Bug 39132

There are known stencil issues on IVB.  Is this a regression?

This is a random smattering of issues (maybe a tracker bug).

* glsl-fs-convolution-1:

Fails because I haven't implemented register spilling yet.  This is compounded by the fact that I use g111-g127 as fake MRFs...but I don't think I actually properly reserve them, so the register allocator may dole them out for other uses.  That needs to be fixed first.

* glsl-fs-pointcoord:

Fixed by commit 186e37c75454965e1a58298e878a9585f4866749, pushed this morning.

* texture borders:

Like Ironlake and Sandybridge, these work on the simulator but not on hardware.  We need to talk to the hardware team and find out why.

* stencil issues:

We were treating the separate stencil buffer as Y-tiled memory, when it turns out it's actually W-tiled.  Chad has patches on the mailing list which fix some of these (likely a great deal of them).  There may be a few remaining issues.

* others:

Likely one-off bugs.  Would need to do analysis.

(In reply to comment #2)
> * stencil issues:
> 
> We were treating the separate stencil buffer as Y-tiled memory, when it turns
> out it's actually W-tiled.  Chad has patches on the mailing list which fix some
> of these (likely a great deal of them).  There may be a few remaining issues.

On ILK with separate stencil enabled, I believe all the fbo-stencil tests are fixed by my patches. I will have access to an IVB machine tomorrow, and then I will verify the patches fix the tests on IVB also.

Verified. All fbo-stencil tests above pass on Ivybridge with my patches.

For completeness' sake, my stencil patches fix this subset of the failing tests:

bugs_fdo23670-drawpix_stencil

general_stencil-drawpixels

spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX16-copypixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX16-drawpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX16-readpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX1-copypixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX1-drawpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX1-readpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX4-copypixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX4-drawpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX4-readpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX8-copypixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX8-drawpixels
spec_EXT_framebuffer_object_fbo-stencil-GL_STENCIL_INDEX8-readpixels

spec_EXT_packed_depth_stencil_fbo-stencil-GL_DEPTH24_STENCIL8-copypixels
spec_EXT_packed_depth_stencil_fbo-stencil-GL_DEPTH24_STENCIL8-readpixels

spec_EXT_packed_depth_stencil_readpixels-24_8

Zhao,

The majority of these tests should be fixed on Mesa 7.11 and xf86-video-intel master. Could you verify this and update the list of failing tests?

shaders/glsl-fs-shadow2dproj and shaders/glsl-fs-shadow2dproj-bias were failing due to missing "this->result = reg_undef" lines; we were reusing the same register for both the projected shadow comparitor and part of the projected coordinate.  Patch on mailing list awaiting review.

texturing/texdepth is broken because I didn't update shadow sampling in the old brw_wm_emit backend.  Running piglit now; will send out the patch for review soon.

(In reply to comment #6)
> Zhao,
> The majority of these tests should be fixed on Mesa 7.11 and xf86-video-intel
> master. Could you verify this and update the list of failing tests?

Hi chad, 
I have tested with both mesa master and mesa 7.11 branch, both of them have great improve while still have some failures. The detail is as following: 

On mesa master there are 5 cases still failed: 
general_framebuffer-srgb    
shaders_fp-fog    
shaders_glsl-fs-convolution-1    
spec_EXT_packed_depth_stencil_fbo-clear-formats    
spec_EXT_texture_compression_rgtc_texwrap-2D-GL_COMPRESSED_SIGNED_RED_RGTC1-bordercolor 

On mesa 7.11 branch there are 8 cases still failed: 
general_framebuffer-srgb
glean_vp1-Z-write_test
shaders_fp-fog
shaders_glsl-fs-convolution-1
shaders_glsl-fs-shadow2dproj-bias
spec_EXT_packed_depth_stencil_fbo-clear-formats
spec_EXT_texture_compression_rgtc_texwrap-2D-GL_COMPRESSED_SIGNED_RED_RGTC1-bordercolor
texturing_texdepth

Now that we have Stencil and HiZ patches, what is the status of those tests?

I just did a piglit run last night.  With Mesa master + my "i965: Set SURFACE_STATE vertical alignment bit on Ivybridge." patch on the mailing list, I see the following delta from Sandybridge:

- fbo-array
- fbo-depth-array
- framebuffer-srgb
- fp-fog
- fbo-generatemipmap-array
- texwrap-2D-GL_COMPRESSED_SIGNED_RED_RGTC1-bordercolor
  (but this seems like a fluke, bordercolor hasn't properly worked since Gen4)
- array-texture

I'm guessing all the array bugs are related.  I'll try and get to these soon.

I just ran a piglit comparison between Sandybridge and Ivybridge, and there's only one regression now: glean/fbo.  This is a recent break caused by one of my commits; I've filed #45221 to track it.

Closing this bug in favor of #45221.  We're essentially at parity.

Verified it on Mesa 8.0 branch(65526d54aa) and Mesa master branch(d2c54fb5223e).