[Mesa-dev] llvmpipe and transform feedback of position

[Dave Airlie]

On Thu, Dec 13, 2012 at 2:48 PM, Dave Airlie <airlied at gmail.com> wrote:
>> Hi,
>>
>> So I've been playing around a bit more with transform feedback with
>> llvmpipe, and it seems the position output from the vertex shader is
>> different from what softpipe produces.
>>
>> using from piglit ./bin/ext_transform_feedback-builtin-varyings gl_Position
>>
>> the position with softpipe for the first call is
>> 0       0.00390625      0.0078125       0.01171875
>> for llvmpipe it is
>> 16      10.666666       0.833333373        85.3333359
>>
>> so is there some intermediate storage or transform that isn't
>> happening in the llvm vertex shader, that should?
>>
>> I start to realise how little I know about the vertex shading side of
>> things, when I hit stuff like this!
>
> Ah okay, its the viewport transform that is happening before feedback,
>
> we probably need to figure out either how to do the feedback with llvm code,
> or split things so clipping/viewport can happen separately. I expect
> the first option to be better, even if it decreases chances of me
> working it out!

Okay now I realise I've been in this code before, and I've stashed the
pre-clip/viewport position into the vertex header
so I'll figure out how to extract that and stick that into the TF
buffer instead.

We probably want to think about codegen TFP at some point though.

Dave.