search for: depbar

...(st 0xf wr 0x0 rd 0x1 wt 0x3) (st 0xf wr 0x1 wt 0x2) ipa $r2 a[0x90] $r0 0x0 0x1 tex nodep $r1 $r2 0x0 0x1 t2d 0x8 ipa $r3 a[0x84] $r0 0x0 0x1 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) fmul ftz $r3 $r0 $r1 mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x1) (st 0xf) (st 0x0) mov $r0 $r3 0xf exit #endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/ex...

Hi, On Wed, Jun 28, 2017 at 12:53 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote: > BTW, you can drop those explicit "depbar" ops. I think they're only > needed when you're doing something weird with barriers. Blob doesn't > use them (anymore) > Gotcha. Should I remove them in the same patch or a different one? It seems like the depbar removal is different than what the commit message describes...

...(st 0xf wr 0x0 rd 0x1 wt 0x3) (st 0xf wr 0x1 wt 0x2) ipa $r2 a[0x90] $r0 0x0 0x1 tex nodep $r1 $r2 0x0 0x1 t2d 0x8 ipa $r3 a[0x84] $r0 0x0 0x1 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) fmul ftz $r3 $r0 $r1 mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x1) (st 0xf) (st 0x0) mov $r0 $r3 0xf exit #endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/ex...

...(st 0xf wr 0x0 rd 0x1 wt 0x3) (st 0xf wr 0x1 wt 0x2) ipa $r2 a[0x90] $r0 0x0 0x1 tex nodep $r1 $r2 0x0 0x1 t2d 0x8 ipa $r3 a[0x84] $r0 0x0 0x1 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) fmul ftz $r3 $r0 $r1 mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6) (st 0xf) (st 0x0) mov $r0 $r3 0xf exit #endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/ex...

BTW, you can drop those explicit "depbar" ops. I think they're only needed when you're doing something weird with barriers. Blob doesn't use them (anymore) On Tue, Jun 27, 2017 at 11:16 AM, Aaryaman Vasishta <jem456.vasishta at gmail.com> wrote: > v4: Updated the wait dependancy bars based on tex component masks...

...>> +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) >> ipa $r2 a[0x80] $r0 0x0 0x1 >> tex nodep $r0 $r2 0x0 0x0 t2d 0x8 >> > > Out of curiosity, what didn't you add a read-dep-bar on $r2:$r3 here? Missed it, thanks for pointing it out. > > > depbar le 0x5 0x0 0x0 >> -sched (st 0x0) (st 0x0) (st 0x0) >> +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) >> fmul ftz $r3 $r0 $r1 >> mov $r2 $r3 0xf >> > > You can stall for only one cycle here, but the 6 cycles on fmul is needed. > > mov $r1 $r3 0xf >> -...

...(st 0xf wr 0x0 rd 0x1 wt 0x3) (st 0xf wr 0x1 wt 0x2) ipa $r2 a[0x90] $r0 0x0 0x1 tex nodep $r1 $r2 0x0 0x1 t2d 0x8 ipa $r3 a[0x84] $r0 0x0 0x1 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) fmul ftz $r3 $r0 $r1 mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6) (st 0xf) (st 0x0) mov $r0 $r3 0xf exit #endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/ex...

...t 0x2) > ipa $r2 a[0x90] $r0 0x0 0x1 > tex nodep $r1 $r2 0x0 0x1 t2d 0x8 > ipa $r3 a[0x84] $r0 0x0 0x1 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) > ipa $r2 a[0x80] $r0 0x0 0x1 > tex nodep $r0 $r2 0x0 0x0 t2d 0x8 > depbar le 0x5 0x0 0x0 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) > fmul ftz $r3 $r0 $r1 > mov $r2 $r3 0xf > mov $r1 $r3 0xf > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x1) (st 0xf) (st 0x0) > mov $r0 $r3 0xf > exit > #endi...

...0x1 wt 0x2) > ipa $r2 a[0x90] $r0 0x0 0x1 > tex nodep $r1 $r2 0x0 0x1 t2d 0x8 > ipa $r3 a[0x84] $r0 0x0 0x1 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) > ipa $r2 a[0x80] $r0 0x0 0x1 > tex nodep $r0 $r2 0x0 0x0 t2d 0x8 > depbar le 0x5 0x0 0x0 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) > fmul ftz $r3 $r0 $r1 > mov $r2 $r3 0xf > mov $r1 $r3 0xf > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x1) (st 0xf) (st 0x0) > mov $r0 $r3 0xf > exit > #endif >...

...2d 0x8 > ipa $r3 a[0x84] $r0 0x0 0x1 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) > ipa $r2 a[0x80] $r0 0x0 0x1 > tex nodep $r0 $r2 0x0 0x0 t2d 0x8 Out of curiosity, what didn't you add a read-dep-bar on $r2:$r3 here? > depbar le 0x5 0x0 0x0 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) > fmul ftz $r3 $r0 $r1 > mov $r2 $r3 0xf You can stall for only one cycle here, but the 6 cycles on fmul is needed. > mov $r1 $r3 0xf > -sched (st 0x0) (st 0x0) (st 0x0) > +sch...

...r2:r3, which are written on by the two 'ipa' above it, have already been waited on in this tex, and both of them read $r0 so we can safely assume that since the two 'ipa' instructions are already waited on, $r0 will be ready? > > > >> >> >> depbar le 0x5 0x0 0x0 >> -sched (st 0x0) (st 0x0) (st 0x0) >> +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) >> fmul ftz $r3 $r0 $r1 >> mov $r2 $r3 0xf >> >> >> You can stall for only one cycle here, but the 6 cycles on fmul is...

...--git a/src/shader/exac8nv110.fp b/src/shader/exac8nv110.fp index ce78036..220d7e5 100644 --- a/src/shader/exac8nv110.fp +++ b/src/shader/exac8nv110.fp @@ -36,12 +36,11 @@ ipa $r3 a[0x84] $r0 0x0 0x1 sched (st 0x0) (st 0x0) (st 0x0) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 -depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) fmul ftz $r3 $r0 $r1 +sched (st 0x0) (st 0x0) (st 0x0) mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) mov $r0 $r3 0xf +sched (st 0x0) (st 0x0) (st 0x0) exit #endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/exac8nv11...

...> > Missed it, thanks for pointing it out. You don't have to. 'tex' reads two sources ($r2:$r3) and writes into $r0, but as $r2:$r3 are NOT re-used before $r0 is read, you can assume that $r0 will be ready and don't need any read-dep-bar. > > > > depbar le 0x5 0x0 0x0 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x6 wt 0x3) (st 0x6) (st 0x1) > fmul ftz $r3 $r0 $r1 > mov $r2 $r3 0xf > > > You can stall for only one cycle here, but the 6 cycles on fmul is > needed. > &gt...

...] $r0 0x0 0x1 > tex nodep $r1 $r2 0x0 0x1 t2d 0x8 > ipa $r3 a[0x84] $r0 0x0 0x1 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0xf wr 0x2) (st 0xf wr 0x1 wt 0x6) (st 0xf) > ipa $r2 a[0x80] $r0 0x0 0x1 > tex nodep $r0 $r2 0x0 0x0 t2d 0x8 > depbar le 0x5 0x0 0x0 > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) > fmul ftz $r3 $r0 $r1 > mov $r2 $r3 0xf > mov $r1 $r3 0xf > -sched (st 0x0) (st 0x0) (st 0x0) > +sched (st 0x1) (st 0xf) (st 0x0) > mov $r0...

...(st 0x2 wr 0x1 rd 0x0 wt 0x3) (st 0x1 wr 0x0 wt 0x1) ipa $r2 a[0x90] $r0 0x0 0x1 tex nodep $r1 $r2 0x0 0x1 t2d 0x8 ipa $r3 a[0x84] $r0 0x0 0x1 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0xf wr 0x2) (st 0x1 wr 0x0 wt 0x5) (st 0xe) ipa $r2 a[0x80] $r0 0x0 0x1 tex nodep $r0 $r2 0x0 0x0 t2d 0x8 depbar le 0x5 0x0 0x0 -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x6 wt 0x3) (st 0x1) (st 0x1) fmul ftz $r3 $r0 $r1 mov $r2 $r3 0xf mov $r1 $r3 0xf -sched (st 0x0) (st 0x0) (st 0x0) +sched (st 0x1) (st 0xf wt 0x3f) (st 0x1) mov $r0 $r3 0xf exit +nop 0x0 #endif diff --git a/src/shader/exac8nv110.fp...

