Halide
Halide copied to clipboard
halide
[Hexagon] Profiling changes for abadams/per_instance_profiling
I still see a crash on device when offloading, the culprit being IR like this:
...
let profiler_shared_sampling_token = (void *)alloca(4)
halide_profiler_init_sampling_token(profiler_shared_sampling_token, 0)
...
<Hexagon> for (output) {
halide_profiler_acquire_sampling_token(profiler_shared_sampling_token, profiler_local_sampling_token) // profiler_shared_sampling_token is not a valid read address here
}
Oh yeah, that sampling token didn't account for crossing a host-device boundaries inside a parallel loop. Hrm.
