Mesa (main): nir/algebraic: Basic patterns for dot_4x8

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 806cd2341c8b6756a4c8a9d9c6e61ece6a5e604b
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=806cd2341c8b6756a4c8a9d9c6e61ece6a5e604b

Author: Ian Romanick <ian.d.romanick at intel.com>
Date:   Tue Feb 23 19:12:49 2021 -0800

nir/algebraic: Basic patterns for dot_4x8

v2: Add and modify patterns to let constant folding do better.

v3: Remove '(is_not_zero)' from the patterns that try to combine
addends.  I honestly don't know why I had it there in the first place,
and nothing in my deep git logs could help clue me in.  Noticed by
Alyssa.  Remover patterns that detect open-coded udot_4x8.  Suggested by
Alyssa and Jason.  Add missing sudot_4x8 patterns.

Reviewed-by: Jason Ekstrand <jason at jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12142>

---

 src/compiler/nir/nir_opt_algebraic.py | 29 +++++++++++++++++++++++++++++
 1 file changed, 29 insertions(+)

diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
index 001a86918d4..2f0a044e2a8 100644
--- a/src/compiler/nir/nir_opt_algebraic.py
+++ b/src/compiler/nir/nir_opt_algebraic.py
@@ -192,6 +192,35 @@ optimizations = [
 
    # flrp(a, a + b, c) => a + flrp(0, b, c) => a + (b * c)
    (('~flrp', a, ('fadd(is_used_once)', a, b), c), ('fadd', ('fmul', b, c), a)),
+
+   (('sdot_4x8_iadd', a, 0, b), b),
+   (('udot_4x8_uadd', a, 0, b), b),
+   (('sdot_4x8_iadd_sat', a, 0, b), b),
+   (('udot_4x8_uadd_sat', a, 0, b), b),
+
+   # sudot_4x8_iadd is not commutative at all, so the patterns must be
+   # duplicated with zeros on each of the first positions.
+   (('sudot_4x8_iadd', a, 0, b), b),
+   (('sudot_4x8_iadd', 0, a, b), b),
+   (('sudot_4x8_iadd_sat', a, 0, b), b),
+   (('sudot_4x8_iadd_sat', 0, a, b), b),
+
+   (('iadd', ('sdot_4x8_iadd(is_used_once)', a, b, '#c'), '#d'), ('sdot_4x8_iadd', a, b, ('iadd', c, d))),
+   (('iadd', ('udot_4x8_uadd(is_used_once)', a, b, '#c'), '#d'), ('udot_4x8_uadd', a, b, ('iadd', c, d))),
+   (('iadd', ('sudot_4x8_iadd(is_used_once)', a, b, '#c'), '#d'), ('sudot_4x8_iadd', a, b, ('iadd', c, d))),
+
+   # Try to let constant folding eliminate the dot-product part.  These are
+   # safe because the dot product cannot overflow 32 bits.
+   (('iadd', ('sdot_4x8_iadd', 'a(is_not_const)', b, 0), c), ('sdot_4x8_iadd', a, b, c)),
+   (('iadd', ('udot_4x8_uadd', 'a(is_not_const)', b, 0), c), ('udot_4x8_uadd', a, b, c)),
+   (('iadd', ('sudot_4x8_iadd', 'a(is_not_const)', b, 0), c), ('sudot_4x8_iadd', a, b, c)),
+   (('iadd', ('sudot_4x8_iadd', a, 'b(is_not_const)', 0), c), ('sudot_4x8_iadd', a, b, c)),
+   (('sdot_4x8_iadd', '#a', '#b', 'c(is_not_const)'), ('iadd', ('sdot_4x8_iadd', a, b, 0), c)),
+   (('udot_4x8_uadd', '#a', '#b', 'c(is_not_const)'), ('iadd', ('udot_4x8_uadd', a, b, 0), c)),
+   (('sudot_4x8_iadd', '#a', '#b', 'c(is_not_const)'), ('iadd', ('sudot_4x8_iadd', a, b, 0), c)),
+   (('sdot_4x8_iadd_sat', '#a', '#b', 'c(is_not_const)'), ('iadd_sat', ('sdot_4x8_iadd', a, b, 0), c), '!options->lower_add_sat'),
+   (('udot_4x8_uadd_sat', '#a', '#b', 'c(is_not_const)'), ('uadd_sat', ('udot_4x8_uadd', a, b, 0), c), '!options->lower_add_sat'),
+   (('sudot_4x8_iadd_sat', '#a', '#b', 'c(is_not_const)'), ('iadd_sat', ('sudot_4x8_iadd', a, b, 0), c), '!options->lower_add_sat'),
 ]
 
 # Float sizes