[Mesa-dev] Summer of Code ideas (maybe just an idea wishlist?)

[Ilia Mirkin]

On Thu, Mar 26, 2015 at 10:17 PM, Jan Vesely <jan.vesely at rutgers.edu> wrote:
> On Thu, 2015-03-26 at 21:11 -0400, Rob Clark wrote:
>> On Thu, Mar 26, 2015 at 8:59 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> > On Thu, Mar 26, 2015 at 8:50 PM, Rob Clark <robdclark at gmail.com> wrote:
>> >> On Thu, Mar 26, 2015 at 3:10 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>> >>> On Thu, Mar 26, 2015 at 3:02 PM, Jason Ekstrand <jason at jlekstrand.net> wrote:
>> >>>> On Thu, Mar 26, 2015 at 12:00 PM, Aditya Avinash
>> >>>> <adityaavinash1 at gmail.com> wrote:
>> >>>>> I mean, implementing fp64 on single precision systems.
>> >>>>
>> >>>> Ok, That makes more sense!  Having lowering passes for various FP64
>> >>>> operations would be great.
>> >>>
>> >>> We should make sure that there are customers of such work. It only
>> >>> makes sense to do this when everything else but FP64 are supported for
>> >>> GL 4.0 (including tess). r600 (eg/ni) is an obvious customer, although
>> >>> that may be easier to do in sb (I believe GlennK has been looking at
>> >>> it, not sure how far he's gotten). Other than that... not sure. Could
>> >>> be that Adreno A420 can do tess && has no native fp64 support, still
>> >>> need to figure that out (hasn't been RE'd yet).
>> >>
>> >> fwiw, fp64 lowering is almost certainly useful for a3xx when we get to
>> >> the point of doing compute..
>> >
>> > Why does compute need fp64? Do the blob drivers expose the fp64 capability?
>> >
>>
>> opencl (beyond embedded profile) requires fp64.. blob driver on a3xx
>> only supports embedded profile opencl (roughly everything *except*
>> fp64)..
>
> fp64 is required only for OpenCL 1.2+, embedded also makes int64, and
> atomic functions optional + precision of some built-in operations is
> relaxed.

I'm reading the OpenCL 2.0 spec... it says

Double precision is an optional feature so the mandated minimum double
precision floating point capability is 0.

and also in the bit about changes from OpenCL 1.2:

Double precision is now an optional core feature instead of an extension

What makes you say that fp64 is required by CL 1.2+?

  -ilia