Host code shares physical memory with both OpenCL™ devices: the CPU and the Intel® Graphics. So consider using combination of
instead of calls to
. The recommendation applies to the CPU OpenCL device, Intel Graphics OpenCL device, and also to the shared (CPU and Intel Graphics devices) context.
Notice that there are two ways to ensure zero-copy path on memory objects mapping. The preferred way is to request the OpenCL runtime to allocate memory with
, so it is originally mirrored on the host in the efficient way.
Another way is to allocate properly aligned and sized memory yourself and share the pointer with the OpenCL framework by using
flag. This is a viable option, if your application uses a specific memory management algorithm, or if you want to wrap existing native application memory allocations. The
flag enables your application to share its memory allocation directly with the OpenCL runtime implementation, and avoid memory copies of the buffer.
For efficiency reasons such a host-side pointer must be allocated for the conditions:
The amount of memory you allocate and the size of the corresponding OpenCL™ buffer must be multiple of the cache line sizes (64 bytes).
Always use 4k alignment (page alignment) when you allocate the host memory for sharing with OpenCL devices.
Consider the following pseudo-code example:
int cachelineSize = clGetDeviceInfo(device, …CL_DEVICE_GLOBAL_MEM_CACHELINE);//bytes
int arraySizeAligned = cachelineSize*(1+(arraySize-1)/cachelineSize);//aligned
void* inputArray = _aligned_malloc(arraySizeAligned, 4096);
cl_mem inputBuf = clCreateBuffer(…CL_MEM_USE_HOST_PTR, arraySizeAligned, inputArray);
Similarly, page-align host pointers for the API calls that accept the pointers:
void* dstArray = _aligned_malloc(arraySize, 4096);
// example of reading a buffer back from Intel® Graphics device (single-device or shared context), notice that clEnqueueMapBuffer is a better solution
clEnqueueReadBuffer(queue, buffer, FALSE, 0, arraySize, dstArray,0, NULL, NULL);
You can map image objects as well. For a context containing only the Intel Graphics device, the mapping of images is less efficient, since the images are tiled and cannot be mapped directly.
You can find additional details in the