void DeviceMatrix3D_copyFromDevice(const DeviceMatrix3D& self, float* dst)
{
if ((self.dim_x == 0) || (self.dim_y == 0) || (self.dim_t == 0)) {
// Bail early if there is nothing to copy
return;
}
if (self.pitch_t == self.dim_y * self.pitch_y) {
// Shortcut if we're packed in the t direction
const size_t widthInBytes = self.dim_x * sizeof(float);
CUDA_SAFE_CALL_NO_SYNC
(cudaMemcpy2D(dst, widthInBytes,
self.data, self.pitch_y * sizeof(float),
widthInBytes, self.dim_y * self.dim_t,
cudaMemcpyDeviceToHost));
return;
}
// Do a series of copies to fill in the 3D array
for (size_t t=0; t < self.dim_t; t++) {
const size_t widthInBytes = self.dim_x * sizeof(float);
float* host_start = dst + t * self.dim_y * self.dim_x;
float* device_start = self.data + t * self.pitch_t;
CUDA_SAFE_CALL_NO_SYNC
(cudaMemcpy2D(host_start, widthInBytes,
device_start, self.pitch_y * sizeof(float),
widthInBytes, self.dim_y,
cudaMemcpyDeviceToHost));
}
}
void SimBufferCuda::Alloc(size_t size)
{
mSize = size;
if (mAllocedSize > 0)
{
if (mAllocedSize == mSize)
return;
Free();
}
cudaError_t result;
switch (mBufferLocation)
{
case Host:
result = mSimCudaAllocator->AllocateHost(&mPtr, mSize);
break;
case HostPinned:
result = mSimCudaAllocator->AllocateHostPinned(&mPtr, mSize);
break;
case Device:
result = mSimCudaAllocator->Allocate(&mPtr, mSize);
break;
}
CUDA_SAFE_CALL_NO_SYNC(result)
mAllocedSize = mSize;
}
void DeviceMatrix_copyFromDevice(const DeviceMatrix& self, float* dst)
{
if ((self.width > 0) && (self.height > 0)) {
const size_t widthInBytes = self.width * sizeof(float);
CUDA_SAFE_CALL_NO_SYNC
(cudaMemcpy2D(dst, widthInBytes,
self.data, self.pitch * sizeof(float),
widthInBytes, self.height,
cudaMemcpyDeviceToHost));
}
}
void DeviceMatrix_copyToDevice(DeviceMatrix& self, const float* data)
{
if ((self.width > 0) && (self.height > 0)) {
const size_t widthInBytes = self.width * sizeof(float);
CUDA_SAFE_CALL_NO_SYNC
(cudaMemcpy2D(self.data, self.pitch * sizeof(float),
data, widthInBytes,
widthInBytes, self.height,
cudaMemcpyHostToDevice));
}
}
void DeviceMatrix3D_copyToDevice(DeviceMatrix3D& self, const float* data)
{
if ((self.dim_x > 0) && (self.dim_y > 0) && (self.dim_t > 0)) {
const size_t widthInBytes = self.dim_x * sizeof(float);
CUDA_SAFE_CALL_NO_SYNC
(cudaMemcpy2D(self.data, self.pitch_y * sizeof(float),
data, widthInBytes,
widthInBytes, self.dim_y * self.dim_t,
cudaMemcpyHostToDevice));
}
}
void SimBufferCuda::Free()
{
if (mAllocedSize > 0)
{
cudaError_t result;
switch (mBufferLocation)
{
case Host:
result = mSimCudaAllocator->FreeHost(&mPtr);
break;
case HostPinned:
result = mSimCudaAllocator->FreeHostPinned(&mPtr);
break;
case Device:
result = mSimCudaAllocator->Free(&mPtr);
break;
}
CUDA_SAFE_CALL_NO_SYNC(result)
mAllocedSize = 0;
}
}