Add floats to storage buffers with the GPU's own atomic add instead of a compare-and-swap loop #2336

Closed

opened 2026-11-11 15:10:05 +00:00 by mooki000 · 1 comment

mooki000 commented

2026-11-11 15:10:05 +00:00

(Migrated from github.com)

The idea

Compute programs that run every frame in Kakariko and on Death Mountain Trail add floats to storage buffers, and the shader translator makes each of those adds a compare-and-swap loop. On an RTX 4090 the busiest of them takes 77 to 80 attempts per add on average. With the GPU's own float add (VK_EXT_shader_atomic_float, shaderBufferFloat32AtomicAdd) the frame's GPU time there is 0.4 to 0.6 ms lower, about 7%.

GPUs that can add floats atomically should do it in one instruction; the others keep the loop.

Anything else

Claim: #2306 is open with it, its tests and the measurements. By docs/workflow.md: S3 (1), a small frame-rate gain x reach (GPUs whose driver offers the feature; measured on NVIDIA only) x confidence 1.0 for the RTX 4090.
The idea is yassine-milal's, from #886, which was closed to keep the queue clear. #2306 is written again on current main, with no code copied from it.
Measured on an RTX 4090 only; AMD and Intel are not measured, as the pull request says.
I can test on Linux with an RTX 4090 (NVIDIA 615.78.08) and an Intel UHD 770 (Mesa 26.2.5), not on Windows or AMD.

### The idea Compute programs that run every frame in Kakariko and on Death Mountain Trail add floats to storage buffers, and the shader translator makes each of those adds a compare-and-swap loop. On an RTX 4090 the busiest of them takes 77 to 80 attempts per add on average. With the GPU's own float add (`VK_EXT_shader_atomic_float`, `shaderBufferFloat32AtomicAdd`) the frame's GPU time there is 0.4 to 0.6 ms lower, about 7%. GPUs that can add floats atomically should do it in one instruction; the others keep the loop. ### Anything else - Claim: #2306 is open with it, its tests and the measurements. By `docs/workflow.md`: S3 (1), a small frame-rate gain x reach (GPUs whose driver offers the feature; measured on NVIDIA only) x confidence 1.0 for the RTX 4090. - The idea is yassine-milal's, from #886, which was closed to keep the queue clear. #2306 is written again on current `main`, with no code copied from it. - Measured on an RTX 4090 only; AMD and Intel are not measured, as the pull request says. - I can test on Linux with an RTX 4090 (NVIDIA 615.78.08) and an Intel UHD 770 (Mesa 26.2.5), not on Windows or AMD.

NotRexed commented

2026-11-11 21:31:32 +00:00

(Migrated from github.com)