Draft becasue this exposes an existing bug on Valhall.
On Bifrost, we have a single "load float varying" instruction that controls the
bit size of the result, allowing us to fold a f2f16 into the load. However, the
larger benefit is that 16-bit varying loads are interpolated at 16-bit. Arm
claims that the varying unit has a 32-bit data path, allowing 16-bit varyings to
be interpolated in half the cycles from 32-bit. This change should therefore
improve performance for workloads that are varying units. This means we want to
be aggressive about 16-bit varying loads, even if it costs some extra f2f32
instructions.
glmark2 total score on Mali-G52 up from 1173fps to 1218fps with particular wins
in -brefract, -bshadow, -bjellyfish, and -bshading.
total instructions in shared programs: 2432246 -> 2423668 (-0.35%)
instructions in affected programs: 516056 -> 507478 (-1.66%)
helped: 3641
HURT: 432
helped stats (abs) min: 1.0 max: 12.0 x̄: 2.91 x̃: 2
helped stats (rel) min: 0.08% max: 54.55% x̄: 9.88% x̃: 5.71%
HURT stats (abs) min: 1.0 max: 42.0 x̄: 4.71 x̃: 4
HURT stats (rel) min: 0.23% max: 200.00% x̄: 12.58% x̃: 6.37%
95% mean confidence interval for instructions value: -2.21 -2.00
95% mean confidence interval for instructions %-change: -7.92% -7.07%
Instructions are helped.
total tuples in shared programs: 1941309 -> 1934647 (-0.34%)
tuples in affected programs: 353169 -> 346507 (-1.89%)
helped: 3233
HURT: 453
helped stats (abs) min: 1.0 max: 14.0 x̄: 2.46 x̃: 2
helped stats (rel) min: 0.12% max: 50.00% x̄: 9.90% x̃: 5.56%
HURT stats (abs) min: 1.0 max: 25.0 x̄: 2.85 x̃: 2
HURT stats (rel) min: 0.22% max: 150.00% x̄: 8.96% x̃: 5.26%
95% mean confidence interval for tuples value: -1.89 -1.72
95% mean confidence interval for tuples %-change: -8.01% -7.15%
Tuples are helped.
total clauses in shared programs: 357354 -> 356610 (-0.21%)
clauses in affected programs: 25794 -> 25050 (-2.88%)
helped: 994
HURT: 317
helped stats (abs) min: 1.0 max: 3.0 x̄: 1.16 x̃: 1
helped stats (rel) min: 1.49% max: 33.33% x̄: 10.78% x̃: 10.00%
HURT stats (abs) min: 1.0 max: 4.0 x̄: 1.31 x̃: 1
HURT stats (rel) min: 1.19% max: 50.00% x̄: 13.56% x̃: 8.33%
95% mean confidence interval for clauses value: -0.63 -0.50
95% mean confidence interval for clauses %-change: -5.63% -4.16%
Clauses are helped.
total cycles in shared programs: 167697.96 -> 167431.15 (-0.16%)
cycles in affected programs: 12638.29 -> 12371.48 (-2.11%)
helped: 2652
HURT: 350
helped stats (abs) min: 0.04166399999999726 max: 0.75 x̄: 0.11 x̃: 0
helped stats (rel) min: 0.12% max: 100.00% x̄: 14.39% x̃: 5.04%
HURT stats (abs) min: 0.041665999999999315 max: 0.5833329999999997 x̄: 0.11 x̃: 0
HURT stats (rel) min: 0.00% max: 75.00% x̄: 7.90% x̃: 4.71%
95% mean confidence interval for cycles value: -0.09 -0.08
95% mean confidence interval for cycles %-change: -12.56% -11.02%
Cycles are helped.
total arith in shared programs: 74169.46 -> 73891.71 (-0.37%)
arith in affected programs: 13885.87 -> 13608.12 (-2.00%)
helped: 3215
HURT: 445
helped stats (abs) min: 0.04166399999999726 max: 0.5416680000000014 x̄: 0.10 x̃: 0
helped stats (rel) min: 0.12% max: 100.00% x̄: 14.16% x̃: 6.67%
HURT stats (abs) min: 0.041665999999999315 max: 1.125 x̄: 0.12 x̃: 0
HURT stats (rel) min: 0.00% max: 100.00% x̄: 9.76% x̃: 5.49%
95% mean confidence interval for arith value: -0.08 -0.07
95% mean confidence interval for arith %-change: -11.91% -10.59%
Arith are helped.
total texture in shared programs: 11936 -> 11931 (-0.04%)
texture in affected programs: 20 -> 15 (-25.00%)
helped: 10
HURT: 0
helped stats (abs) min: 0.5 max: 0.5 x̄: 0.50 x̃: 0
helped stats (rel) min: 14.29% max: 100.00% x̄: 45.71% x̃: 33.33%
95% mean confidence interval for texture value: -0.50 -0.50
95% mean confidence interval for texture %-change: -73.16% -18.26%
Texture are helped.
total vary in shared programs: 4180.88 -> 3447.19 (-17.55%)
vary in affected programs: 2109.88 -> 1376.19 (-34.77%)
helped: 2202
HURT: 39
helped stats (abs) min: 0.0625 max: 1.4375 x̄: 0.34 x̃: 0
helped stats (rel) min: 2.38% max: 66.67% x̄: 40.43% x̃: 50.00%
HURT stats (abs) min: 0.125 max: 0.375 x̄: 0.26 x̃: 0
HURT stats (rel) min: 0.00% max: 300.00% x̄: 92.54% x̃: 23.08%
95% mean confidence interval for vary value: -0.34 -0.32
95% mean confidence interval for vary %-change: -39.22% -37.01%
Vary are helped.
total quadwords in shared programs: 1689664 -> 1684852 (-0.28%)
quadwords in affected programs: 265522 -> 260710 (-1.81%)
helped: 2864
HURT: 447
helped stats (abs) min: 1.0 max: 14.0 x̄: 2.10 x̃: 2
helped stats (rel) min: 0.15% max: 31.58% x̄: 6.05% x̃: 4.65%
HURT stats (abs) min: 1.0 max: 22.0 x̄: 2.67 x̃: 2
HURT stats (rel) min: 0.27% max: 38.46% x̄: 6.79% x̃: 4.55%
95% mean confidence interval for quadwords value: -1.54 -1.37
95% mean confidence interval for quadwords %-change: -4.55% -4.08%
Quadwords are helped.
total threads in shared programs: 53656 -> 53688 (0.06%)
threads in affected programs: 32 -> 64 (100.00%)
helped: 32
HURT: 0
helped stats (abs) min: 1.0 max: 1.0 x̄: 1.00 x̃: 1
helped stats (rel) min: 100.00% max: 100.00% x̄: 100.00% x̃: 100.00%
95% mean confidence interval for threads value: 1.00 1.00
95% mean confidence interval for threads %-change: 100.00% 100.00%
Threads are helped.
total preloads in shared programs: 116212 -> 103476 (-10.96%)
preloads in affected programs: 45222 -> 32486 (-28.16%)
helped: 3022
HURT: 11
helped stats (abs) min: 1.0 max: 11.0 x̄: 4.23 x̃: 4
helped stats (rel) min: 7.14% max: 68.75% x̄: 30.39% x̃: 25.00%
HURT stats (abs) min: 2.0 max: 4.0 x̄: 3.45 x̃: 4
HURT stats (rel) min: 14.29% max: 50.00% x̄: 25.93% x̃: 25.00%
95% mean confidence interval for preloads value: -4.26 -4.14
95% mean confidence interval for preloads %-change: -30.68% -29.69%
Preloads are helped.