Do the benchmarks measure the time to transfer the images from the host (CPU) to the GPU and back about halideautogpu HOT 2 CLOSED

Hi there,

Halide should only move the inputs once to the GPU then keep them there for follow-up calls to the same pipeline.
The device_sync in the code above forces any functions running on the GPU to finish before taking a measurement, so it takes synchronization costs into account .

In my experience, replacing that with a copy_to_host didn't have much of an impact (as long as device_sync stays there of course).
Removing it altogether may cause the benchmark function to only report the time needed to queue requests to the GPU.

See the definition of benchmark at:

https://github.com/halide/Halide/blob/master/tools/halide_benchmark.h

from halideautogpu.

That makes sense. Thanks!

from halideautogpu.