Ian Ameline (ian.ameline@(NoSpam).gmail.com) on 4/18/11 wrote:
---------------------------
>Really?
>
>*Fully* compliant? As in <= 1/2 ULP precision guarantees for add, subtract, divide, multiply *and* sqrt?
>
>Denormal support?
>
>+-0 and +- infinity correctly generated and handled?
>
>I'm very skeptical.
>
>-- Ian.
>

Right now, OpenCL 1.1 requires basically the following:

1. round to nearest even, other rounding modes are optional. Dynamic rounding mode selection is not supported.
2. INF and (non signaling) NaN must be supported.
3. Denormalized number is optional.
4. No exception handling is supported right now.
5. +, -, * must be infinitely precise. Division and sqrt are not required to be infinitely precise.

In the case of newer NVIDIA devices (2.x in CUDA):

1. All four rounding modes are supported for most operations, but only static (no dynamic selection).
2. INF and (non signaling) NaN is supported (but absolute and negation of NaN are not conforming)
3. Denormalized number is supported for both single precision and double precision (only double precision in 1.x devices).
4. +, -, * are infinitely precise. Div and sqrt can be made infinitely precise through compiler options (for single precision only, double precision is always infinitely precise).