[Mesa-dev] [RFC PATCH] Add support for ARB_vertex_type_2_10_10_10_rev on i965

[Kenneth Graunke]

On 10/13/2012 07:13 PM, Kenneth Graunke wrote:
> On 10/09/2012 01:06 PM, Kenneth Graunke wrote:
>> On 10/09/2012 02:33 AM, Chris Forbes wrote:
>>> This series almost adds support for ARB_vertex_type_2_10_10_10_rev on
>>> i965.
>>> Working from the Sandybridge and Ivybridge PRMs, this ought to work
>>> (what look
>>> like the correct dwords end up in the batch, etc).
>>>
>>> Details pieced together from both PRMs, since the Sandybridge PRM
>>> references
>>> surface format encodings which aren't in its master table, and the
>>> Ivybridge
>>> PRM is missing the data on which formats are usable in vertex buffers.
>>>
>>> Any hints on what I'm missing -- and trying it on an Ivybridge
>>> machine, where
>>> the formats *definitely* ought to all exist -- would be appreciated.
>>
>> Hi Chris,
>>
>> I looked over your patches this morning and at first glance, they look
>> good!   Unfortunately, I ran the piglit tests and got failures.  Not
>> entirely sure what's going on there.  Endianness issues perhaps?
>>
>> I'll try and take a more detailed look later this week.
>>
>> Thanks so much for doing this!
>>
>> --Ken
>
> Unfortunately, it's looking like the 10_10_10_2 USCALED/SSCALED formats
> don't work in hardware, at least on Ivybridge...the vertices get
> converted to floats incorrectly.  Apparently the workaround is to pass
> them in as UINT and emit shader code to fix it up (similar to how we do
> for GL_FIXED).  I'll try and investigate more and send patches...

Scratch one mystery:

If you look in the Sandybridge PRMs, volume 4 part 1, page 90 and 
following, you'll see that not all of the 10-10-10-2 formats have a "Y" 
in the "Input Vertex Buffer" column.

The ones that are supported are:
* R10G10B10A2_UNORM
* R10G10B10A2_UINT
* R10G10B10_SNORM_A2_UNORM
* R10G10B10X2_USCALED
(and AFAICT these should be supported all the way back to Gen4)

Unfortunately, the Ivybridge docs appear to have lost the information 
about what's unsupported.  Looking at the internal docs[*], it appears 
the following formats are *not* supported for vertex buffer data on 
current hardware:

* R10G10B10A2_SNORM
* R10G10B10A2_USCALED
* R10G10B10A2_SSCALED
* R10G10B10A2_SINT
* B10G10R10A2_UNORM
* B10G10R10A2_SNORM
* B10G10R10A2_USCALED
* B10G10R10A2_SSCALED
* B10G10R10A2_UINT
* B10G10R10A2_SINT

This explains why I saw a lot of test failures with your patches: 
they're trying to use formats which aren't supported.

So it does seem to be necessary to use either R10G10B10A2_UINT or 
R10G10B10_SNORM_A2_UNORM and emit shader code to convert the data.

I apologize for the documentation...this really wasn't clear at all.

--Ken

[*] For Intel folks reading along, this is based on the table in 
vol2a.02 / Vertex Fetch (VF) Stage / 3D Primitive Processing / 
FormatConversion.