[Mesa-dev] [PATCH 8/8] gallium: remove TGSI opcode KILL
imirkin at alum.mit.edu
Sun Aug 20 16:24:51 UTC 2017
On Sun, Aug 20, 2017 at 6:51 AM, Marek Olšák <maraeo at gmail.com> wrote:
> On Sun, Aug 20, 2017 at 4:29 AM, Roland Scheidegger <sroland at vmware.com> wrote:
>> For 1-3/8, I don't think anything ever used these, and 5/8 (DPH) we
>> don't need neither, so
>> Reviewed-by: Roland Scheidegger <sroland at vmware.com>
>> 4,6,7, well we can live fine without them, just translating them away
>> the same as you do here too. (I thought some hw could do them natively,
>> hence the driver would have to recognize them to get optimal results,
>> but if they aren't emitted anyway who cares.)
>> Can we keep KILL though please? From a shader api cleanness point of
>> view, I'd rather have this simple instruction than upgrading it to a
>> rather complex instruction involving per-channel float comparisons (plus
>> combining the results), even if it's trivial to do so and the compiler
>> should be able to throw out all the conditional mess.
>> From a conceptual point of view, we don't have conditional ret, cont,
>> etc. neither (much less ones which use built-in per channel float
>> comparisons...). Probably nuking KILL_IF instead isn't a good idea (as
>> this really translates to old-style kill) so IMHO we should just keep both.
>> (But you can really kill breakc if you want - that's going to be less
>> work than any of the dp2a/xpd/scs ones here :-)).
> Yeah, I can keep KILL and then kill BREAKC.
> XPD etc. were really trivial. There is only one or two places that
> need to unwind the instructions. The rest is just removed code.
I can't say I care *strongly*, but some of those DP* variants have hw
instructions that correspond both in nv30 and r300, and I believe
neither have the compilers to properly recombine them.
More information about the mesa-dev