Notes on OpenCL implementation of HEVC motion compensation filters.

This filter was designed to maximize performance of our HEVC video codecs, using OpenCL acceleration. Typically interpolation takes up to 60% of HEVC decoder complexity and thus it's a first candidate for acceleration. But obtained results still not so impressive.

Performance numbers for 1920x1080p luma-only interpolation:

GPU:

• AMD Radeon HD 7750 - 200 FPS (8 cores) 
• Intel HD Graphics 4000 - 60 FPS (16 cores) 
• NVidia ION GPU - 27 FPS (2 cores) 

CPU:

• Intel CPU driver - 90 FPS (one core, 3570K@3400MHz) 
• AMD CPU driver -20 FPS (one core, 3570K@3400MHz) 
• Non-OpenCL implementation (reference SSE4.2) - 175 FPS, one core(synthetic test). 

As one can see, that actually only AMD GPU and Intel CPU can be used for real-time acceleration with OpenCL. Our "hand-made" SSE4.2 optimizations of the same functions, provide interpolation in only 600 clocks/block 8x8, i.e. ~ 175 FPS using only one core of Intel i5.

Probably I missed something. So, I will continue my attempts to optimize these filters, because it should be significantly faster. For now filter was implemented using vector instructions (16x). And only 3 vector multiplications performed per pixel.

So, theoretical peak is for various devices should be:

• For AMD it should be ~ 200 FPS. 
• For NVidia ION ~50 FPS. 
• For Intel CPU & AMD CPU driver - as we already noticed ~175 FPS, one core. 
• And for Intel HD 4000 it should be ~125 FPS. 

Let's see...


UPD: some improvements was done and now results for AMD GPU is about twice better:
Updated results for 4k motion compensation