[Mesa-dev] [PATCH 10/12] st/nine: Remove usage of SQRT in ff code

[Axel Davy]

On 08/02/2016 00:21, Ilia Mirkin wrote:
> On Sun, Feb 7, 2016 at 6:13 PM, Axel Davy <axel.davy at ens.fr> wrote:
>> SQRT is not supported everywhere, so replace
>> it by RSQ + RCP
>>
>> Signed-off-by: Axel Davy <axel.davy at ens.fr>
>> ---
>>   src/gallium/state_trackers/nine/nine_ff.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/gallium/state_trackers/nine/nine_ff.c b/src/gallium/state_trackers/nine/nine_ff.c
>> index a5466a7..894fc63 100644
>> --- a/src/gallium/state_trackers/nine/nine_ff.c
>> +++ b/src/gallium/state_trackers/nine/nine_ff.c
>> @@ -563,7 +563,8 @@ nine_ff_build_vs(struct NineDevice9 *device, struct vs_build_ctx *vs)
>>           struct ureg_src cPsz2 = ureg_DECL_constant(ureg, 27);
>>
>>           ureg_DP3(ureg, tmp_x, ureg_src(r[1]), ureg_src(r[1]));
>> -        ureg_SQRT(ureg, tmp_y, _X(tmp));
>> +        ureg_RSQ(ureg, tmp_y, _X(tmp));
>> +        ureg_RCP(ureg, tmp_y, _Y(tmp));
> I'd recommend doing
>
> ureg_MUL(ureg, tmp_y, _Y(tmp), _X(tmp))
>
> instead. That should be (a) more numerically stable (rcp doesn't have
> great precision), and (b) not blow up for 0.
Ok for the precision, but I'm not sure for 0

With the mul version, with 0, it ends up computing inf * 0 = NaN,
whereas with the rcp version, it does 1/inf == 0 (as far as I know),
which is the expected result.

>
>>           ureg_MAD(ureg, tmp_x, _Y(tmp), _YYYY(cPsz2), _XXXX(cPsz2));
>>           ureg_MAD(ureg, tmp_x, _Y(tmp), _X(tmp), _WWWW(cPsz1));
>>           ureg_RCP(ureg, tmp_x, ureg_src(tmp));
>> --
>> 2.7.0
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev