Mesa (main): pan/bi: Implement b2i with MUX

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: e57dfed419d930849e86514a95c2e71a05f58c56
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=e57dfed419d930849e86514a95c2e71a05f58c56

Author: Alyssa Rosenzweig <alyssa at collabora.com>
Date:   Thu Jun  2 20:13:14 2022 -0400

pan/bi: Implement b2i with MUX

The result_type modifier propagation looks for MUX instructions, so using this
canonical b2i implementation allows the sequence b2i(cmp) to be fused.
It's also faster on its own: on Valhall, MUX may be implemented as CSEL on the
CVT unit, while AND may only be implemented on the SFU unit. So in case this
doesn't get fused, we expect 4x better throughput for b2i with this
implementation. Similarly, on Bifrost, MUX may be scheduled to either unit (as
CSEL on FMA or MUX on ADD), whereas AND may only be scheduled to FMA.

Results on Mali-G52:

total instructions in shared programs: 2419171 -> 2414814 (-0.18%)
instructions in affected programs: 272203 -> 267846 (-1.60%)
helped: 767
HURT: 0
helped stats (abs) min: 1.0 max: 138.0 x̄: 5.68 x̃: 2
helped stats (rel) min: 0.12% max: 15.57% x̄: 2.09% x̃: 0.68%
95% mean confidence interval for instructions value: -6.68 -4.68
95% mean confidence interval for instructions %-change: -2.37% -1.82%
Instructions are helped.

total tuples in shared programs: 1932822 -> 1929234 (-0.19%)
tuples in affected programs: 76485 -> 72897 (-4.69%)
helped: 380
HURT: 3
helped stats (abs) min: 1.0 max: 138.0 x̄: 9.46 x̃: 1
helped stats (rel) min: 0.14% max: 15.96% x̄: 3.81% x̃: 0.92%
HURT stats (abs)   min: 1.0 max: 6.0 x̄: 2.67 x̃: 1
HURT stats (rel)   min: 0.38% max: 8.57% x̄: 3.80% x̃: 2.44%
95% mean confidence interval for tuples value: -11.30 -7.44
95% mean confidence interval for tuples %-change: -4.27% -3.22%
Tuples are helped.

total clauses in shared programs: 356094 -> 355992 (-0.03%)
clauses in affected programs: 3264 -> 3162 (-3.12%)
helped: 80
HURT: 0
helped stats (abs) min: 1.0 max: 9.0 x̄: 1.27 x̃: 1
helped stats (rel) min: 0.81% max: 50.00% x̄: 4.83% x̃: 3.39%
95% mean confidence interval for clauses value: -1.49 -1.06
95% mean confidence interval for clauses %-change: -6.23% -3.43%
Clauses are helped.

total cycles in shared programs: 167337.10 -> 167329.19 (<.01%)
cycles in affected programs: 510.08 -> 502.17 (-1.55%)
helped: 80
HURT: 2
helped stats (abs) min: 0.041665999999999315 max: 0.7916659999999993 x̄: 0.10 x̃: 0
helped stats (rel) min: 0.51% max: 13.64% x̄: 2.12% x̃: 1.34%
HURT stats (abs)   min: 0.041665999999999315 max: 0.0416669999999999 x̄: 0.04 x̃: 0
HURT stats (rel)   min: 0.39% max: 2.78% x̄: 1.58% x̃: 1.58%
95% mean confidence interval for cycles value: -0.12 -0.07
95% mean confidence interval for cycles %-change: -2.59% -1.48%
Cycles are helped.

total arith in shared programs: 73819.54 -> 73669.25 (-0.20%)
arith in affected programs: 2840.54 -> 2690.25 (-5.29%)
helped: 383
HURT: 3
helped stats (abs) min: 0.041665999999999315 max: 5.75 x̄: 0.39 x̃: 0
helped stats (rel) min: 0.33% max: 18.81% x̄: 4.39% x̃: 0.98%
HURT stats (abs)   min: 0.041665999999999315 max: 0.25 x̄: 0.11 x̃: 0
HURT stats (rel)   min: 0.39% max: 8.96% x̄: 4.04% x̃: 2.78%
95% mean confidence interval for arith value: -0.47 -0.31
95% mean confidence interval for arith %-change: -4.93% -3.71%
Arith are helped.

total quadwords in shared programs: 1679798 -> 1676259 (-0.21%)
quadwords in affected programs: 72826 -> 69287 (-4.86%)
helped: 381
HURT: 15
helped stats (abs) min: 1.0 max: 142.0 x̄: 9.35 x̃: 1
helped stats (rel) min: 0.25% max: 18.87% x̄: 4.33% x̃: 1.13%
HURT stats (abs)   min: 1.0 max: 6.0 x̄: 1.47 x̃: 1
HURT stats (rel)   min: 0.30% max: 6.25% x̄: 0.77% x̃: 0.35%
95% mean confidence interval for quadwords value: -10.76 -7.11
95% mean confidence interval for quadwords %-change: -4.71% -3.56%
Quadwords are helped.

Results on Mali-G57:

total instructions in shared programs: 2704193 -> 2699317 (-0.18%)
instructions in affected programs: 293366 -> 288490 (-1.66%)
helped: 758
HURT: 5
helped stats (abs) min: 1.0 max: 151.0 x̄: 6.45 x̃: 2
helped stats (rel) min: 0.11% max: 22.22% x̄: 2.05% x̃: 0.64%
HURT stats (abs)   min: 1.0 max: 7.0 x̄: 2.20 x̃: 1
HURT stats (rel)   min: 0.22% max: 1.69% x̄: 0.87% x̃: 1.08%
95% mean confidence interval for instructions value: -7.42 -5.36
95% mean confidence interval for instructions %-change: -2.27% -1.79%
Instructions are helped.

total cycles in shared programs: 141711.73 -> 141711.84 (<.01%)
cycles in affected programs: 214.36 -> 214.47 (0.05%)
helped: 4
HURT: 42
helped stats (abs) min: 0.015625 max: 0.359375 x̄: 0.20 x̃: 0
helped stats (rel) min: 1.85% max: 12.78% x̄: 9.12% x̃: 10.93%
HURT stats (abs)   min: 0.015625 max: 0.09375 x̄: 0.02 x̃: 0
HURT stats (rel)   min: 0.17% max: 17.65% x̄: 0.84% x̃: 0.34%
95% mean confidence interval for cycles value: -0.02 0.03
95% mean confidence interval for cycles %-change: -1.23% 1.17%
Inconclusive result (value mean confidence interval includes 0).

total cvt in shared programs: 14479.14 -> 14474.19 (-0.03%)
cvt in affected programs: 2877.05 -> 2872.09 (-0.17%)
helped: 508
HURT: 209
helped stats (abs) min: 0.015625 max: 0.453125 x̄: 0.02 x̃: 0
helped stats (rel) min: 0.25% max: 16.67% x̄: 1.23% x̃: 0.37%
HURT stats (abs)   min: 0.015625 max: 0.296875 x̄: 0.03 x̃: 0
HURT stats (rel)   min: 0.15% max: 18.18% x̄: 1.70% x̃: 0.34%
95% mean confidence interval for cvt value: -0.01 -0.00
95% mean confidence interval for cvt %-change: -0.57% -0.18%
Cvt are helped.

total sfu in shared programs: 7875.69 -> 7590.75 (-3.62%)
sfu in affected programs: 1567.38 -> 1282.44 (-18.18%)
helped: 906
HURT: 0
helped stats (abs) min: 0.0625 max: 8.625 x̄: 0.31 x̃: 0
helped stats (rel) min: 2.38% max: 100.00% x̄: 16.80% x̃: 5.63%
95% mean confidence interval for sfu value: -0.37 -0.26
95% mean confidence interval for sfu %-change: -18.43% -15.17%
Sfu are helped.

total quadwords in shared programs: 1468152 -> 1465800 (-0.16%)
quadwords in affected programs: 37104 -> 34752 (-6.34%)
helped: 161
HURT: 2
helped stats (abs) min: 8.0 max: 80.0 x̄: 14.71 x̃: 8
helped stats (rel) min: 1.67% max: 20.00% x̄: 8.05% x̃: 7.69%
HURT stats (abs)   min: 8.0 max: 8.0 x̄: 8.00 x̃: 8
HURT stats (rel)   min: 3.57% max: 3.85% x̄: 3.71% x̃: 3.71%
95% mean confidence interval for quadwords value: -16.29 -12.57
95% mean confidence interval for quadwords %-change: -8.58% -7.22%
Quadwords are helped.

Signed-off-by: Alyssa Rosenzweig <alyssa at collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16857>

---

 src/panfrost/bifrost/bifrost_compile.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c
index 5a376da963d..123671a8eeb 100644
--- a/src/panfrost/bifrost/bifrost_compile.c
+++ b/src/panfrost/bifrost/bifrost_compile.c
@@ -2797,7 +2797,7 @@ bi_emit_alu(bi_builder *b, nir_alu_instr *instr)
         case nir_op_b2i8:
         case nir_op_b2i16:
         case nir_op_b2i32:
-                bi_lshift_and_to(b, sz, dst, s0, bi_imm_uintN(1, sz), bi_imm_u8(0));
+                bi_mux_to(b, sz, dst, bi_imm_u8(0), bi_imm_uintN(1, sz), s0, BI_MUX_INT_ZERO);
                 break;
 
         case nir_op_f2b16: