OpenCL 算法加速

Question

I wrote a simple Opencl algorithm that uses my GPU to apply a filter to an image. Everything works fine so I decided to write a C version of the algorithm that basically does the same task (in single core) to compare the different execution speeds. I ran the two algorithms 1000 times each and for the OpenCL version I get an average execution time of 1 ms whereas for the serial version I get an average of 36 ms, that's a huge difference so I was wondering if it's plausible such an improvement.

Answer 1

You already answered your question: Your test showed a speedup of 36x. That is not an uncommon result. When going from a single-core CPU to a GPU implementation you may see no speedup at all (in your case if the image size would be very small, then PCIe latency would be more than the compute speedup you get) all the way to about 2000x (large image / perfectly parallelizable algorithm without communication between threads) depending on your hardware. If you want to figure out how good exactly your implementation is, do a roofline analysis .