Repository navigation
This update brings up to 4x speedup for running FluidX3D on CPUs. Looking at the x86 assembly, I stumbled upon two major performance issues on CPU: vload_half()/vstore_half_rte() OpenCL functions were not vectorized by Intel CPU Runtime for OpenCL - an oversight in the compiler - however plain (half)<->(float) datatype casting gets the load/store vectorized to vcvtph2ps/vcvtps2ph, and that brings massive performance speedup. The second issue was that the guard clause early in the stream_collide() kernel causes thread divergence, which again breaks vectorization. Turning the guard clause into a mask for the remaining kernel resolved that and brings some speedup for all FP32/FP16S/FP16C storage types.
This update also contains an important fix for Nvidia GPUs: Nvidia quietly enshittified their drivers starting at version 610, by making OpenCL single buffer allocations >1/4 of VRAM capacity fail. This meant FluidX3D could only use a fraction of the GPU's VRAM. A hidden, undocumented environment variable NVCL_DISABLE_MAX_MEM_ALLOC_SIZE_CHECK=1 is now applied as a workaround.
Improvements
- enabled vectorized FP16S load/store and branchless guard clause on CPUs, FP16S on CPUs is now up to 4x faster, FP16C up to 1.5x faster, FP32 up to 1.2x faster
- moved index tables from scratchpad to constant memory space and other minor micro-optimizations
- added dumping for AMD ISA, Intel vISA and CPU x86 assembly, along Nvidia PTX
- display max workgroup size in device infos
- updated driver install instructions
Bug fixes
- fixed maximum buffer allocation size limit for Nvidia GPUs with driver 610 or newer
- fixed temperature advection artifacting at
beta=0with volume force - fixed streamline rendering in 3D with
Nz=1 - fixed integer overflow in
graphics_q() - fixed wrong OpenCL barriers
- hardened
read_bmp()andread_stl_raw()against corrupt files
Have fun with the software!
-- Moritz