summaryrefslogtreecommitdiffstats
path: root/liba52
diff options
context:
space:
mode:
authormichael <michael@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-12-25 17:29:12 +0000
committermichael <michael@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-12-25 17:29:12 +0000
commit1c6fe451890076a648cf7f3e309f7828fab27721 (patch)
treeb355113c541c6050e069e8fb8d21c2767c7b354a /liba52
parent26e195ff8927d0946d77e06fb6f60273c68c4b69 (diff)
downloadmpv-1c6fe451890076a648cf7f3e309f7828fab27721.tar.bz2
mpv-1c6fe451890076a648cf7f3e309f7828fab27721.tar.xz
mix21to2 & mix21toS in SSE
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@3740 b3059339-0415-0410-9bf9-f77b7e298cf2
Diffstat (limited to 'liba52')
-rw-r--r--liba52/downmix.c43
1 files changed, 43 insertions, 0 deletions
diff --git a/liba52/downmix.c b/liba52/downmix.c
index 527f38efdd..d6991fffba 100644
--- a/liba52/downmix.c
+++ b/liba52/downmix.c
@@ -454,11 +454,32 @@ static void mix21to2 (sample_t * left, sample_t * right, sample_t bias)
int i;
sample_t common;
+#ifdef HAVE_SSE
+ asm volatile(
+ "movlps %2, %%xmm7 \n\t"
+ "shufps $0x00, %%xmm7, %%xmm7 \n\t"
+ "movl $-1024, %%esi \n\t"
+ "1: \n\t"
+ "movaps 1024(%1, %%esi), %%xmm0 \n\t"
+ "addps %%xmm7, %%xmm0 \n\t" //common
+ "movaps (%0, %%esi), %%xmm1 \n\t"
+ "movaps (%1, %%esi), %%xmm2 \n\t"
+ "addps %%xmm0, %%xmm1 \n\t"
+ "addps %%xmm0, %%xmm2 \n\t"
+ "movaps %%xmm1, (%0, %%esi) \n\t"
+ "movaps %%xmm2, (%1, %%esi) \n\t"
+ "addl $16, %%esi \n\t"
+ " jnz 1b \n\t"
+ :: "r" (left+256), "r" (right+256), "m" (bias)
+ : "%esi"
+ );
+#else
for (i = 0; i < 256; i++) {
common = right[i + 256] + bias;
left[i] += common;
right[i] += common;
}
+#endif
}
static void mix21toS (sample_t * samples, sample_t bias)
@@ -466,11 +487,33 @@ static void mix21toS (sample_t * samples, sample_t bias)
int i;
sample_t 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 2048(%0, %%esi), %%xmm0 \n\t" // surround
+ "movaps (%0, %%esi), %%xmm1 \n\t"
+ "movaps 1024(%0, %%esi), %%xmm2 \n\t"
+ "addps %%xmm7, %%xmm1 \n\t"
+ "addps %%xmm7, %%xmm2 \n\t"
+ "subps %%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++) {
surround = samples[i + 512];
samples[i] += bias - surround;
samples[i + 256] += bias + surround;
}
+#endif
}
static void mix31to2 (sample_t * samples, sample_t bias)