[Mesa-dev] [PATCH] st/glsl_to_tgsi: fix mismatch between TGSI BFI/BFE and GLSL

[Ilia Mirkin]

On Mon, Oct 24, 2016 at 12:11 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 24.10.2016 17:16, Ilia Mirkin wrote:
>>
>> On Mon, Oct 24, 2016 at 11:12 AM, Nicolai Hähnle <nhaehnle at gmail.com>
>> wrote:
>>>
>>> On 24.10.2016 16:44, Ilia Mirkin wrote:
>>>>
>>>> In any case, the GLSL way is backwards-compatible with the DX11 way.
>>>> It just specifies some unspecified situations.
>>>
>>>
>>> No, it isn't -- that's the whole problem :)
>>>
>>> Both GLSL and SM5 specify clearly what should happen for the offset=0,
>>> bits=32 case, but they disagree.
>>
>>
>> Oh. Interesting. Do we have tests for that? (Do they fail on nvc0?
>> Pretty sure they were passing last I checked...)
>>
>> I find this very surprising BTW - please double-check.
>
>
> For SM5:
> https://msdn.microsoft.com/en-us/library/windows/desktop/hh446837(v=vs.85).aspx
> is pretty clear that only the 5 LSBs matter, i.e. bits=32 should be the same
> as bits=0.
>
> For GLSL, the quote is:
>
>   Extracts bits [offset, offset + bits - 1] from value,
>   returning them in the least significant bits of the result.
>   [...]
>   If bits is zero, the result will be zero. The result will
>   be undefined if offset or bits is negative, or if the sum
>   of offset and bits is _greater_than_ the number of bits used
>   to store the operand.
>
> None of this would be a problem if the spec said "greater than or equal to"
> instead.
>
> We have piglit tests for that, which I realize now I forgot to mention in
> the commit message:
> arb_gpu_shader5/execution/built-in-functions/{fs,vs}-bitfield{extract,insert}.
>
> There's also a GL CTS test that checks for this.

Oh. Very sad. So yeah - on NVIDIA hw it's easy to represent either
policy. So they must have made the underlying op work the "GLSL" way
for 32, and on DX11 they probably just take the lower 5 bits of the
offset. (The "BFI" op takes a combo bits/offset argument that has to
be composed first, and that composition is usually done with another
BFI op with fixed bits/offset.)

  -ilia