ref: a364126820f4d7d1786d79b08340149a38dc00a3
parent: 2eaabafce3b01fc291b570d54612260fd2c163b1
author: Martin Storsjö <martin@martin.st>
date: Tue Oct 8 05:40:17 EDT 2019
arm64: looprestoration: Use ld2r instead of ld1+dup+dup
--- a/src/arm/64/looprestoration.S
+++ b/src/arm/64/looprestoration.S
@@ -1894,12 +1894,12 @@
// const int16_t wt[2]);
function sgr_weighted2_neon, export=1
ldr x8, [sp]
- ld1 {v31.s}[0], [x8]cmp x7, #2
add x10, x0, x1
add x11, x2, x3
add x12, x4, #2*FILTER_OUT_STRIDE
add x13, x5, #2*FILTER_OUT_STRIDE
+ ld2r {v30.8h, v31.8h}, [x8] // wt[0], wt[1]mov x8, #4*FILTER_OUT_STRIDE
lsl x1, x1, #1
lsl x3, x3, #1
@@ -1908,8 +1908,6 @@
sub x1, x1, x9
sub x3, x3, x9
sub x8, x8, x9, lsl #1
- dup v30.8h, v31.h[0] // wt[0]
- dup v31.8h, v31.h[1] // wt[1]
mov x9, x6
b.lt 2f
1: