA_INIT_ORDER")) {
static bool first = true;
if (first) {
first = false;
std::cerr << "DEBUG: caffe2::NumCudaDevices() invoked for the first time"
<< std::endl;
}
}
// It logs warnings on first run
return c10::cuda::device_count();
}
namespace {
int gDefaultGPUID = 0;
} // namespace
void SetDefaultGPUID(const int deviceid) {
CAFFE_ENFORCE_LT(
deviceid,
NumCudaDevices(),
"The default gpu id should be smaller than the number of gpus "
"on this machine: ",
deviceid,
" vs ",
NumCudaDevices());
gDefaultGPUID = deviceid;
}
int GetDefaultGPUID() { return gDefaultGPUID; }
int CaffeCudaGetDevice() {
int gpu_id = 0;
CUDA_ENFORCE(cudaGetDevice(&gpu_id));
return gpu_id;
}
void CaffeCudaSetDevice(const int id) {
CUDA_ENFORCE(cudaSetDevice(id));
}
int GetGPUIDForPointer(const void* ptr) {
cudaPointerAttributes attr;
cudaError_t err = cudaPointerGetAttributes(&attr, ptr);
if (err == cudaErrorInvalidValue) {
// Occurs when the pointer is in the CPU address space that is
// unmanaged by CUDA; make sure the last error state is cleared,
// since it is persistent
err = cudaGetLastError();
CHECK(err == cudaErrorInvalidValue);
return -1;
}
// Otherwise, there must be no error
CUDA_ENFORCE(err);
if (attr.CAFFE2_CUDA_PTRATTR_MEMTYPE == cudaMemoryTypeHost) {
return -1;
}
return attr.device;
}
struct CudaDevicePropWrapper {
CudaDevicePropWrapper() : props(NumCudaDevices()) {
for (int i = 0; i < NumCudaDevices(); ++i) {
CUDA_ENFORCE(cudaGetDeviceProperties(&props[i], i));
}
}
vector<cudaDeviceProp> props;
};
const cudaDeviceProp& GetDeviceProperty(const int deviceid) {
// According to C++11 standard section 6.7, static local variable init is
// thread safe. See
// https://stackoverflow.com/questions/8102125/is-local-static-variable-initialization-thread-safe-in-c11
// for details.
static CudaDevicePropWrapper props;
CAFFE_ENFORCE_LT(
deviceid,
NumCudaDevices(),
"The gpu id should be smaller than the number of gpus ",
"on this machine: ",
deviceid,
" vs ",
NumCudaDevices());
return props.props[deviceid];
}
void DeviceQuery(const int device) {
const cudaDeviceProp& prop = GetDeviceProperty(device);
std::stringstream ss;
ss << std::endl;
ss << "Device id: " << device << std::endl;
ss << "Major revision number: " << prop.major << std::endl;
ss << "Minor revision number: " << prop.minor << std::endl;
ss << "Name: " << prop.name << std::endl;
ss << "Total global memory: " << prop.totalGlobalMem << std::endl;
ss << "Total shared memory per block: " << prop.sharedMemPerBlock
<< std::endl;
ss << "Total registers per block: " << prop.regsPerBlock << std::endl;
ss << "Warp size: " << prop.warpSize << std::endl;
#if !defined(USE_ROCM)
ss << "Maximum memory pitch: " << prop.memPitch << std::endl;
#endif
ss << "Maximum threads per block: " << prop.maxThreadsPerBlock
<< std::endl;
ss << "Maximum dimension of block: "
<< prop.maxThreadsDim[0] << ", " << prop.maxThreadsDim[1] << ", "
<< prop.maxThreadsDim[2] << std::endl;
ss << "Maximum dimension of grid: "
<< prop.maxGridSize[0] << ", " << prop.maxGridSize[1] << ", "
<< prop.maxGridSize[2] << std::endl;
ss << "Clock rate: " << prop.clockRate << std::endl;
ss << "Total constant memory: " << prop.totalConstMem << std::endl;
#if !defined(USE_ROCM)
ss << "Texture alignment: " << prop.textureAlignment << std::endl;
ss << "Concurrent copy and execution: "
<< (prop.deviceOverlap ? "Yes" : "No") << std::endl;
#endif
ss << "Number of multiprocessors: " << prop.multiProcessorCount
<< std::endl;
#if !defined(USE_ROCM)
ss << "Kernel execution timeout: "
<< (prop.kernelExecTimeoutEnabled ? "Yes" : "No") << std::endl;
