void GPUDataTransferer<ElemType>::CopyCPUToGPUAsync(ElemType* cpuBuffer, size_t numElements, ElemType* gpuBuffer)
{
PrepareDevice(m_deviceId);
cudaMemcpyAsync(gpuBuffer, cpuBuffer, numElements * sizeof(ElemType), cudaMemcpyHostToDevice, m_assignStream) || "cudaMemcpyAsync failed";
cudaEventRecord(m_assignCompleteEvent, m_assignStream) || "cudaEventRecord failed";
}
void GPUDataTransferer<ElemType>::CopyGPUToCPUAsync(ElemType* gpuBuffer, size_t numElements, ElemType* cpuBuffer)
{
PrepareDevice(m_deviceId);
cudaMemcpyAsync(cpuBuffer, gpuBuffer, numElements * sizeof(ElemType), cudaMemcpyDeviceToHost, m_fetchStream) || "cudaMemcpyAsync failed";
cudaEventRecord(m_fetchCompleteEvent, m_fetchStream) || "cudaEventRecord failed";
}
GPUDataTransferer<ElemType>::GPUDataTransferer(int deviceId, bool useConcurrentStreams)
: m_deviceId(deviceId)
{
PrepareDevice(m_deviceId);
// events
// Note: Do NOT use cudaEventBlockingSync (which supposedly yields the process)--it will totally break cudaEventSynchronize(), causing it to take 50 or 100 ms randomly.
cudaEventCreateWithFlags(&m_fetchCompleteEvent, cudaEventDisableTiming) || "cudaEventCreateWithFlags failed";
cudaEventCreateWithFlags(&m_assignCompleteEvent, cudaEventDisableTiming) || "cudaEventCreateWithFlags failed";
#pragma warning(disable : 4127)
if (useConcurrentStreams && (m_fetchStream == NULL))
{
cudaStreamCreateWithFlags(&m_fetchStream, cudaStreamNonBlocking) || "cudaStreamCreateWithFlags failed";
cudaStreamCreateWithFlags(&m_assignStream, cudaStreamNonBlocking) || "cudaStreamCreateWithFlags failed";
}
}
PrefetchGPUDataTransferer::~PrefetchGPUDataTransferer()
{
try
{
PrepareDevice(m_deviceId);
}
catch (...)
{
// the error is already logged
return;
}
auto code = cudaStreamDestroy(m_stream);
if (code != cudaSuccess)
{
std::cerr << "cudaStreamDestroy failed (PrefetchGPUDataTransferer dtor): "
<< cudaGetErrorString(code) << " (cuda error " << code << ")"<< std::endl;
}
}
void GranularGPUDataTransferer::CopyCPUToGPUAsync(const void* cpuBuffer, size_t numElements, size_t elementSize, void* gpuBuffer)
{
PrepareDevice(m_deviceId);
cudaMemcpyAsync(gpuBuffer, cpuBuffer, numElements * elementSize, cudaMemcpyHostToDevice, m_assignStream) || "cudaMemcpyAsync failed";
}
void GranularGPUDataTransferer::WaitForCopyGPUToCPU()
{
PrepareDevice(m_deviceId);
cudaEventSynchronize(m_fetchCompleteEvent) || "cudaEventSynchronize failed";
}
void GPUDataTransferer::WaitForCopyCPUToGPUAsync()
{
PrepareDevice(m_inner->m_deviceId);
SyncEvent(m_inner->m_assignCompleteEvent);
}
void GranularGPUDataTransferer::WaitForSyncPointOnAssignStreamAsync()
{
PrepareDevice(m_deviceId);
cudaStreamWaitEvent(m_assignStream, m_syncEvent, 0 /*flags 'must be 0'*/) || "cudaStreamWaitEvent failed";
}
void GranularGPUDataTransferer::RecordComputeStreamSyncPoint()
{
PrepareDevice(m_deviceId);
cudaEventRecord(m_syncEvent, GetStream()) || "cudeEventRecord failed";
}
void GPUDataTransferer<ElemType>::WaitForCopyGPUToCPUAsync()
{
PrepareDevice(m_deviceId);
SyncEvent(m_fetchCompleteEvent);
}