Possible reasons of higher memory cost when using OpenCL

I have an image processing program which iterates each pixel and does some kind of filtering computing on it. For such kind of loop algorithm, when each pixel's computation is independent, using OpenCL to accelerate is a good idea. However, when I use OpenCL to process each pixel as a thread, I found the memory cost become 30% higher than totally serialized version.

Further I did another experiment, i.e. subdivide the images into 3 x 3 blocks, and applied OpenCL module to each block. Then I found the memory cost decreased, but still higher than the serialized version. With the number of blocks subdivided more, the memory cost decreased and speed also decreased.

Could anyone intuitively think about why OpenCL will cause more memory? Actually I don't have allocating memory process in each thread, so I cannot see any reason for it.

Recent Questions...