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/aarch64/h264cmc_neon.S | 59 ++++++++++++++++++++++++++++++++++++---
 1 file changed, 55 insertions(+), 4 deletions(-)

(limited to 'libavcodec/aarch64')

diff --git a/libavcodec/aarch64/h264cmc_neon.S b/libavcodec/aarch64/h264cmc_neon.S
index 50d12b4e20..d1025c7a25 100644
--- a/libavcodec/aarch64/h264cmc_neon.S
+++ b/libavcodec/aarch64/h264cmc_neon.S
@@ -95,9 +95,10 @@ function ff_\type\()_\codec\()_chroma_mc8_neon, export=1
         b.gt            1b
         ret
 
-2:      tst             w6,  w6
-        add             w12, w12, w6
+2:      adds            w12, w12, w6
         dup             v0.8B, w4
+        b.eq            5f
+        tst             w6,  w6
         dup             v1.8B, w12
         b.eq            4f
 
@@ -161,6 +162,33 @@ function ff_\type\()_\codec\()_chroma_mc8_neon, export=1
         st1             {v17.8B}, [x0], x2
         b.gt            4b
         ret
+
+5:      ld1             {v4.8B}, [x1], x2
+        ld1             {v5.8B}, [x1], x2
+        prfm            pldl1strm, [x1]
+        subs            w3,  w3,  #2
+        umull           v16.8H, v4.8B, v0.8B
+        umull           v17.8H, v5.8B, v0.8B
+        prfm            pldl1strm, [x1, x2]
+  .ifc \codec,h264
+        rshrn           v16.8B, v16.8H, #6
+        rshrn           v17.8B, v17.8H, #6
+  .else
+        add             v16.8H, v16.8H, v22.8H
+        add             v17.8H, v17.8H, v22.8H
+        shrn            v16.8B, v16.8H, #6
+        shrn            v17.8B, v17.8H, #6
+  .endif
+  .ifc \type,avg
+        ld1             {v20.8B}, [x8], x2
+        ld1             {v21.8B}, [x8], x2
+        urhadd          v16.8B, v16.8B, v20.8B
+        urhadd          v17.8B, v17.8B, v21.8B
+  .endif
+        st1             {v16.8B}, [x0], x2
+        st1             {v17.8B}, [x0], x2
+        b.gt            5b
+        ret
 endfunc
 .endm
 
@@ -238,9 +266,10 @@ function ff_\type\()_\codec\()_chroma_mc4_neon, export=1
         b.gt            1b
         ret
 
-2:      tst             w6,  w6
-        add             w12, w12, w6
+2:      adds            w12, w12, w6
         dup             v30.8B, w4
+        b.eq            5f
+        tst             w6,  w6
         dup             v31.8B, w12
         trn1            v0.2S,  v30.2S, v31.2S
         trn2            v1.2S,  v30.2S, v31.2S
@@ -303,6 +332,28 @@ function ff_\type\()_\codec\()_chroma_mc4_neon, export=1
         st1             {v16.S}[1], [x0], x2
         b.gt            4b
         ret
+
+5:      ld1             {v4.S}[0], [x1], x2
+        ld1             {v4.S}[1], [x1], x2
+        umull           v18.8H, v4.8B,  v30.8B
+        subs            w3,  w3,  #2
+        prfm            pldl1strm, [x1]
+  .ifc \codec,h264
+        rshrn           v16.8B, v18.8H, #6
+  .else
+        add             v18.8H, v18.8H, v22.8H
+        shrn            v16.8B, v18.8H, #6
+  .endif
+  .ifc \type,avg
+        ld1             {v20.S}[0], [x8], x2
+        ld1             {v20.S}[1], [x8], x2
+        urhadd          v16.8B, v16.8B, v20.8B
+  .endif
+        prfm            pldl1strm, [x1]
+        st1             {v16.S}[0], [x0], x2
+        st1             {v16.S}[1], [x0], x2
+        b.gt            5b
+        ret
 endfunc
 .endm
 
-- 
cgit v1.2.3