On Friday 15 May 2015, Fredrik Höglund wrote:
> On Friday 15 May 2015, Michel Dänzer wrote:
> > On 14.05.2015 22:52, fred...@kemper.freedesktop.org (Fredrik HXXglund)
> > wrote:
> > >
> > > URL:
> > http://cgit.freedesktop.org/mesa/mesa/commit/?id=6b284f08ab399154ad10e2166440b44cbbdcb2c5
> > > Author: Laura Ekstrand <la...@jlekstrand.net>
> > > Date: Tue Feb 3 14:47:00 2015 -0800
> > >
> > > main: _mesa_blit_framebuffer updates its arbitrary framebuffers.
> > >
> > > Previously, we used _mesa_update_state to update the currently bound
> > > framebuffers prior to performing a blit. Now that
> > > _mesa_blit_framebuffer
> > > uses arbitrary framebuffers, _mesa_update_state is not specific
> > > enough.
> > >
> > > Reviewed-by: Fredrik Höglund <fred...@kde.org>
> > > Signed-off-by: Fredrik Höglund <fred...@kde.org>
> >
> > This commit broke the piglit test
> > spec@ext_framebuffer_multisample@bitmap with the radeonsi driver:
> >
> > Probe color at (224,0)
> > Left: 0.000000 0.000000 0.000000 1.000000
> > Right: 1.000000 1.000000 1.000000 1.000000
> >
> > Looks like it's because the bottom right squares of the Xs are missing,
> > see the attached picture.
> >
> > Any ideas?
>
> I did notice that failure as well, but when I ran the test manually it
> passed for me, leading me to think that it was a spurious failure.
>
> The output looks exactly the same for me. But the test works by
> comparing the left and right halves of the framebuffer, so if the
> bottom right squares are missing on both sides, the test should
> pass.
>
> The left side is the test image, and the right side is the reference
> image.
Okay, so I've finished analyzing the failure. The glBindFramebuffer()
calls that precede the glBlitFramebuffer() call cause _NEW_BUFFERS to
be set. This used to result in _mesa_update_state() being called from
_mesa_blit_framebuffer(), which in turn would call st_invalidate_state(),
which would mark the framebuffer state as invalid. st_BlitFramebuffer()
would then call st_validate_state() which would call
update_framebuffer_state() which would call st_flush_bitmap_cache().
That call to st_flush_bitmap_cache() is what caused the rendering by the
last glBitmap() call to land in the framebuffer.
The attached patch fixes the problem by restoring the
_mesa_update_state() call. The bug can also be fixed by calling
st_flush_bitmap_cache() from st_BlitFramebuffer().
Fredrik
From 9bb1d8faaa0e943292d333ba315c21b96cb4ca4d Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Fredrik=20H=C3=B6glund?= <fred...@kde.org>
Date: Fri, 15 May 2015 17:35:52 +0200
Subject: [PATCH] mesa: Restore the _mesa_update_state call in
_mesa_blit_framebuffer
We have to set _NEW_BUFFERS and call _mesa_update_state() to ensure
that rendering is flushed to the read framebuffer before the blit.
---
src/mesa/main/blit.c | 15 ++++++++++++---
1 file changed, 12 insertions(+), 3 deletions(-)
diff --git a/src/mesa/main/blit.c b/src/mesa/main/blit.c
index fac9724..bc01ad7 100644
--- a/src/mesa/main/blit.c
+++ b/src/mesa/main/blit.c
@@ -161,13 +161,22 @@ _mesa_blit_framebuffer(struct gl_context *ctx,
GL_DEPTH_BUFFER_BIT |
GL_STENCIL_BUFFER_BIT);
- FLUSH_VERTICES(ctx, 0);
+ /* We have to set _NEW_BUFFERS to make drivers aware that we're about
+ * to render to/from a different set of buffers.
+ */
+ FLUSH_VERTICES(ctx, (readFb != ctx->ReadBuffer ||
+ drawFb != ctx->DrawBuffer) ? _NEW_BUFFERS : 0);
+
+ if (ctx->NewState)
+ _mesa_update_state(ctx);
/* Update completeness status of readFb and drawFb. */
- _mesa_update_framebuffer(ctx, readFb, drawFb);
+ if (readFb != ctx->ReadBuffer || drawFb != ctx->DrawBuffer)
+ _mesa_update_framebuffer(ctx, readFb, drawFb);
/* Make sure drawFb has an initialized bounding box. */
- _mesa_update_draw_buffer_bounds(ctx, drawFb);
+ if (drawFb != ctx->DrawBuffer)
+ _mesa_update_draw_buffer_bounds(ctx, drawFb);
if (!readFb || !drawFb) {
/* This will normally never happen but someday we may want to
--
2.1.4
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
