2017-01-30 12:52 GMT+01:00 Christian Gmeiner <christian.gmeiner@xxxxxxxxx>: > Hi Lucas, > > 2017-01-30 12:48 GMT+01:00 Lucas Stach <l.stach@xxxxxxxxxxxxxx>: >> Am Mittwoch, den 18.01.2017, 12:25 +0100 schrieb Lucas Stach: >>> Hi all, >>> >>> the following patches introduce a cmduf suballocator in the Etnaviv >>> kernel driver, which has the following benefits: >>> >>> 1. Allocating and freeing a CMA buffer for each user command submission >>> is taking a big toll on the CPU, as CMA is not exactly low overhead. >>> By suballocating a single buffer we avoid all this overhead. >>> >>> 2. Less TLB flushes on MMUv2. Each time a new buffer gets mapped into >>> the GPU address space on MMUv2 the TLBs need to be flushed. Mapping >>> the suballocated area once allows to skip the TLB flushes (at least >>> as long as userspace re-uses existing buffers). >>> >>> 3. No workarounds for GC3000 required anymore. The FE TLB flush on >>> GC3000 doesn't work reliably, which required us to map the cmdbufs >>> into the GPU address space at specific positions, which also isn't >>> guaranteed to work if the address space is already crowded. Having >>> a single static area for the cmdbufs side-steps the erratum completely. >>> >>> If I can get reviews and/or enough testing, I would like to include this >>> code in kernel 4.11. >> >> If this helps in testing this change, I have pushed my queue out into my >> public git repo at: >> >> https://git.pengutronix.de/git/lst/linux drm-etnaviv-next >> >> The branch is based on v4.10-rc1, but is trivial to rebase onto >> something more recent. >> > > Will have a look at it later the day. > I really like the idea of a suballocator - nice work! Series is Reviewed-by: Christian Gmeiner <christian.gmeiner@xxxxxxxxx> greets -- Christian Gmeiner, MSc https://www.youtube.com/user/AloryOFFICIAL https://soundcloud.com/christian-gmeiner _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