...ss $r0 a[0x7c] 0x0 0x0 0x1 +mufu rcp $r0 $r0 +ipa $r3 a[0x94] $r0 0x0 0x1 +sched (st 0x0) (st 0x0) (st 0x0) +ipa $r2 a[0x90] $r0 0x0 0x1 +tex nodep $r1 $r2 0x0 0x1 t2d 0x8 +ipa $r3 a[0x84] $r0 0x0 0x1 +sched (st 0x0) (st 0x0) (st 0x0) +ipa $r2 a[0x80] $r0 0x0 0x1 +tex nodep $r0 $r2 0x0 0x0 t2d 0x8 +depbar le 0x5 0x0 0x0 +sched (st 0x0) (st 0x0) (st 0x0) +fmul ftz $r3 $r0 $r1 +mov $r2 $r3 0xf +mov $r1 $r3 0xf +sched (st 0x0) (st 0x0) (st 0x0) +mov $r0 $r3 0xf +exit +#endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/exac8nv110.fpc new file mode 100644 index 0000000..4aa1368 --- /dev/null +++...

...ss $r0 a[0x7c] 0x0 0x0 0x1 +mufu rcp $r0 $r0 +ipa $r3 a[0x94] $r0 0x0 0x1 +sched (st 0x0) (st 0x0) (st 0x0) +ipa $r2 a[0x90] $r0 0x0 0x1 +tex nodep $r1 $r2 0x0 0x1 t2d 0x8 +ipa $r3 a[0x84] $r0 0x0 0x1 +sched (st 0x0) (st 0x0) (st 0x0) +ipa $r2 a[0x80] $r0 0x0 0x1 +tex nodep $r0 $r2 0x0 0x0 t2d 0x8 +depbar le 0x5 0x0 0x0 +sched (st 0x0) (st 0x0) (st 0x0) +fmul ftz $r3 $r0 $r1 +mov $r2 $r3 0xf +mov $r1 $r3 0xf +sched (st 0x0) (st 0x0) (st 0x0) +mov $r0 $r3 0xf +exit +#endif diff --git a/src/shader/exac8nv110.fpc b/src/shader/exac8nv110.fpc new file mode 100644 index 0000000..4aa1368 --- /dev/null +++...

...> +ipa $r3 a[0x94] $r0 0x0 0x1 > +sched (st 0x0) (st 0x0) (st 0x0) > +ipa $r2 a[0x90] $r0 0x0 0x1 > +tex nodep $r1 $r2 0x0 0x1 t2d 0x8 > +ipa $r3 a[0x84] $r0 0x0 0x1 > +sched (st 0x0) (st 0x0) (st 0x0) > +ipa $r2 a[0x80] $r0 0x0 0x1 > +tex nodep $r0 $r2 0x0 0x0 t2d 0x8 > +depbar le 0x5 0x0 0x0 > +sched (st 0x0) (st 0x0) (st 0x0) > +fmul ftz $r3 $r0 $r1 > +mov $r2 $r3 0xf > +mov $r1 $r3 0xf > +sched (st 0x0) (st 0x0) (st 0x0) > +mov $r0 $r3 0xf > +exit > +#endif > diff --git a/src/shader/exac8nv110.fpc b/src/shader/exac8nv110.fpc > new file mod...

I believe I've addressed all the feedback from the first time around, and also made fixes necessary for GM20x based on testing results. I believe now it should actually work for all GM10x and GM20x. Further, GP10x should be very easy to add, but without someone to actually test I didn't want to claim support for it. Ilia Mirkin (7): exa: add GM10x acceleration support hwdefs: update

v5: Rebased on depbar removal patch; removed a redundant read dep-bar. This patch adds proper delays to maxwell exa shaders. rendercheck tests seem consistent with/without this patch. I haven't extensively tested them though. Trello: https://trello.com/c/6LPB2EIS/174-update-maxwell-shaders-with-proper-delays Sign...