[Mesa-dev] TGSI 16-bit support
maraeo at gmail.com
Thu Aug 24 16:17:31 UTC 2017
On Thu, Aug 24, 2017 at 3:39 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 24.08.2017 14:19, Erik Faye-Lund wrote:
>> On Wed, Aug 23, 2017 at 6:19 PM, Marek Olšák <maraeo at gmail.com> wrote:
>>> On Wed, Aug 23, 2017 at 3:08 PM, Nicolai Hähnle <nhaehnle at gmail.com>
>>>> Here's another question: What does "low precision" mean on a texture
>>>> instruction? Are the offsets low precision or is it the output? Maybe we
>>>> punt on this for now -- at least GCN doesn't have low precision there
>>> HalfPrecision means that all dst and src sources can be 16-bit.
>>> If the consumer of a TEX instruction is 16-bit, TEX should return
>>> 16-bit automatically. If a source of a TEX instruction is 16-bit, TEX
>>> should accept 16-bit automatically.
>> This sounds inconsistent with how lowp works; a texture-sampler
>> declared as lowp in GLSL only have low precision output AFAIK.
> lowp and mediump are entirely optional concepts. Their precision can be as
> high as highp operations.
FWIW, we won't probably treat lowp differently from mediump. Its
precision requirements are low and weird: float could be an 11-bit
fixed-point type with 8 fractional bits, and int/uint should have at
least 9 bits.
More information about the mesa-dev