El 09/06/2025 a las 12:39, Ivan Matek escribió:
On Mon, Jun 9, 2025 at 10:31 AM Joaquin M López Muñoz via Boost <boost@lists.boost.org> wrote:
I'll use -native as you suggest. As for the difference between the original hash production scheme and the one proposed by Kostas (cells marked with *), numbers are not very conclusive, but looks like Kostas's approach incurs a slight degradation in execution time. I hope we can see this more clearly with the upcoming GHA benchmarks on dedicated machines.
btw one more thing that I noticed, but did not investigate in detail: gcc 14 numbers are much much worse than clang 20 on my machine, but just for some values in tables, while others are super close. I would not bother you for tiny performance difference, but this is huge.
At first I thought it is just I am doing something wrong, i.e. number of elements, but other numbers match closely. I do not know nice way to show you ratio of numbers from 2 HTML documents, but so you can search for 13.71 and 13.64 in result_1M_num_release_withoutpragma_gcc_run0.html , you will see the slow values.
This is not one off discrepancy, i.e. on multiple runs I get large difference although it does fluctuate a bit.
I see... well, who knows. FWIW I run the test locally with GCC 13.2 on MSYS2 and these anomalies don't show up. It may be a codegen issue, if you're keen on investigating this further try reducing the number of filter configurations being tested (for instance, remove all rows except #3 and #4). Joaquin M Lopez Munoz