Mesa (staging/22.1): nir/algebraic: Fix NaN-unsafe fcsel patterns

[GitLab Mirror]

Module: Mesa
Branch: staging/22.1
Commit: ac10f5a3ab11ae1422452f4cac1099ca27313d2b
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=ac10f5a3ab11ae1422452f4cac1099ca27313d2b

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Thu Mar 25 13:13:21 2021 -0700

nir/algebraic: Fix NaN-unsafe fcsel patterns

For example, the proof for this pattern

    (('bcsel', ('flt', 'a at 32', 0), 'b at 32', 'c at 32'), ('fcsel_ge', a, c, b)),

would be

    bcsel(a < 0, b, c)
    bcsel(!(a < 0), c, b)
    bcsel(a >= 0, c, b)
    fcsel_ge(a, c, b)

However, !(a < 0) => (a >= 0) is well known to produce different
results if `a` is NaN.

Instead of that replacement, use this replacement:

    bcsel(a < 0, b, c)
    bcsel(-0 < -a, b, c)
    bcsel(0 < -a, b, c)
    fcsel_gt(-a, b, c)

This is NaN-safe and exact.

Reviewed-by: Alyssa Rosenzweig <alyssa at collabora.com>
Reviewed-by: Jason Ekstrand <jason.ekstrand at collabora.com>
Fixes: 0f5b3c37c5d ("nir: Add opcodes for fused comp + csel and optimizations")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17048>
(cherry picked from commit a2a2fbc5101d6e6b5d18903c7a0bd84037dbdddc)

---

 .pick_status.json                     | 2 +-
 src/compiler/nir/nir_opt_algebraic.py | 8 ++++----
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/.pick_status.json b/.pick_status.json
index 9481b740c65..987e35235c9 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -319,7 +319,7 @@
         "description": "nir/algebraic: Fix NaN-unsafe fcsel patterns",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": "0f5b3c37c5d757f6ffe994bae24071c0462bb13f"
     },
diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index d45adbfe147..4cc415131d5 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -1972,9 +1972,9 @@ optimizations.extend([
    (('imul24', a, 0), (0)),
 
    (('fcsel', ('slt', 0, a), b, c), ('fcsel_gt', a, b, c), "options->has_fused_comp_and_csel"),
-   (('fcsel', ('slt', a, 0), b, c), ('fcsel_ge', a, c, b), "options->has_fused_comp_and_csel"),
+   (('fcsel', ('slt', a, 0), b, c), ('fcsel_gt', ('fneg', a), b, c), "options->has_fused_comp_and_csel"),
    (('fcsel', ('sge', a, 0), b, c), ('fcsel_ge', a, b, c), "options->has_fused_comp_and_csel"),
-   (('fcsel', ('sge', 0, a), b, c), ('fcsel_gt', a, c, b), "options->has_fused_comp_and_csel"),
+   (('fcsel', ('sge', 0, a), b, c), ('fcsel_ge', ('fneg', a), b, c), "options->has_fused_comp_and_csel"),
 
    (('bcsel', ('ilt', 0, 'a at 32'), 'b at 32', 'c at 32'), ('i32csel_gt', a, b, c), "options->has_fused_comp_and_csel"),
    (('bcsel', ('ilt', 'a at 32', 0), 'b at 32', 'c at 32'), ('i32csel_ge', a, c, b), "options->has_fused_comp_and_csel"),
@@ -1982,9 +1982,9 @@ optimizations.extend([
    (('bcsel', ('ige', 0, 'a at 32'), 'b at 32', 'c at 32'), ('i32csel_gt', a, c, b), "options->has_fused_comp_and_csel"),
 
    (('bcsel', ('flt', 0, 'a at 32'), 'b at 32', 'c at 32'), ('fcsel_gt', a, b, c), "options->has_fused_comp_and_csel"),
-   (('bcsel', ('flt', 'a at 32', 0), 'b at 32', 'c at 32'), ('fcsel_ge', a, c, b), "options->has_fused_comp_and_csel"),
+   (('bcsel', ('flt', 'a at 32', 0), 'b at 32', 'c at 32'), ('fcsel_gt', ('fneg', a), b, c), "options->has_fused_comp_and_csel"),
    (('bcsel', ('fge', 'a at 32', 0), 'b at 32', 'c at 32'), ('fcsel_ge', a, b, c), "options->has_fused_comp_and_csel"),
-   (('bcsel', ('fge', 0, 'a at 32'), 'b at 32', 'c at 32'), ('fcsel_gt', a, c, b), "options->has_fused_comp_and_csel"),
+   (('bcsel', ('fge', 0, 'a at 32'), 'b at 32', 'c at 32'), ('fcsel_ge', ('fneg', a), b, c), "options->has_fused_comp_and_csel"),
 
 ])