[Mesa-dev] [PATCH v3 06/29] glsl: add ARB_gpu_shader_fp64 to the glsl extensions. (v2)
adityaavinash1 at gmail.com
Sun Feb 8 15:45:56 PST 2015
On Sun, Feb 8, 2015 at 6:30 PM, Matt Turner <mattst88 at gmail.com> wrote:
> On Sun, Feb 8, 2015 at 2:51 PM, Dave Airlie <airlied at gmail.com> wrote:
> > On 9 February 2015 at 08:44, Aditya Avinash <adityaavinash1 at gmail.com>
> >> 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
> >> 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.
> i965 hardware needs varying degrees of help to do doubles (the
> round-down instruction doesn't work on doubles, for instance), so
> we'll need some amount of this.
> But I'm curious about what you think the cost:benefit is to writing
> what sounds like a potentially significant amount of code  to
> support doubles on old cards? For a feature that no one seems to use,
> no less.
I think the current sweet spot for AMD GPUs is NI. I can't buy less than
that on Amazon (high end!).
Full support for SI to start with and NI would be enough for now.
Is it possible to borrow the techniques from OpenCL as it should support
doubles for all compatible cards?
I think using softfloat at GLSL compilation stage will be a good idea
(atleast saves time on runtime compiler). Having both soft and hard floats
in the binary can help as you can discard which ever you want depending on
the backend. This also removes the stress on runtime compiler.
> Actually, looking at the link you gave, Wikipedia says they just
> support GL 3.3. Does fglrx expose ARB_gpu_shader_fp64 but not GL 4.0
> on them?
>  Take for instance a WIP implementation of floor:
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the mesa-dev