On Tue, May 17, 2016 at 12:49 PM, Mateusz <[email protected]> wrote:
> > Using numNumaNodes * MAX_POOL_THREADS as the upper-bound on the # > threads per numa node isn't the cleanest. > > My intention was to use the upper-bound for whole threads number (on all > NUMAs). > > This patch is only for case '--pools N' where N is a number, for example > '--pools 55' create exactly 55 threads (with affinity for all logical CPUs > on all NUMAs). Old logic for '--pools 55' was to create min(55, #CPU on > NUMA 0) threads on NUMA 0 and 0 threads on all remaining NUMAs. To use old > logic with this patch you can add comma: '--pools 55,' > In a machine with N nodes, the logical thing to do is to specify N values to the --pools option. The effect above that you describe when only one value is specified is just coincidence and certainly not intentional. Enabling support with '--pools 55,' for example, isn't really essential. > If I use '--pools 6144 -F16' option with this patch on i5 CPU, I get > x265 [info]: Thread pool created using 64 threads > x265 [info]: frame threads / pool features : 16 / wpp(12 rows) > > Without this patch I get only 4 threads. This is the problem that I feared - since the HW has only 4 threads, even though you specified 6144 as the # SW threads you wanted to create, we should've spawned only 4 SW threads also. Using min CPUs across all NUMA nodes will ensure that you only spawn 4 SW threads. You are currently clipped at 64 threads because that is the max # threads in on pool (limited by the fact that we use a 64-bit bit vector for a mask and hence can only track 64 threads per pool). > Also, can you please include the patch as inline in future so that it > becomes easier to review? > > I will do that, thanks for info. My first attempt to do this fail due to > line breaking in my e-mail client. > > > _______________________________________________ > x265-devel mailing list > [email protected] > https://mailman.videolan.org/listinfo/x265-devel >
_______________________________________________ x265-devel mailing list [email protected] https://mailman.videolan.org/listinfo/x265-devel
