On Thu, 2017-10-05 at 12:17 -0700, Matt Arsenault wrote:
> > On Oct 5, 2017, at 08:09, Jan Vesely <[email protected]> wrote:
> > 
> > On Mon, 2017-10-02 at 10:32 -0700, Matt Arsenault wrote:
> > > ping
> > > 
> > > > On Sep 19, 2017, at 19:25, Matt Arsenault <[email protected]> wrote:
> > > > 
> > > > These aren't great since they need more test
> > > > values, and generating half results is kind
> > > > of a pain from any other tool. The perfect
> > > > values used don't really stress the conversions,
> > > > but this as at least enough to make sure the
> > > > encoding is correct.
> > 
> > what is this test supposed to test? generic fp16 encoding
> > /functionality? if so, why is it better than vstore_half/vload_half
> > tests? 
> > does it test fp16 version of clamp? why is it better than generated
> > clamp tests?
> > 
> > Jan
> > 
> 
> It’s specifically testing v_mad_mix_f32, v_mad_mixlo_f16,
> v_mad_mixhi_f16 instructions added in gfx9

ah, that makes more sense. Do you mind if I add it to the commit
message? (I'll also fix the formatting nits) with that

Reviewed-by: Jan Vesely <[email protected]>

out of curiosity what's the use of having these in piglit? supposedly
the instruction selection and encoding part is tested in llvm lit. Is
this testing whether the instruction works correctly? shouldn't the hw
design team have tests for that?

Jan

Attachment: signature.asc
Description: This is a digitally signed message part

_______________________________________________
Piglit mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/piglit

Reply via email to