[vlc-commits] packetizer: add SSE2 based AnnexB startcode helper

[Francois Cartegnie]

vlc | branch: master | Francois Cartegnie <fcvlcdev at free.fr> | Tue Jan  5 17:39:28 2016 +0100| [975713e89c0be2410e34834bb5a098b2fbcd2222] | committer: Francois Cartegnie

packetizer: add SSE2 based AnnexB startcode helper

Improves even more from previous commit (by ~2x on 4K)

> http://git.videolan.org/gitweb.cgi/vlc.git/?a=commit;h=975713e89c0be2410e34834bb5a098b2fbcd2222
---

 modules/packetizer/startcode_helper.h |   87 +++++++++++++++++++++++++++------
 1 file changed, 72 insertions(+), 15 deletions(-)

diff --git a/modules/packetizer/startcode_helper.h b/modules/packetizer/startcode_helper.h
index 9db67fd..81a5b79 100644
--- a/modules/packetizer/startcode_helper.h
+++ b/modules/packetizer/startcode_helper.h
@@ -20,10 +20,75 @@
 #ifndef _STARTCODE_HELPER_H
 #define _STARTCODE_HELPER_H 1
 
+#if defined(HAVE_SSE2_INTRINSICS)
+   #include <emmintrin.h>
+#endif
+
 /* Looks up efficiently for an AnnexB startcode 0x00 0x00 0x01
- * by using a 4 times faster trick than single byte lookup.
- *
- * That code is adapted from libav's ff_avc_find_startcode_internal
+ * by using a 4 times faster trick than single byte lookup. */
+
+#define TRY_MATCH(p,a) {\
+     if (p[a+1] == 0) {\
+            if (p[a+0] == 0 && p[a+2] == 1)\
+                return a+p;\
+            if (p[a+2] == 0 && p[a+3] == 1)\
+                return a+p+1;\
+        }\
+        if (p[a+3] == 0) {\
+            if (p[a+2] == 0 && p[a+4] == 1)\
+                return a+p+2;\
+            if (p[a+4] == 0 && p[a+5] == 1)\
+                return a+p+3;\
+        }\
+    }
+
+#if defined(HAVE_SSE2_INTRINSICS)
+
+__attribute__((target("sse2")))
+static inline const uint8_t * startcode_FindAnnexB( const uint8_t *p, const uint8_t *end )
+{
+    /* First align to 16 */
+    /* Skipping this step and doing unaligned loads isn't faster */
+    const uint8_t *alignedend = p + 16 - ((intptr_t)p & 15);
+    for (end -= 3; p < alignedend && p < end; p++) {
+        if (p[0] == 0 && p[1] == 0 && p[2] == 1)
+            return p;
+    }
+
+    if( p == end )
+        return NULL;
+
+    alignedend = end - ((intptr_t) end & 15);
+    if( alignedend > p )
+    {
+        __m128i zeros = _mm_set1_epi8( 0x00 );
+        for( ; p < alignedend; p += 16)
+        {
+            __m128i v = _mm_load_si128((__m128i*)p);
+            __m128i res = _mm_cmpeq_epi8( zeros, v );
+            uint16_t match = _mm_movemask_epi8( res ); /* mask will be in reversed match order */
+
+            if( match & 0x000F )
+                TRY_MATCH(p, 0);
+            if( match & 0x00F0 )
+                TRY_MATCH(p, 4);
+            if( match & 0x0F00 )
+                TRY_MATCH(p, 8);
+            if( match & 0xF000 )
+                TRY_MATCH(p, 12);
+        }
+    }
+
+    for (; p < end; p++) {
+        if (p[0] == 0 && p[1] == 0 && p[2] == 1)
+            return p;
+    }
+
+    return NULL;
+}
+
+#else
+/* That code is adapted from libav's ff_avc_find_startcode_internal
  * and i believe the trick originated from
  * https://graphics.stanford.edu/~seander/bithacks.html#ZeroInWord
  */
@@ -41,18 +106,7 @@ static inline const uint8_t * startcode_FindAnnexB( const uint8_t *p, const uint
         if ((x - 0x01010101) & (~x) & 0x80808080)
         {
             /* matching DW isn't faster */
-            if (p[1] == 0) {
-                if (p[0] == 0 && p[2] == 1)
-                    return p;
-                if (p[2] == 0 && p[3] == 1)
-                    return p+1;
-            }
-            if (p[3] == 0) {
-                if (p[2] == 0 && p[4] == 1)
-                    return p+2;
-                if (p[4] == 0 && p[5] == 1)
-                    return p+3;
-            }
+            TRY_MATCH(p, 0);
         }
     }
 
@@ -63,5 +117,8 @@ static inline const uint8_t * startcode_FindAnnexB( const uint8_t *p, const uint
 
     return NULL;
 }
+#endif
+
+#undef TRY_MATCH
 
 #endif