Skip to content

Performance improvements and automations

Tarush Bansal requested to merge dev into main
  • Adding frame rate throttling feature to all benchmarks (enabled using '--max_fps' command line flag)
  • Switching Tensorflow tag to v2.20.0 with SME support. Turning off ArmNN builds.
  • Adding automated benchmarking scripts that collect perfetto/simpleperf traces
  • Fixing segmentation fault on Android in CollapseNet convolve_and_blur_za32 function
  • Moving heap intermediate buffer allocations to stack due to page faulting overhead on Android (new/delete usage for data buffers inside hot compute functions is unsafe)
Edited by Tarush Bansal

Merge request reports