[Beignet] double precision support

[Zhigang Gong]

On Tue, Mar 10, 2015 at 10:59:05AM -0700, Matt Turner wrote:
> On Tue, Mar 10, 2015 at 2:19 AM, Zhigang Gong
> <zhigang.gong at linux.intel.com> wrote:
> > 2. The double support is not fully supported. For example, all the math
> >    functions and even the divide instruction is not supported.
> 
> You're right that the hardware doesn't natively do most of the math
> operations on doubles (it even doesn't do floor/ceil/trunc!), but this
> BSpec page [0] does describe using features new to Broadwell to get
> IEEE-compliant fdiv and sqrt for both single-precision and
> double-precision.
> 
> It uses the new INVM and RSQRTM math operations, the new MADM
> instruction, and the additional accumulator registers.
> 
> It seems that INVM/RSQRTM always write the flag register (the
> math.eo.f0 apparently means "early out", it only seems to be
> documented in passing on that page) in order to skip some instructions
> when not necessary.
> 
> [0] 3D-Media-GPGPU Engine > EU Overview > ISA Introduction >
> Instruction Set Reference > EUISA Instructions > math – Extended Math
> Function [SNB+]

Thanks for pointing this out. We confirmed that to implement double
precision support on BDW+ is feasible. We will consider to implement
it after we get most of the OpenCL 2.0 feature done.

Thanks,
Zhigang Gong.