55 minutes ago, struts said:

Recently stumbled over these threads about GPU acceleration but I have not found any explanation of how the benefit is actually derived.  To my simple understanding (EE degree from 1987!) DSP is an essentially sequential computation that doesn't really lend itself to the parallel processing strengths of a GPU.  But then CUDA Programming hadn't been invented then and I would imagine the state of the art in DSP algorithms has probably advanced a bit in that time as well.  Please could you provide a high level explanation of how the GPU acceleration mechanism works (or link if you've already provided one, sorry I didn't find it), I'm really curious!

Many, but not all, DSP algorithms can be parallelized. It depends on the mathematical constructs whether it is possible or not.

Also on regular CPU's, HQPlayer heavily depends on vector units such as AVX2 that allow multiple data items to be processed with single instruction. AVX2 can compute four 64-bit floating point values with single instruction, per CPU core. Thus 8 core CPU can calculate 32 values simultaneously, for example at 5 GHz CPU speed.

GPUs just take this to the next level by being able to do hundreds or thousands of calculations in parallel. But speed of single calculation is not quite as fast as on CPU. GPU clock speeds are about half compared to CPU clock speeds at around 2.5 GHz.

HQPlayer then splits the work such way that both CPU and GPU can perform calculations at same time, each doing type of work that fits the best.