[Mesa-dev] [PATCH 5/8] i965: Implement FS backend for ARB_sample_shading

[Kenneth Graunke]

On 10/15/2013 05:42 PM, Anuj Phogat wrote:
[snip]
>>> +   /* WM will be run in MSDISPMODE_PERSAMPLE. So, only SIMD8 mode will be
>>
>> Looking at the docs for 3DSTATE_PS, I believe that SIMD16 is allowed in
>> some cases...
>>
>> SNB: Must be SIMD8 only.
>> IVB:
>>  - 4x MSAA: Either SIMD8 or SIMD16 should work.
>>  - 8x MSAA: Must be SIMD8 only.
>>
>> Maybe I'm crazy, though...I haven't looked into it too closely...
>>
> Yes, you're right. I should have explained it here that I've chosen
> to use 'SIMD8 only' for 4x and 8x MSAA on both SNB and IVB.
> I decided to use SIMD8 over SIMD16 in case of 4x MSAA after
> reading following lines in IVB PRM vol 2, part 1, page 330:
> "Object Size: If the number of very small objects (e.g., covering 2
>  subspans or fewer) is  expected to comprise a significant portion
>  of the workload, supporting the 8-pixel dispatch  mode may be
>  advantageous Otherwise there could be a large number of 16-pixel
>  dispatches  with only 1 or 2 valid subspans, resulting in low
>  efficiency for those threads"

I don't see any text on that page to indicate that SIMD8-only is
advantageous for per-sample shading.  You can actually program up to
three pixel shader kernels in 3DSTATE_PS the GPU will dynamically select
the best one based on the object size.  Valid combinations are listed in
the table on page 330.

This means that if you have a small polygon which spans only a few
pixels, it'll use the 8-wide program.  But if you have a large polygon,
covering a large area of pixels, it'll use the 16-wide (or 32-wide)
program.  Processing more pixels at a time amortizes the cost of
instruction dispatch over twice as many pixels.

This is why we always program both an 8-wide and 16-wide version, when
possible.  (We'd also like to experiment with 32-wide eventually.)

> Do you have other reasons to suggest using SIMD16 for 4x MSAA
> on IVB?

At least with normal per-pixel shading, SIMD16 has proven to be a 10-20%
performance improvement.  So I think it's worth using where possible.

--Ken