summaryrefslogtreecommitdiffstats
path: root/liba52
diff options
context:
space:
mode:
authormichael <michael@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-12-18 02:48:14 +0000
committermichael <michael@b3059339-0415-0410-9bf9-f77b7e298cf2>2001-12-18 02:48:14 +0000
commit0a76350996f576ede23b8ca5376658924010a46a (patch)
tree062fc1dddabcc952887df4007dac32384c3896e7 /liba52
parentc7a0e6827fb1e1f886e05789ac9aa312c0817c14 (diff)
downloadmpv-0a76350996f576ede23b8ca5376658924010a46a.tar.bz2
mpv-0a76350996f576ede23b8ca5376658924010a46a.tar.xz
mmx opt
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@3578 b3059339-0415-0410-9bf9-f77b7e298cf2
Diffstat (limited to 'liba52')
-rw-r--r--liba52/resample.c44
-rw-r--r--liba52/resample_c.c44
-rw-r--r--liba52/resample_mmx.c44
3 files changed, 132 insertions, 0 deletions
diff --git a/liba52/resample.c b/liba52/resample.c
index e78f26f304..356220b128 100644
--- a/liba52/resample.c
+++ b/liba52/resample.c
@@ -320,6 +320,49 @@ int a52_resample(float * _f, int16_t * s16)
}
break;
case A52_2F2R | A52_LFE:
+#ifdef HAVE_MMX
+ asm volatile(
+ "movl $-1024, %%esi \n\t"
+ "movq magicF2W, %%mm7 \n\t"
+// "pxor %%mm6, %%mm6 \n\t"
+ "1: \n\t"
+ "movq 1024(%1, %%esi), %%mm0 \n\t"
+ "movq 2048(%1, %%esi), %%mm1 \n\t"
+ "movq 3072(%1, %%esi), %%mm2 \n\t"
+ "movq 4096(%1, %%esi), %%mm3 \n\t"
+ "movq (%1, %%esi), %%mm5 \n\t"
+ "psubd %%mm7, %%mm0 \n\t"
+ "psubd %%mm7, %%mm1 \n\t"
+ "psubd %%mm7, %%mm2 \n\t"
+ "psubd %%mm7, %%mm3 \n\t"
+ "psubd %%mm7, %%mm5 \n\t"
+ "leal (%%esi, %%esi, 2), %%edi \n\t"
+
+ "packssdw %%mm2, %%mm0 \n\t" // CcAa
+ "packssdw %%mm3, %%mm1 \n\t" // DdBb
+ "packssdw %%mm5, %%mm5 \n\t" // FfFf
+ "movq %%mm0, %%mm2 \n\t" // CcAa
+ "punpcklwd %%mm1, %%mm0 \n\t" // BAba
+ "punpckhwd %%mm1, %%mm2 \n\t" // DCdc
+ "pxor %%mm4, %%mm4 \n\t" // 0000
+ "punpcklwd %%mm5, %%mm4 \n\t" // F0f0
+ "movq %%mm0, %%mm1 \n\t" // BAba
+ "movq %%mm4, %%mm3 \n\t" // F0f0
+ "punpckldq %%mm2, %%mm0 \n\t" // dcba
+ "punpckhdq %%mm1, %%mm1 \n\t" // BABA
+ "punpckldq %%mm1, %%mm4 \n\t" // BAf0
+ "punpckhdq %%mm3, %%mm2 \n\t" // F0DC
+
+ "movq %%mm0, (%0, %%edi) \n\t"
+ "movq %%mm4, 8(%0, %%edi) \n\t"
+ "movq %%mm2, 16(%0, %%edi) \n\t"
+ "addl $8, %%esi \n\t"
+ " jnz 1b \n\t"
+ "emms \n\t"
+ :: "r" (s16+1536), "r" (f+256)
+ :"%esi", "%edi", "memory"
+ );
+#else
for (i = 0; i < 256; i++) {
s16[6*i] = convert (f[i+256]);
s16[6*i+1] = convert (f[i+512]);
@@ -328,6 +371,7 @@ int a52_resample(float * _f, int16_t * s16)
s16[6*i+4] = 0;
s16[6*i+5] = convert (f[i]);
}
+#endif
break;
case A52_3F2R | A52_LFE:
#ifdef HAVE_MMX
diff --git a/liba52/resample_c.c b/liba52/resample_c.c
index e78f26f304..356220b128 100644
--- a/liba52/resample_c.c
+++ b/liba52/resample_c.c
@@ -320,6 +320,49 @@ int a52_resample(float * _f, int16_t * s16)
}
break;
case A52_2F2R | A52_LFE:
+#ifdef HAVE_MMX
+ asm volatile(
+ "movl $-1024, %%esi \n\t"
+ "movq magicF2W, %%mm7 \n\t"
+// "pxor %%mm6, %%mm6 \n\t"
+ "1: \n\t"
+ "movq 1024(%1, %%esi), %%mm0 \n\t"
+ "movq 2048(%1, %%esi), %%mm1 \n\t"
+ "movq 3072(%1, %%esi), %%mm2 \n\t"
