projects
/
platform
/
upstream
/
mesa.git
/ commitdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
| commitdiff |
tree
raw
|
patch
| inline |
side by side
(parent:
b3ed522
)
agx: Lower f2u8/f2i8
author
Alyssa Rosenzweig
<alyssa@rosenzweig.io>
Fri, 11 Aug 2023 13:58:33 +0000
(09:58 -0400)
committer
Marge Bot
<emma+marge@anholt.net>
Fri, 11 Aug 2023 20:31:28 +0000
(20:31 +0000)
The less 8-bit we need to emulate in the backend the better.
Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24635>
src/asahi/compiler/agx_nir_algebraic.py
patch
|
blob
|
history
diff --git
a/src/asahi/compiler/agx_nir_algebraic.py
b/src/asahi/compiler/agx_nir_algebraic.py
index
639cba5
..
9fd4a40
100644
(file)
--- a/
src/asahi/compiler/agx_nir_algebraic.py
+++ b/
src/asahi/compiler/agx_nir_algebraic.py
@@
-39,6
+39,10
@@
lower_pack = [
(('u2f32', ('u2u32', a)), ('u2f32', a)),
(('i2f32', ('i2i32', a)), ('i2f32', a)),
+ # Chew through some 8-bit before the backend has to deal with it
+ (('f2u8', a), ('u2u8', ('f2u16', a))),
+ (('f2i8', a), ('i2i8', ('f2i16', a))),
+
# These are based on the lowerings from nir_opt_algebraic, but conditioned
# on the number of bits not being constant. If the bit count is constant
# (the happy path) we can use our native instruction instead.