[Mesa-dev] [PATCH v3 06/29] glsl: add ARB_gpu_shader_fp64 to the glsl extensions. (v2)

[Ilia Mirkin]

On Sun, Feb 8, 2015 at 5:51 PM, Dave Airlie <airlied at gmail.com> wrote:
> On 9 February 2015 at 08:44, Aditya Avinash <adityaavinash1 at gmail.com> wrote:
>> Ya. I just want to know that part "only some r600".
>> I believe some of the nv0 cards doesn't support double. You have any ideas
>> or suggestions to make it possible?
>
> For AMD
> http://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units
>
> has what cards directly support double precision for r600/700/evergreen
>
> The plan is to implement a soffloat library like the fglrx drivers do for the
> GPUs that don't do doubles just to satisfy the GL4.0 requirement.
>
> Whether we implement this in GLSL, NIR or in the backends is currently
> an open question.
>
> Not sure what the nvidia limitations are if any.

All of the Fermi+ cards have pretty full hw support. The only lacking
area is that they only compute the upper 32 bits of a RSQ and RCP
operation, but that's solvable with a bit of Newton-Raphson magic.

The G200 chip (NVA0) is the only Tesla-family chip which has some fp64
support, but it's missing some useful operations, like any sort of
sqrt or division-related functionality. But it can add/multiply. Some
sort of partial softfloat may be useful there, but... it can be
difficult to care about the single high-end chip of the old family.

  -ilia