topic Jeffrey, thank you so much in OpenCL* for CPU

problem of clEnqueueMapImage

wu__frank — Tue, 14 Mar 2017 02:29:57 GMT

Recently I use VME to derive MVs information and I need Write Source and Reference image into memory. I use clEnqueueMapImage to do this operation as follows: (host to device side)

size_t pitch;

unsigned char* dst = (unsigned char*)clEnqueueMapImage(queue, srcImage, CL_TRUE,CL_MAP_WRITE, offset, region, &pitch,0, 0, NULL, NULL, &err);

unsigned char* src = YBuffer;

for(i, i<height; i++)

{

memcpy(dst, src, sizeof(unsigned char)*width);

src+=src_pitch;

dst += pitch;

}

clEnqueueUnmapMemObject(queue, srcImage, dst,0, NULL, NULL);

The simulation results show wrong data. all MVs are a fix value (-64, -48). However when I use clEnqueueWriteImage, the result is correct. I don't know if the usage of Map/Unmap operation not correct? Thanks a lot.

Depending on how the image

Jeffrey_M_Intel1 — Wed, 15 Mar 2017 06:26:23 GMT

Depending on how the image was created, you may be missing out on the conversion to tiled memory layout that you would get with clEnqueWriteImage.

The driver may be able to help automate the conversion so you can still use mapping. For more information see:

https://software.intel.com/en-us/articles/getting-the-most-from-opencl-12-how-to-increase-performance-by-minimizing-buffer-copies-on-intel-processor-graphics

Jeffrey, thank you so much

wu__frank — Wed, 15 Mar 2017 07:06:32 GMT

Jeffrey, thank you so much for quickly reply. I have found the question. and fixed this problem.