[Mesa-dev] TGSI 16-bit support

[Ilia Mirkin]

On Wed, Aug 23, 2017 at 10:48 AM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 23.08.2017 16:36, Ilia Mirkin wrote:
>>
>> On Wed, Aug 23, 2017 at 10:30 AM, Nicolai Hähnle <nhaehnle at gmail.com>
>> wrote:
>>>
>>> On 23.08.2017 15:15, Nicolai Hähnle wrote:
>>>>
>>>>
>>>> On 22.08.2017 19:32, Marek Olšák wrote:
>>>>>
>>>>>
>>>>> On Tue, Aug 22, 2017 at 7:28 PM, Ilia Mirkin <imirkin at alum.mit.edu>
>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> How do you propose defining the semantics for e.g. loading a 16-bit
>>>>>> value from a constbuf/ssbo? Would those get separate instructions?
>>>>>
>>>>>
>>>>>
>>>>> st/mesa should use UP2H, PK2H and similar opcodes for I16 and U16, and
>>>>> drivers can replace them with MOV if HalfPrecision == 1.
>>>>
>>>>
>>>>
>>>> You mean, if HalfPrecision == 1 for subsequent operations?
>>>>
>>>> How *do* we implement this for LLVM, anyway? Downcast (fptrunc) from
>>>> float
>>>> to half whenever we're loading operands of a HalfPrecision == 1
>>>> instruction,
>>>> and then casting (fpext) back up before storing the result?
>>>>
>>>> LLVM instcombine seems quite capable of seeing through that in simple
>>>> code, but I worry about control flow.
>>>
>>>
>>>
>>> Thinking about this some more, having the precision a property of the
>>> temporaries like Ilia suggested would probably help with emitting LLVM IR
>>> that behaves well across control flow, but complicate st_glsl_to_tgsi.
>>> Hard
>>> to say what the tradeoff is there.
>>
>>
>> Why would it complicate glsl_to_tgsi? At the GLSL level, it's not the
>> operations that have precision, but variables. And those variables map
>> to temp's... we'd have to create a separate pool of high- vs
>> low-precision temps, but that's about it.
>
>
> Well, it may not be so bad in the end. But we definitely have to be more
> careful with things like temporary remapping, peephole, etc.
>
> Also, at least OUT variables could also be affected, right?

And IN as well. Note that some hardware has support for fp16 varyings
and fp16 color outputs (the GLES end of things).