On Tue, 23 Jul 2019 at 17:17, Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote: > > As we never officially write to the scratch buffer, the kernel will > leave it in the CPU read domain upon execution. Our attempt to > invalidate the CPU cache on !llc is therefore skipped as the kernel > doesn't believe the backing store has been invalidated. Use a WC mmap to > avoid the CPU cache for readback, and add an extra sanity check that the > scratch buffer is found at the same location after execution. (This > sanity check does not affect the failure rate on bsw, that is only fixed > after realising that we do not clflush for the invalidate prior to the > read). > > Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111187 > Signed-off-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> Reviewed-by: Matthew Auld <matthew.auld@xxxxxxxxx> _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
