Mesa (master): nir/algebraic: Use value range analysis to convert fmax to fsat

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: af1bc7e0c7dd1f3c4f2226f93e819e410fd7a731
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=af1bc7e0c7dd1f3c4f2226f93e819e410fd7a731

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Mon Feb 10 18:23:54 2020 -0800

nir/algebraic: Use value range analysis to convert fmax to fsat

This is conceptually similar to the 1-fsat(a) <=> fsat(1-a) rearragement
done in:

3b747909419 ("nir/algebraic: Recognize open-coded flrp(a, b, fsat(c))")

2d259713b7 ("nir/algebraic: Commute 1-fsat(a) to fsat(1-a) for all
non-fmul instructions").

Note: this helps the Aztex Ruins shader that was hurt for spills and
fills on Braodwell in the previous commit, but it does not fix the
spills or fills. :(

All Intel platforms had similar results. (Ice Lake shown)
total instructions in shared programs: 14528985 -> 14526116 (-0.02%)
instructions in affected programs: 477300 -> 474431 (-0.60%)
helped: 2332
HURT: 0
helped stats (abs) min: 1 max: 18 x̄: 1.23 x̃: 1
helped stats (rel) min: 0.07% max: 8.89% x̄: 0.88% x̃: 0.64%
95% mean confidence interval for instructions value: -1.27 -1.19
95% mean confidence interval for instructions %-change: -0.92% -0.85%
Instructions are helped.

total cycles in shared programs: 203723684 -> 203692984 (-0.02%)
cycles in affected programs: 4878847 -> 4848147 (-0.63%)
helped: 1764
HURT: 324
helped stats (abs) min: 1 max: 706 x̄: 22.94 x̃: 17
helped stats (rel) min: <.01% max: 17.75% x̄: 1.94% x̃: 1.66%
HURT stats (abs)   min: 1 max: 400 x̄: 30.15 x̃: 10
HURT stats (rel)   min: <.01% max: 17.76% x̄: 1.91% x̃: 0.69%
95% mean confidence interval for cycles value: -16.55 -12.86
95% mean confidence interval for cycles %-change: -1.44% -1.24%
Cycles are helped.

Reviewed-by: Matt Turner <mattst88 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/1359>

---

 src/compiler/nir/nir_opt_algebraic.py | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 76ec1acc056..bb490d2a19c 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -560,6 +560,16 @@ optimizations.extend([
    (('umin', ('umax', ('umin', ('umax', a, b), c), b), c), ('umin', ('umax', a, b), c)),
    (('fmax', ('fsat', a), '#b at 32(is_zero_to_one)'), ('fsat', ('fmax', a, b))),
    (('fmin', ('fsat', a), '#b at 32(is_zero_to_one)'), ('fsat', ('fmin', a, b))),
+
+   # If a in [0,b] then b-a is also in [0,b].  Since b in [0,1], max(b-a, 0) =
+   # fsat(b-a).
+   #
+   # If a > b, then b-a < 0 and max(b-a, 0) = fsat(b-a) = 0
+   #
+   # This should be NaN safe since max(NaN, 0) = fsat(NaN) = 0.
+   (('fmax', ('fadd(is_used_once)', ('fneg', 'a(is_not_negative)'), '#b at 32(is_zero_to_one)'), 0.0),
+    ('fsat', ('fadd', ('fneg',  a), b)), '!options->lower_fsat'),
+
    (('extract_u8', ('imin', ('imax', a, 0), 0xff), 0), ('imin', ('imax', a, 0), 0xff)),
    (('~ior', ('flt(is_used_once)', a, b), ('flt', a, c)), ('flt', a, ('fmax', b, c))),
    (('~ior', ('flt(is_used_once)', a, c), ('flt', b, c)), ('flt', ('fmin', a, b), c)),