Helion autotuner adds NVIDIA CompileIQ ptxas controls to kernel search
Overview
PyTorch Blog says Helion's autotuner can now evaluate pre-tuned Advanced Control Files (ACFs) from NVIDIA CompileIQ alongside its kernel configurations.
The ACFs set ptxas compiler controls such as register allocation, instruction scheduling and code-generation heuristics, which the post says are left at defaults in kernels shipping today. The post says the joint search improves kernels that were already autotuned, but the excerpt does not give the size of the gains.
ACFs are specific to hardware and toolkit: the post says an ACF tuned for sm_90 is not valid on sm_100. It also says each ACF in autotune_search_acf multiplies the number of autotuning trials, and recommends running the search offline on important kernels and then pinning the winning result.
Written by AI from the articles below · updated Oct 19, 12:34 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- PyTorch BlogOfficialHelion autotuner adds NVIDIA CompileIQ ptxas controls to kernel search
AIPyTorch Blog says Helion's autotuner can now evaluate pre-tuned Advanced Control Files from NVIDIA CompileIQ alongside its kernel configurations. The ACFs set ptxas compiler controls such as register allocation, instruction scheduling and code-generation heuristics, which are left at defaults in kernels shipping today. The post says the joint search improves kernels that were already autotuned, though it does not give the size of the gains in the excerpt.
Heat Trend
Not enough continuous observations to show a trend yet.