ref: 532b3a6720beb58eb24aab519ca6398353b2c10b
parent: ebfbf4efe65976ed7294b748420e3dd2e5b1050e
author: Martin Storsjö <martin@martin.st>
date: Thu Jan 31 05:22:47 EST 2019
arm64: looprestoration: Simplify the horizontal filtering of one pixel at a time
--- a/src/arm/64/looprestoration.S
+++ b/src/arm/64/looprestoration.S
@@ -282,19 +282,15 @@
addv h6, v6.8h
addv h7, v7.8h
dup v16.4h, v2.h[3]
- dup v17.4h, v4.h[3]
+ ins v16.h[1], v4.h[3]
+ ins v6.h[1], v7.h[0]
shl v16.4h, v16.4h, #7
- shl v17.4h, v17.4h, #7
sub v16.4h, v16.4h, v30.4h
- sub v17.4h, v17.4h, v30.4h
sqadd v6.4h, v6.4h, v16.4h
- sqadd v7.4h, v7.4h, v17.4h
sshr v6.4h, v6.4h, #3
- sshr v7.4h, v7.4h, #3
add v6.4h, v6.4h, v31.4h
- add v7.4h, v7.4h, v31.4h
st1 {v6.h}[0], [x0], #2- st1 {v7.h}[0], [x12], #2+ st1 {v6.h}[1], [x12], #2subs w5, w5, #1
ext v2.16b, v2.16b, v3.16b, #2
ext v4.16b, v4.16b, v5.16b, #2