mirror of
https://gitlab.freedesktop.org/mesa/mesa.git
synced 2026-05-24 08:28:16 +02:00
05703a49f9
This matches what NIR does. See nir_vector_extract.
This improves code generation for several reasons. First, it only
requires 3 comparisons instead of 4 (vec3(i > 0, i > 1, i > 2) vs
vec4(i == 0, i == 1, i == 2, i == 3)).
Secondly, it shortens the liverange of some values, possibly quite
dramatically. Consider a loop in the old version (after lowering
if-statements to selects):
loop {
...
x = csel(i == 0, a[0], x);
x = csel(i == 1, a[1], x);
x = csel(i == 2, a[2], x);
x = csel(i == 3, a[3], x);
...
}
x is live for the whole loop across iterations. In the new version, x
is only live while its value is needed:
loop {
...
t0 = csel(i > 0 , a[1], a[0]);
t1 = csel(i > 2 , a[3], a[2]);
x = csel(i > 1, t1, t0);
...
}
Outside a loop, this also means more values of the array may have their
liveness reduced sooner (by consuming two values at once).
All Intel platforms had similar results. (Tigerlake shown)
total instructions in shared programs: 21171336 -> 21163615 (-0.04%)
instructions in affected programs: 89680 -> 81959 (-8.61%)
helped: 40
HURT: 4
helped stats (abs) min: 1 max: 450 x̄: 193.68 x̃: 196
helped stats (rel) min: 0.41% max: 13.32% x̄: 6.01% x̃: 6.22%
HURT stats (abs) min: 1 max: 12 x̄: 6.50 x̃: 6
HURT stats (rel) min: 0.50% max: 0.66% x̄: 0.58% x̃: 0.58%
95% mean confidence interval for instructions value: -229.68 -121.28
95% mean confidence interval for instructions %-change: -6.93% -3.89%
Instructions are helped.
total cycles in shared programs: 832879641 -> 829513122 (-0.40%)
cycles in affected programs: 44738430 -> 41371911 (-7.52%)
helped: 35
HURT: 2
helped stats (abs) min: 2 max: 189948 x̄: 96186.49 x̃: 116154
helped stats (rel) min: 0.37% max: 11.08% x̄: 5.88% x̃: 6.47%
HURT stats (abs) min: 4 max: 4 x̄: 4.00 x̃: 4
HURT stats (rel) min: 0.69% max: 0.69% x̄: 0.69% x̃: 0.69%
95% mean confidence interval for cycles value: -112881.94 -69092.06
95% mean confidence interval for cycles %-change: -6.77% -4.27%
Cycles are helped.
total spills in shared programs: 8061 -> 7338 (-8.97%)
spills in affected programs: 873 -> 150 (-82.82%)
helped: 24
HURT: 0
total fills in shared programs: 7501 -> 6388 (-14.84%)
fills in affected programs: 1389 -> 276 (-80.13%)
helped: 24
HURT: 0
Radeon R430
total instructions in shared programs: 2449852 -> 2449136 (-0.03%)
instructions in affected programs: 6285 -> 5569 (-11.39%)
helped: 64
HURT: 0
helped stats (abs) min: 4 max: 12 x̄: 11.19 x̃: 12
helped stats (rel) min: 8.16% max: 21.62% x̄: 12.09% x̃: 10.91%
total consts in shared programs: 1032517 -> 1032482 (<.01%)
consts in affected programs: 966 -> 931 (-3.62%)
helped: 35
HURT: 0
helped stats (abs) min: 1 max: 1 x̄: 1.00 x̃: 1
helped stats (rel) min: 2.94% max: 10.00% x̄: 4.26% x̃: 3.57%
Reviewed-by: Matt Turner <mattst88@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14573>
|
||
|---|---|---|
| .. | ||
| amd | ||
| android_stub | ||
| asahi | ||
| broadcom | ||
| compiler | ||
| drm-shim | ||
| egl | ||
| etnaviv | ||
| freedreno | ||
| gallium | ||
| gbm | ||
| getopt | ||
| glx | ||
| gtest | ||
| hgl | ||
| imgui | ||
| intel | ||
| loader | ||
| mapi | ||
| mesa | ||
| microsoft | ||
| nouveau | ||
| panfrost | ||
| tool | ||
| util | ||
| virtio | ||
| vulkan | ||
| meson.build | ||