Open BattleAxeVR opened 7 months ago
Thanks for the comments.
Regarding the difference in speed, this is a known issue, and I should be able to verify in the next week or so if this will be fixed in the next driver.
@mklefrancois Hi, I was wondering if there was any news about driver updates to fix the performance delta.
I'm excited to get to performance parity with DX12. thanks
I just tested the latest RTXGI 2.0 sample and Vulkan now runs as fast or even slightly faster as DX12. Was the SER issue fixed in the latest driver? I'll re-test this repo again to make sure (unless the RTXGI sample doesn't use SER for either API, I doubt it given how fast it runs).
Just re-did the tests, it's still the same as before. I wonder whether the RTXGI 2.0 sample uses SER at all, since VK + DX12 have performance parity. Not sure how that relates to GLSL vs HLSL shaders vs SER perf, but DX12 certainly uses HLSL DXR shaders.
ser_pathtrace test results using latest drivers and Vulkan SDK (as of today), Nvidia RTX 4090, CPU = 13900KS
GLSL:
SER ON: 260 FPS SER OFF: 134 FPS
HLSL:
SER ON: 189 FPS SER OFF: 121 FPS
SLANG:
SER ON: 198 FPS SER OFF: 122 FPS
BTW HLSL doesn't compile right now.
Had to add this code to get it to compile, not sure if it's right.
Hi, thanks for adding support for Vulkan SER for all the main shader languages! I got it all running and it's helping me tremendously to integrate it into my own HLSL-based engine.
I did some comparisons on my RTX 4090 and was wondering why there is still a significant performance delta between GLSL and HLSL or SLANG shader pipelines. I guess they aren't doing the exact same work, which is fine, but I'd like to see ideally a much closer difference between them, or at least understand why GLSL still appears to benefit from SER significantly more (2x) than HLSL/SLANG do. And has a higher base framerate too, which is less significant but still nearly 10%:
On an RTX 4090:
132 -> 263 FPS WITH SER ON in GLSL
121 -> 191 FPS with SER ON in HLSL
122 -> 195 FPS with SER ON in SLANG
I was careful not to move the cameras because the framerate is highly view-dependent. This is in Release build.
191 vs 263 FPS is still a big performance delta