ref: 53c8af4566c435960263f66d53550411ae1a7827
parent: fe57aa46df2ae9eb54a371d32a3472b31f69efd6
author: zhiliang wang <[email protected]>
date: Thu May 15 11:04:44 EDT 2014
Refine some code
--- a/codec/encoder/core/arm64/pixel_neon_aarch64.S
+++ b/codec/encoder/core/arm64/pixel_neon_aarch64.S
@@ -35,20 +35,16 @@
#include "arm_arch64_common_macro.S"
.macro CALC_AND_STORE_SAD
- uaddlp v2.4s, v2.8h
- addp v2.4s, v2.4s, v2.4s
- addp v2.2s, v2.2s, v2.2s
- umov w0, v2.s[0]
+ saddlv s2, v2.8h
+ fmov w0, s2
.endm
.macro CALC_AND_STORE_SAD_FOUR
- addp v0.8h, v28.8h, v29.8h
- addp v1.8h, v30.8h, v31.8h
- addp v0.8h, v0.8h, v1.8h
- addp v0.8h, v0.8h, v0.8h
- eor v1.8b, v1.8b, v1.8b
- saddl v0.4s, v0.4h, v1.4h
- st1 {v0.4s}, [x4]+ saddlv s28, v28.8h
+ saddlv s29, v29.8h
+ saddlv s30, v30.8h
+ saddlv s31, v31.8h
+ st4 {v28.s, v29.s, v30.s, v31.s}[0], [x4].endm
.macro LOAD_8X8_1
@@ -250,9 +246,8 @@
ld1 {v1.s}[0], [x2], x3uabal v2.8h, v0.8b, v1.8b
.endr
- uaddlp v2.2s, v2.4h
- addp v2.2s, v2.2s, v2.2s
- umov w0, v2.s[0]
+ saddlv s2, v2.4h
+ fmov w0, s2
WELS_ASM_ARCH64_FUNC_END
WELS_ASM_ARCH64_FUNC_BEGIN WelsSampleSad8x8_AArch64_neon