On Wed, Jan 22, 2014 at 11:38:08AM +0000, Tvrtko Ursulin wrote: > > On 01/22/2014 11:21 AM, Chris Wilson wrote: > >On Wed, Jan 22, 2014 at 10:41:38AM +0000, Tvrtko Ursulin wrote: > >>From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> > >> > >>A set of userptr test cases to support the new feature. > > > >Oh, and another thing... > > > >I guess we will want some benchmarks for mmu_notifier scaling since > >tracking the performance implications seems important. > > What would be a realistic benchmark? My first thoughts were something like: Here we can be purely synthetic. :) Have a broad scope of lots of little tests so that we can detect changes easily and later correlate those with realworld measurements. > for each [unsync, sync] > for each N [1, 10, 100, 1000] > create N userptr bos > benchmark for T seconds: malloc/realloc/free > benchmark for T seconds: mmap/unmap > ??? Yes, that will check that creating our userptr bo does not impact on the rest of system/application performance. I think we also want basic measurement of load, so (mmap, create bo, [blit,] munmap) x N mmap region, create N bo, [blit,] munmap region and a bit of random order vs sequential Double checking that speed of pointer access is not affected by creating a userptr bo is also a good idea. > Add this as tests/gem_userptr_benchmark ? Yes. This may be something that Akash would like to contribute ideas to... -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/intel-gfx
