From f9c8122cae4a1089f6fdd32b2e8b67208a400613 Mon Sep 17 00:00:00 2001
From: michael <michael@b3059339-0415-0410-9bf9-f77b7e298cf2>
Date: Tue, 25 Dec 2001 17:36:54 +0000
Subject: [PATCH] mix31to2 & mix31toS in SSE

git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@3741 b3059339-0415-0410-9bf9-f77b7e298cf2
---
 liba52/downmix.c | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 46 insertions(+)

diff --git a/liba52/downmix.c b/liba52/downmix.c
index d6991fffba..5bd51b4e36 100644
--- a/liba52/downmix.c
+++ b/liba52/downmix.c
@@ -521,11 +521,33 @@ static void mix31to2 (sample_t * samples, sample_t bias)
     int i;
     sample_t common;
 
+#ifdef HAVE_SSE
+	asm volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"movl $-1024, %%esi		\n\t"
+		"1:				\n\t"
+		"movaps 1024(%0, %%esi), %%xmm0	\n\t"  
+		"addps 3072(%0, %%esi), %%xmm0	\n\t"  
+		"addps %%xmm7, %%xmm0		\n\t" // common
+		"movaps (%0, %%esi), %%xmm1	\n\t" 
+		"movaps 2048(%0, %%esi), %%xmm2	\n\t"
+		"addps %%xmm0, %%xmm1		\n\t"
+		"addps %%xmm0, %%xmm2		\n\t"
+		"movaps %%xmm1, (%0, %%esi)	\n\t"
+		"movaps %%xmm2, 1024(%0, %%esi)	\n\t"
+		"addl $16, %%esi		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%esi"
+	);
+#else
     for (i = 0; i < 256; i++) {
 	common = samples[i + 256] + samples[i + 768] + bias;
 	samples[i] += common;
 	samples[i + 256] = samples[i + 512] + common;
     }
+#endif
 }
 
 static void mix31toS (sample_t * samples, sample_t bias)
@@ -533,12 +555,36 @@ static void mix31toS (sample_t * samples, sample_t bias)
     int i;
     sample_t common, surround;
 
+#ifdef HAVE_SSE
+	asm volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"movl $-1024, %%esi		\n\t"
+		"1:				\n\t"
+		"movaps 1024(%0, %%esi), %%xmm0	\n\t"  
+		"movaps 3072(%0, %%esi), %%xmm3	\n\t" // surround
+		"addps %%xmm7, %%xmm0		\n\t" // common
+		"movaps (%0, %%esi), %%xmm1	\n\t" 
+		"movaps 2048(%0, %%esi), %%xmm2	\n\t"
+		"addps %%xmm0, %%xmm1		\n\t"
+		"addps %%xmm0, %%xmm2		\n\t"
+		"subps %%xmm3, %%xmm1		\n\t"
+		"addps %%xmm3, %%xmm2		\n\t"
+		"movaps %%xmm1, (%0, %%esi)	\n\t"
+		"movaps %%xmm2, 1024(%0, %%esi)	\n\t"
+		"addl $16, %%esi		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%esi"
+	);
+#else
     for (i = 0; i < 256; i++) {
 	common = samples[i + 256] + bias;
 	surround = samples[i + 768];
 	samples[i] += common - surround;
 	samples[i + 256] = samples[i + 512] + common + surround;
     }
+#endif
 }
 
 static void mix22toS (sample_t * samples, sample_t bias)