+ "movq 4096(%1, %%esi), %%mm3 \n\t"
+ "movq (%1, %%esi), %%mm5 \n\t"
+ "psubd %%mm7, %%mm0 \n\t"
+ "psubd %%mm7, %%mm1 \n\t"
+ "psubd %%mm7, %%mm2 \n\t"
+ "psubd %%mm7, %%mm3 \n\t"
+ "psubd %%mm7, %%mm5 \n\t"
+ "leal (%%esi, %%esi, 2), %%edi \n\t"
+
+ "packssdw %%mm2, %%mm0 \n\t" // CcAa
+ "packssdw %%mm3, %%mm1 \n\t" // DdBb
+ "packssdw %%mm5, %%mm5 \n\t" // FfFf
+ "movq %%mm0, %%mm2 \n\t" // CcAa
+ "punpcklwd %%mm1, %%mm0 \n\t" // BAba
+ "punpckhwd %%mm1, %%mm2 \n\t" // DCdc
+ "pxor %%mm4, %%mm4 \n\t" // 0000
+ "punpcklwd %%mm5, %%mm4 \n\t" // F0f0
+ "movq %%mm0, %%mm1 \n\t" // BAba
+ "movq %%mm4, %%mm3 \n\t" // F0f0
+ "punpckldq %%mm2, %%mm0 \n\t" // dcba
+ "punpckhdq %%mm1, %%mm1 \n\t" // BABA
+ "punpckldq %%mm1, %%mm4 \n\t" // BAf0
+ "punpckhdq %%mm3, %%mm2 \n\t" // F0DC
+
+ "movq %%mm0, (%0, %%edi) \n\t"
+ "movq %%mm4, 8(%0, %%edi) \n\t"
+ "movq %%mm2, 16(%0, %%edi) \n\t"
+ "addl $8, %%esi \n\t"
+ " jnz 1b \n\t"
+ "emms \n\t"
+ :: "r" (s16+1536), "r" (f+256)
+ :"%esi", "%edi", "memory"
+ );
+#else
for (i = 0; i < 256; i++) {
s16[6*i] = convert (f[i+256]);
s16[6*i+1] = convert (f[i+512]);
@@ -328,6 +371,7 @@ int a52_resample(float * _f, int16_t * s16)
s16[6*i+4] = 0;
s16[6*i+5] = convert (f[i]);
}
+#endif
break;
case A52_3F2R | A52_LFE:
#ifdef HAVE_MMX
diff --git a/liba52/resample_mmx.c b/liba52/resample_mmx.c
index e78f26f304..356220b128 100644
--- a/liba52/resample_mmx.c
+++ b/liba52/resample_mmx.c
@@ -320,6 +320,49 @@ int a52_resample(float * _f, int16_t * s16)
}
break;
case A52_2F2R | A52_LFE:
+#ifdef HAVE_MMX
+ asm volatile(
+ "movl $-1024, %%esi \n\t"
+ "movq magicF2W, %%mm7 \n\t"
+// "pxor %%mm6, %%mm6 \n\t"
+ "1: \n\t"
+ "movq 1024(%1, %%esi), %%mm0 \n\t"
+ "movq 2048(%1, %%esi), %%mm1 \n\t"
+ "movq 3072(%1, %%esi), %%mm2 \n\t"
+ "movq 4096(%1, %%esi), %%mm3 \n\t"
+ "movq (%1, %%esi), %%mm5 \n\t"
+ "psubd %%mm7, %%mm0 \n\t"
+ "psubd %%mm7, %%mm1 \n\t"
+ "psubd %%mm7, %%mm2 \n\t"
+ "psubd %%mm7, %%mm3 \n\t"
+ "psubd %%mm7, %%mm5 \n\t"
+ "leal (%%esi, %%esi, 2), %%edi \n\t"
+
+ "packssdw %%mm2, %%mm0 \n\t" // CcAa
+ "packssdw %%mm3, %%mm1 \n\t" // DdBb
+ "packssdw %%mm5, %%mm5 \n\t" // FfFf
+ "movq %%mm0, %%mm2 \n\t" // CcAa
+ "punpcklwd %%mm1, %%mm0 \n\t" // BAba
+ "punpckhwd %%mm1, %%mm2 \n\t" // DCdc
+ "pxor %%mm4, %%mm4 \n\t" // 0000
+ "punpcklwd %%mm5, %%mm4 \n\t" // F0f0
+ "movq %%mm0, %%mm1 \n\t" // BAba
+ "movq %%mm4, %%mm3 \n\t" // F0f0
+ "punpckldq %%mm2, %%mm0 \n\t" // dcba
+ "punpckhdq %%mm1, %%mm1 \n\t" // BABA
+ "punpckldq %%mm1, %%mm4 \n\t" // BAf0
+ "punpckhdq %%mm3, %%mm2 \n\t" // F0DC
+
+ "movq %%mm0, (%0, %%edi) \n\t"
+ "movq %%mm4, 8(%0, %%edi) \n\t"
+ "movq %%mm2, 16(%0, %%edi) \n\t"
+ "addl $8, %%esi \n\t"
+ " jnz 1b \n\t"
+ "emms \n\t"
+ :: "r" (s16+1536), "r" (f+256)
+ :"%esi", "%edi", "memory"
+ );
+#else
for (i = 0; i < 256; i++) {
s16[6*i] = convert (f[i+256]);
s16[6*i+1] = convert (f[i+512]);
@@ -328,6 +371,7 @@ int a52_resample(float * _f, int16_t * s16)
s16[6*i+4] = 0;
s16[6*i+5] = convert (f[i]);
}
+#endif
break;
case A52_3F2R | A52_LFE:
#ifdef HAVE_MMX