From f23d26a6864128001b03876b0b92fffe131f2060 Mon Sep 17 00:00:00 2001
From: Janne Grunau <janne-libav@jannau.net>
Date: Wed, 18 Jun 2014 17:10:42 +0200
Subject: h264: avoid using uninitialized memory in NEON chroma mc

Adapt commit 982b596ea6640bfe218a31f6c3fc542d9fe61c31 for the arm and
aarch64 NEON asm. 5-10% faster on Cortex-A9.
---
 libavcodec/arm/h264cmc_neon.S | 60 ++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 56 insertions(+), 4 deletions(-)

(limited to 'libavcodec/arm')

diff --git a/libavcodec/arm/h264cmc_neon.S b/libavcodec/arm/h264cmc_neon.S
index 3183dd8d82..ee7011b00b 100644
--- a/libavcodec/arm/h264cmc_neon.S
+++ b/libavcodec/arm/h264cmc_neon.S
@@ -96,9 +96,10 @@ T       cmp             r7,  #0
 
         pop             {r4-r7, pc}
 
-2:      tst             r6,  r6
-        add             r12, r12, r6
+2:      adds            r12, r12, r6
         vdup.8          d0,  r4
+        beq             5f
+        tst             r6,  r6
         vdup.8          d1,  r12
 
         beq             4f
@@ -163,6 +164,33 @@ T       cmp             r7,  #0
         vst1.8          {d17}, [r0,:64], r2
         bgt             4b
 
+        pop             {r4-r7, pc}
+
+5:      vld1.8          {d4}, [r1], r2
+        vld1.8          {d5}, [r1], r2
+        pld             [r1]
+        subs            r3,  r3,  #2
+        vmull.u8        q8,  d4,  d0
+        vmull.u8        q9,  d5,  d0
+        pld             [r1, r2]
+  .ifc \codec,h264
+        vrshrn.u16      d16, q8,  #6
+        vrshrn.u16      d17, q9,  #6
+  .else
+        vadd.u16        q8,  q8,  q11
+        vadd.u16        q9,  q9,  q11
+        vshrn.u16       d16, q8,  #6
+        vshrn.u16       d17, q9,  #6
+  .endif
+  .ifc \type,avg
+        vld1.8          {d20}, [lr,:64], r2
+        vld1.8          {d21}, [lr,:64], r2
+        vrhadd.u8       q8,  q8,  q10
+  .endif
+        vst1.8          {d16}, [r0,:64], r2
+        vst1.8          {d17}, [r0,:64], r2
+        bgt             5b
+
         pop             {r4-r7, pc}
 endfunc
 .endm
@@ -245,9 +273,10 @@ T       cmp             r7,  #0
 
         pop             {r4-r7, pc}
 
-2:      tst             r6,  r6
-        add             r12, r12, r6
+2:      adds            r12, r12, r6
         vdup.8          d0,  r4
+        beq             5f
+        tst             r6,  r6
         vdup.8          d1,  r12
         vtrn.32         d0,  d1
 
@@ -310,6 +339,29 @@ T       cmp             r7,  #0
         vst1.32         {d16[1]}, [r0,:32], r2
         bgt             4b
 
+        pop             {r4-r7, pc}
+
+5:      vld1.32         {d4[0]},  [r1], r2
+        vld1.32         {d4[1]},  [r1], r2
+        vmull.u8        q8,  d4,  d0
+        subs            r3,  r3,  #2
+        pld             [r1]
+  .ifc \codec,h264
+        vrshrn.u16      d16, q8,  #6
+  .else
+        vadd.u16        q8,  q8,  q11
+        vshrn.u16       d16, q8,  #6
+  .endif
+  .ifc \type,avg
+        vld1.32         {d20[0]}, [lr,:32], r2
+        vld1.32         {d20[1]}, [lr,:32], r2
+        vrhadd.u8       d16, d16, d20
+  .endif
+        pld             [r1]
+        vst1.32         {d16[0]}, [r0,:32], r2
+        vst1.32         {d16[1]}, [r0,:32], r2
+        bgt             5b
+
         pop             {r4-r7, pc}
 endfunc
 .endm
-- 
cgit v1.2.3