On 16 August 2016 at 20:10, Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> wrote: > On Tue, Aug 16, 2016 at 08:04:35PM +0100, Matthew Auld wrote: >> > + if (dst_needs_clflush & CLFLUSH_BEFORE) >> > + batch_len = roundup(batch_len, boot_cpu_data.x86_clflush_size); >> hmm, this bit doesn't seem obvious to me. What am I missing? > > The code is optimized to work on cachelines (to work on a partial > cacheline requires a flush before the read). We know that the batch is > in whole pages and so we can always round up to the end of the next > cacheline. We also know that we can read more than the declared length > of the batch into the local page as we only validate as much as > required. So it is safe to read from beyond the end of the batch, and to > do so avoids having to insert clflushes before the read. Ah, right ofc. Reviewed-by: Matthew Auld <matthew.auld@xxxxxxxxx> _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
