Closed ftessier closed 3 years ago
@ftessier I believe you should have merge rights. Once you're happy with the PR feel free to merge :slightly_smiling_face:
Thanks for noting the typos! 🙏 I had noted the change to the timer decorator to account the non-blocking jit calls, but if I understand correctly, the task-clock:u
count captures the run time properly, while the original timer function provides a sense of the (non-blocked) code speed, not counting the (non-blocked) jit calls. I though it might be useful to see both in the perf
run.
I created a markdown log regarding the very early performance testing for basic codes that simply update arrays with random numbers, as discussed earlier. I also include all the source code files discussed in the document (they are also listed in the
.md
for the record), in case anyone wants to play with it. I am just short now of the interesting part, i.e., testing on the GPU; I need to complete a local CUDA installation first, might prove tricky on the work computer which I don't fully manage. @SimonBiggs feel free to merge this PR or leave it open for now, as you see fit.