#endif
LOG(INFO) << ss.str();
return;
}
bool GetCudaPeerAccessPattern(vector<vector<bool> >* pattern) {
int gpu_count;
if (cudaGetDeviceCount(&gpu_count) != cudaSuccess) return false;
pattern->clear();
pattern->resize(gpu_count, vector<bool>(gpu_count, false));
for (int i = 0; i < gpu_count; ++i) {
for (int j = 0; j < gpu_count; ++j) {
int can_access = true;
if (i != j) {
if (cudaDeviceCanAccessPeer(&can_access, i, j)
!= cudaSuccess) {
return false;
}
}
(*pattern)[i][j] = static_cast<bool>(can_access);
}
}
return true;
}
bool TensorCoreAvailable() {
int device = CaffeCudaGetDevice();
auto& prop = GetDeviceProperty(device);
return prop.major >= 7;
}
const char* cublasGetErrorString(cublasStatus_t error) {
switch (error) {
case CUBLAS_STATUS_SUCCESS:
return "CUBLAS_STATUS_SUCCESS";
case CUBLAS_STATUS_NOT_INITIALIZED:
return "CUBLAS_STATUS_NOT_INITIALIZED";
case CUBLAS_STATUS_ALLOC_FAILED:
return "CUBLAS_STATUS_ALLOC_FAILED";
case CUBLAS_STATUS_INVALID_VALUE:
return "CUBLAS_STATUS_INVALID_VALUE";
case CUBLAS_STATUS_ARCH_MISMATCH:
return "CUBLAS_STATUS_ARCH_MISMATCH";
case CUBLAS_STATUS_INTERNAL_ERROR:
return "CUBLAS_STATUS_INTERNAL_ERROR";
#if !defined(USE_ROCM)
case CUBLAS_STATUS_MAPPING_ERROR:
return "CUBLAS_STATUS_MAPPING_ERROR";
case CUBLAS_STATUS_EXECUTION_FAILED:
return "CUBLAS_STATUS_EXECUTION_FAILED";
case CUBLAS_STATUS_NOT_SUPPORTED:
return "CUBLAS_STATUS_NOT_SUPPORTED";
case CUBLAS_STATUS_LICENSE_ERROR:
return "CUBLAS_STATUS_LICENSE_ERROR";
#else
case rocblas_status_invalid_size:
return "rocblas_status_invalid_size";
case rocblas_status_perf_degraded:
return "rocblas_status_perf_degraded";
case rocblas_status_size_query_mismatch:
return "rocblas_status_size_query_mismatch";
case rocblas_status_size_increased:
return "rocblas_status_size_increased";
case rocblas_status_size_unchanged:
return "rocblas_status_size_unchanged";
default:
return "unrecognized_rocblas_error";
#endif
}
// To suppress compiler warning.
return "Unrecognized cublas error string";
}
const char* curandGetErrorString(curandStatus_t error) {
switch (error) {
case CURAND_STATUS_SUCCESS:
return "CURAND_STATUS_SUCCESS";
case CURAND_STATUS_VERSION_MISMATCH:
return "CURAND_STATUS_VERSION_MISMATCH";
case CURAND_STATUS_NOT_INITIALIZED:
return "CURAND_STATUS_NOT_INITIALIZED";
case CURAND_STATUS_ALLOCATION_FAILED:
return "CURAND_STATUS_ALLOCATION_FAILED";
case CURAND_STATUS_TYPE_ERROR:
return "CURAND_STATUS_TYPE_ERROR";
case CURAND_STATUS_OUT_OF_RANGE:
return "CURAND_STATUS_OUT_OF_RANGE";
case CURAND_STATUS_LENGTH_NOT_MULTIPLE:
return "CURAND_STATUS_LENGTH_NOT_MULTIPLE";
case CURAND_STATUS_DOUBLE_PRECISION_REQUIRED:
return "CURAND_STATUS_DOUBLE_PRECISION_REQUIRED";
case CURAND_STATUS_LAUNCH_FAILURE:
return "CURAND_STATUS_LAUNCH_FAILURE";
case CURAND_STATUS_PREEXISTING_FAILURE:
return "CURAND_STATUS_PREEXISTING_FAILURE";
case CURAND_STATUS_INITIALIZATION_FAILED:
return "CURAND_STATUS_INITIALIZATION_FAILED";
case CURAND_STATUS_ARCH_MISMATCH:
return "CURAND_STATUS_ARCH_MISMATCH";
case CURAND_STATUS_INTERNAL_ERROR:
return "CURAND_STATUS_INTERNAL_ERROR";
#if defined(USE_ROCM)
case HIPRAND_STATUS_NOT_IMPLEMENTED:
return "HIPRAND_STATUS_NOT_IMPLEMENTED";
#endif
}
// To suppress compiler warning.
return "Unrecognized curand error string";
}
// Turn on the flag g_caffe2_has_cuda_linked to true for HasCudaRuntime()
// function.
namespace {
class CudaRuntimeFlagFlipper {
public:
CudaRuntimeFlagFlipper() {
internal::SetCudaRuntimeFlag();
}
};
static CudaRuntimeFlagFlipper g_flipper;
} // namespace
} // namespace caffe2
|