As much as memset clears amuse me, I have to nak this patch as is. The
memset you are using doesn't properly account for the Y-tiled format of the
MCs buffer for certain surface sizes. I can explain more later but the
short version is that the total_height parameter we have in
intel_mipmap_tree may not be the full height of the tiled buffer. The
total_height is only 4-aligned while the buffer is going to be 32-aligned
(I think). The result is a partially-cleared surface. Incidentally, I
think the memset we use when we create the MCs is wrong too.
--Jason
On Sep 25, 2014 11:28 AM, "Kenneth Graunke" <kenn...@whitecape.org> wrote:
> Fast color clears simply fill the MCS buffer with 0xff. If the MCS BO
> is not busy, we can map it without stalling and simply use memset to do
> that, without having to reprogram the entire GPU, draw rectangles, and
> invoke pixel shaders. This is presumably cheaper.
>
> This code was swiped from intel_miptree_alloc_mcs, which clears the
> buffer using memset when allocating the MCS.
>
> Improves performance in GpuTest Triangle by 5.09439% +/- 1.27536%.
> Xonotic, EgyptHD, and many others also hit this path, but I was not
> able to measure any performance increase in other applications.
>
> Signed-off-by: Kenneth Graunke <kenn...@whitecape.org>
>
> Concerns I have which hopefully reviewers can address:
> - Do we need to worry about clipping/scissoring/etc? I think our
> fast color clear implementation only handles full surface clears, so we
> should be OK.
> - Do we need to check if the miptree BO is referenced/busy? Or is
> checking the MCS really sufficient?
> ---
> src/mesa/drivers/dri/i965/brw_meta_fast_clear.c | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> b/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> index b4e75a7..43df2d0 100644
> --- a/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> +++ b/src/mesa/drivers/dri/i965/brw_meta_fast_clear.c
> @@ -503,6 +503,19 @@ brw_meta_fast_clear(struct brw_context *brw, struct
> gl_framebuffer *fb,
> if (irb->mt->fast_clear_state == INTEL_FAST_CLEAR_STATE_CLEAR)
> continue;
>
> + /* If the MCS isn't busy, just do a CPU-side memset rather than
> + * reprogramming the whole GPU.
> + */
> + if (!drm_intel_bo_references(brw->batch.bo,
> irb->mt->mcs_mt->bo) &&
> + !drm_intel_bo_busy(irb->mt->mcs_mt->bo)) {
> + void *data = intel_miptree_map_raw(brw, irb->mt->mcs_mt);
> + memset(data, 0xff,
> + irb->mt->mcs_mt->total_height *
> irb->mt->mcs_mt->pitch);
> + intel_miptree_unmap_raw(brw, irb->mt->mcs_mt);
> + irb->mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_CLEAR;
> + continue;
> + }
> +
> /* Set fast_clear_state to RESOLVED so we don't try resolve them
> when
> * we draw, in case the mt is also bound as a texture.
> */
> --
> 2.1.0
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
