Unoptimized library implementation causing CUDA API slow.#165

Open

nishitnshah wants to merge 10 commits intoProject-HAMi:mainfrom

nishitnshah:perf/reduce-hijack-overhead

Commits on Mar 24, 2026

perf: cache log level to eliminate per-call getenv()/atoi() overhead
nishitnshah
committed
perf: use cached my_slot pointer in wait_status_self()
nishitnshah
committed
perf: replace time() syscall with vDSO clock and add double-checked locking in pre_launch_kernel()
nishitnshah
committed
perf: optimize rate_limiter() by caching limits and removing redundant shared memory ops
nishitnshah
committed
perf: remove dead cuDeviceGetCount call from oom_check()
nishitnshah
committed
docs: add performance optimization documentation
nishitnshah
committed
fix: move log level globals to separate log_utils.c for link visibility
nishitnshah
committed
fix: restore unconditional set_recent_kernel(2) in rate_limiter()
nishitnshah
committed
Revert "fix: restore unconditional set_recent_kernel(2) in rate_limiter()"
nishitnshah
committed
fix: restore set_recent_kernel(2) in rate_limiter() to match original behavior
nishitnshah
committed