unsigned int GPUInterface::GetAvailableMemory() {
#if CUDA_VERSION >= 3020
size_t availableMem = 0;
size_t totalMem = 0;
SAFE_CUPP(cuMemGetInfo(&availableMem, &totalMem));
#else
unsigned int availableMem = 0;
unsigned int totalMem = 0;
SAFE_CUPP(cuMemGetInfo(&availableMem, &totalMem));
#endif
return availableMem;
}
void GPUInterface::SynchronizeDevice() {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::SynchronizeDevice\n");
#endif
SAFE_CUPP(cuEventRecord(cudaEvents[numStreams], 0));
SAFE_CUPP(cuStreamWaitEvent(0, cudaEvents[numStreams], 0));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::SynchronizeDevice\n");
#endif
}
void* GPUInterface::AllocatePinnedHostMemory(size_t memSize, bool writeCombined, bool mapped) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::AllocatePinnedHostMemory\n");
#endif
void* ptr;
unsigned int flags = 0;
if (writeCombined)
flags |= CU_MEMHOSTALLOC_WRITECOMBINED;
if (mapped)
flags |= CU_MEMHOSTALLOC_DEVICEMAP;
SAFE_CUPP(cuMemHostAlloc(&ptr, memSize, flags));
#ifdef BEAGLE_DEBUG_VALUES
fprintf(stderr, "Allocated pinned host (CPU) memory %ld to %lu .\n", (long)ptr, ((long)ptr + memSize));
#endif
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::AllocatePinnedHostMemory\n");
#endif
return ptr;
}
void GPUInterface::SynchronizeDeviceWithIndex(int streamRecordIndex, int streamWaitIndex) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::SynchronizeDeviceWithIndex\n");
#endif
CUstream streamRecord = NULL;
CUstream streamWait = NULL;
if (streamRecordIndex >= 0)
streamRecord = cudaStreams[streamRecordIndex % numStreams];
if (streamWaitIndex >= 0)
streamWait = cudaStreams[streamWaitIndex % numStreams];
SAFE_CUPP(cuEventRecord(cudaEvents[numStreams], streamRecord));
SAFE_CUPP(cuStreamWaitEvent(streamWait, cudaEvents[numStreams], 0));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::SynchronizeDeviceWithIndex\n");
#endif
}
void GPUInterface::FreeMemory(GPUPtr dPtr) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::FreeMemory\n");
#endif
SAFE_CUPP(cuMemFree(dPtr));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::FreeMemory\n");
#endif
}
void GPUInterface::FreePinnedHostMemory(void* hPtr) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::FreePinnedHostMemory\n");
#endif
SAFE_CUPP(cuMemFreeHost(hPtr));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::FreePinnedHostMemory\n");
#endif
}
void GPUInterface::SynchronizeHost() {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::SynchronizeHost\n");
#endif
SAFE_CUPP(cuCtxSynchronize());
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::SynchronizeHost\n");
#endif
}
void GPUInterface::MemcpyDeviceToDevice(GPUPtr dest,
GPUPtr src,
size_t memSize) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::MemcpyDeviceToDevice\n");
#endif
SAFE_CUPP(cuMemcpyDtoD(dest, src, memSize));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::MemcpyDeviceToDevice\n");
#endif
}
void GPUInterface::MemsetShort(GPUPtr dest,
unsigned short val,
size_t count) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::MemsetShort\n");
#endif
SAFE_CUPP(cuMemsetD16(dest, val, count));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::MemsetShort\n");
#endif
}
void GPUInterface::MemcpyDeviceToHost(void* dest,
const GPUPtr src,
size_t memSize) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::MemcpyDeviceToHost\n");
#endif
SAFE_CUPP(cuMemcpyDtoHAsync(dest, src, memSize, cudaStreams[0]));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::MemcpyDeviceToHost\n");
#endif
}
GPUPtr GPUInterface::GetDevicePointer(void* hPtr) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::GetDevicePointer\n");
#endif
GPUPtr dPtr;
SAFE_CUPP(cuMemHostGetDevicePointer(&dPtr, hPtr, 0));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::GetDevicePointer\n");
#endif
return dPtr;
}
GPUFunction GPUInterface::GetFunction(const char* functionName) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tEntering GPUInterface::GetFunction\n");
#endif
GPUFunction cudaFunction;
SAFE_CUPP(cuModuleGetFunction(&cudaFunction, cudaModule, functionName));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr,"\t\t\tLeaving GPUInterface::GetFunction\n");
#endif
return cudaFunction;
}
void GPUInterface::MemcpyHostToDevice(GPUPtr dest,
const void* src,
size_t memSize) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::MemcpyHostToDevice\n");
#endif
// SAFE_CUPP(cuMemcpyHtoDAsync(dest, src, memSize, cudaStreams[0]));
SAFE_CUPP(cuMemcpyHtoD(dest, src, memSize));
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::MemcpyHostToDevice\n");
#endif
}
GPUPtr GPUInterface::AllocateIntMemory(size_t length) {
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tEntering GPUInterface::AllocateIntMemory\n");
#endif
GPUPtr ptr;
SAFE_CUPP(cuMemAlloc(&ptr, SIZE_INT * length));
#ifdef BEAGLE_DEBUG_VALUES
fprintf(stderr, "Allocated GPU memory %llu to %llu.\n", (unsigned long long)ptr, (unsigned long long)(ptr + length));
#endif
#ifdef BEAGLE_DEBUG_FLOW
fprintf(stderr, "\t\t\tLeaving GPUInterface::AllocateIntMemory\n");
#endif
return ptr;
}