[Mesa-dev] [PATCH 00/17] Rework Mesa/ST shader compilation logic (Part 1)

[Marek Olšák]

On Wed, Oct 7, 2015 at 9:54 PM, Brian Paul <brianp at vmware.com> wrote:
> On 10/07/2015 12:34 PM, Brian Paul wrote:
>>
>> On 10/05/2015 07:26 PM, Marek Olšák wrote:
>>>
>>> Hi,
>>>
>>> This is a start of reworking how st/mesa translates and creates
>>> shaders. The result of this patch series is this:
>>>
>>> In LinkShader or ProgramStringNotify, the shader is translated to TGSI
>>> as-is. There are no shader code modifications for shader variants. The
>>> glsl-to-tgsi visitor is released.
>>>
>>> We can't release the GLSL IR at this point yet, because core Mesa
>>> needs it for program resource queries at least.
>>>
>>> That's the only place st/mesa interacts with Mesa IR and GLSL IR.
>>> After that, it only works with the TGSI representation.
>>>
>>> When we get to a draw call, tgsi_transform_shader is used to modify
>>> shaders to produce shader variants. It does:
>>> - feature emulation (color clamping, per-sample shading, edgeflags)
>>> - adding glDrawPixels shader code
>>> - adding glBitmap shader code
>>>
>>> All the shader transformations are rewritten completely. There is no
>>> patching with Mesa IR or glsl-to-tgsi anymore.
>>>
>>> It shouldn't be so difficult to add support for some new IR or phase
>>> out Mesa IR and GLSL IR support at some point.
>>>
>>> The main motivations for this series are:
>>> - translate to TGSI as soon as possible
>>> - possibility to create gallium shaders after that if drivers don't
>>> need many variants
>>> - if variants are needed, the path to get them is much simpler
>>>
>>> There will be part 2 when it's ready.
>>>
>>> Please review. Thanks.
>>
>>
>> I did a quick review and it looks good.  Some nice clean-ups for sure.
>>
>> Unfortunately, it looks like the conform drawpix.c test is now failing
>> with this series.  It uses glPixelMap so maybe that's it.  I'll see if I
>> can learn more.
>
>
> Actually, it seems to be glDrawPixels(GL_STENCIL_INDEX) that's failing.
>
> Piglit's ext_packed_depth_stencil-readdrawpixels is failing also.  Can you
> check that out?

That test passes with radeonsi, softpipe, but doesn't with llvmpipe.
I'll take a look.

Marek