nickserv, 1 month ago

@ProjectPhysX Do you have any thoughts on the Snapdragon 855 iGPU in this Chips and Cheese's FluidX3D test? Can you elaborate on the total failure by comparison to Intel?

https://chipsandcheese.com/2024/05/01/inside-the-snapdragon-855s-igpu/

ProjectPhysX, 1 month ago

@nickserv that's a bug in #ARM's #OpenCL runtime: fused-multiply-add (fma) is somehow emulated with terrible performance. This is very similar to what @niconiconi found on Nvidia CMP 170HX, where fma was disabled in the driver.
I've just fixed this in #FluidX3D, by macro-replacing fma with a*b+c. Performance went up by 8-13x on my Samsung S9+ (ARM Mali-G72 MP18) with this workaround.
https://github.com/ProjectPhysX/FluidX3D/commit/9ce2caecfc85e4fda50fed3350304b75b223b06b
cc @chipsandcheese

niconiconi, 1 month ago

@ProjectPhysX @nickserv @chipsandcheese Could it be possible that the GPU only supports mad() but not fma()? That may explain the poor-performance emulation by the driver. It's perhaps worth trying to provide three options: unfused, mad, and fma.

ProjectPhysX, 1 month ago

@niconiconi @chipsandcheese @nickserv mad is equally slow as fma 🐌

niconiconi, 1 month ago

@ProjectPhysX @chipsandcheese @nickserv ​:blobcatfacepalm:​