The latter is 1 cycle faster on a cortex-53 and since the operands are
bytewise (or larger) bitmask (impossible to overflow to zero) both are
equivalent.
---
libavcodec/aarch64/vp9lpf_neon.S | 31 ++++++++++++++++++++-----------
1 file changed, 20 insertions(+), 11 deletions(-)
diff --git a/libavcodec/aarch64/vp9lpf_neon.S b/libavcodec/aarch64/vp9lpf_neon.S
index 392794b..e9c7d9e 100644
--- a/libavcodec/aarch64/vp9lpf_neon.S
+++ b/libavcodec/aarch64/vp9lpf_neon.S
@@ -218,13 +218,15 @@
xtn_sz v5, v6.8h, v7.8h, \sz
and v4\sz, v4\sz, v5\sz // fm
+ // If no pixels need filtering, just exit as soon as possible
mov x5, v4.d[0]
.ifc \sz, .16b
mov x6, v4.d[1]
- orr x5, x5, x6
-.endif
- // If no pixels need filtering, just exit as soon as possible
+ adds x5, x5, x6
+ b.eq 9f
+.else
cbz x5, 9f
+.endif
.if \wd >= 8
movi v0\sz, #1
@@ -344,15 +346,17 @@
bit v22\sz, v0\sz, v5\sz // if (!hev && fm &&
!flat8in)
bit v25\sz, v2\sz, v5\sz
+ // If no pixels need flat8in, jump to flat8out
+ // (or to a writeout of the inner 4 pixels, for wd=8)
.if \wd >= 8
mov x5, v6.d[0]
.ifc \sz, .16b
mov x6, v6.d[1]
- orr x5, x5, x6
-.endif
- // If no pixels need flat8in, jump to flat8out
- // (or to a writeout of the inner 4 pixels, for wd=8)
+ adds x5, x5, x6
+ b.eq 6f
+.else
cbz x5, 6f
+.endif
// flat8in
uaddl_sz \tmp1\().8h, \tmp2\().8h, v20, v21, \sz
@@ -406,20 +410,25 @@
mov x5, v2.d[0]
.ifc \sz, .16b
mov x6, v2.d[1]
- orr x5, x5, x6
+ adds x5, x5, x6
+ b.ne 1f
+.else
+ cbnz x5, 1f
.endif
// If no pixels needed flat8in nor flat8out, jump to a
// writeout of the inner 4 pixels
- cbnz x5, 1f
br x14
1:
+
mov x5, v7.d[0]
.ifc \sz, .16b
mov x6, v7.d[1]
- orr x5, x5, x6
+ adds x5, x5, x6
+ b.ne 1f
+.else
+ cbnz x5, 1f
.endif
// If no pixels need flat8out, jump to a writeout of the inner 6 pixels
- cbnz x5, 1f
br x15
1:
--
2.10.2
