softgpu: Optimize (bi-)linear texture filtering #17609
Conversation
Seeing as SampleLinearLevel is near the top in the profiler, optimize actual bilinear filtering using SSE2. Solid win in the synthetic benchmark (https://godbolt.org/z/fqh3xvbGx, also doubles as correctness check), no visible difference in actual PPSSPP. Note: profiler suggests that hot part of SampleLinearLevel is elsewhere.
|
I keep making various optimizations myself that locally look like great wins but seems to have barely a measurable effect overall... but it's hard to measure. Machines clock up and down according to load, etc. This one has to be a win on some dimension, maybe power consumption :P I'm all for merging it, though I'll let @unknownbrackets click merge. |
|
Well, in my case "observable difference" would constitute going from 7 FPS on average to 8 - 12.5% improvement, pretty significant for a single function change. It oscillating between 6 and 9 FPS does not help measuring. Offtopic, but while eyeing softgpu for more optimization opportunities, I have several questions, which I'm not sure where to ask. The discord would seem a logical choice... if the damn thing would actually work for me. Maybe I'll just create "softgpu optimization opportunities" issue, or something. |
|
This would only apply for 32-bit Intel, you're not going to end up in this function on x86_64. So it probably won't actually make any difference for most users. I'd tried to avoid over optimizing this code for SSE given that we're already using a jit for it that is much faster (especially with AVX2.) -[Unknown] |
|
Oh right, forgot about that, hah. Do feel free to create a discussion issue if you want. |
|
Oh, looks like I'm blind. I somehow thought that |
Seeing as SampleLinearLevel is near the top in the profiler, optimize actual bilinear filtering using SSE2. Solid win in the synthetic benchmark (https://godbolt.org/z/fqh3xvbGx, also doubles as correctness check), no visible difference in actual PPSSPP. Note: profiler suggests that hot part of SampleLinearLevel is elsewhere.