[Mesa-dev] Difference between TransformFeedback Gallium <-> Vulkan (V3D)

[Marek Olšák]

On Mon, Aug 26, 2019 at 6:51 PM Jason Ekstrand <jason at jlekstrand.net> wrote:

> On Mon, Aug 26, 2019 at 6:43 AM Daniel Stone <daniel at fooishbar.org> wrote:
>
>> Hi Andreas,
>>
>> On Sun, 25 Aug 2019 at 21:11, Andreas Bergmeier <abergmeier at gmx.net>
>> wrote:
>> > For a few weeks now I am working on implementing Vulkan for VideoCore 6
>> AKA 42 (using V3D/DRM). Don't hold you breath ;)
>>
>> Great! I can't say much about the specifics of VideoCore hardware, but
>> at least for some of the common parts ...
>>
>> > Currently I am trying to understand what is necessary or how to
>> interact with V3D. So I am looking at TransformFeedback because it
>> interacts with quite a few other parts of the pipeline and still seems less
>> mangled into the big logic than other features. So I am comparing how
>> Gallium (V3D) is handling TF in the state tracker VS how Vulkan (Intel) is
>> handling the Extension.
>> >
>> > The following is what I so far think I gathered:
>> > 1. GV3D is handling TransformFeedback directly with other bound parts
>> of the pipeline (e.g. `emit_state` in _v3dx_emit.c_). It seems to look into
>> the shader and associated TF specs. It seems to use "streamout targets",
>> although I do not yet understand what these are really. Then it passes all
>> the offsets, indices and finally resources to V3D.
>>
>> 'Stream out' is basically just what DirectX calls its version of
>> transform feedback. The idea is the same: capturing output from
>> vertex/geometry stages.
>>
>> > 2. The Vulkan Extension only knows about CounterBuffers and iterates
>> over these. Intel seems to call TF -> XF? and subsequently the buffers XFB.
>> Have also not yet gathered what is the difference and what all the
>> gazillion acronyms mean.
>>
>> 'XFB' is the most widely-used acronym for transform, as 'trans' or
>> 'cross' can abbreviate to X. 'TF' is not as widely used as XFB.
>>
>> > So far my idea would be to implement TF similar to Intel and instead of
>> iterating over "streamout targets" I would iterate XFBs.
>>
>> So these would really be the same thing. A streamout target is where
>> the VC4 writes its output stream of data from these shading stages,
>> and a counter buffer is where Vulkan writes the output stream of data
>> from these shading stages.
>>
>
> Be warned that the way the Intel streamout hardware works is really
> weird.  It's designed from the perspective of something walking a buffer
> and trying to figure out which outputs to grab.  This is completely
> backwards (or inside-out, whichever is weirder) from the API which is
> written from the perspective of a shader shoving data into a buffer.  If
> you're looking at the ANV code, it's really easy to get turned around.
>

Yeah that is unpleasant. The core of transform feedback is really an atomic
add instruction that returns in the thread ID order and can fail if the
remaining free range is too small. With that, transform feedback can be
nicely emulated in the shader. The fact that it's the only way to do buffer
append in the thread ID order makes it pretty unique, and perhaps it would
be more useful if it was supported by compute shaders.

Marek
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20190827/742d0404/attachment.html>