annotate liba52/liba52_changes.diff @ 29601:cd3ae709054f

Disable liba52 SSE imdct for x86_64+PIC (e.g. OSX 64 bit) since it will not compile for that combination.

--- include/a52.h	2006-06-12 15:04:57.000000000 +0200
+++ liba52/a52.h	2006-06-05 02:23:02.000000000 +0200
@@ -59,4 +66,9 @@
 int a52_block (a52_state_t * state);
 void a52_free (a52_state_t * state);

+void* a52_resample_init(uint32_t mm_accel,int flags,int chans);
+extern int (* a52_resample) (float * _f, int16_t * s16);
+
+uint16_t crc16_block(uint8_t *data,uint32_t num_bytes);
+
 #endif /* A52_H */
--- liba52/a52_internal.h	2006-06-12 15:05:07.000000000 +0200
+++ liba52/a52_internal.h	2006-06-05 02:23:02.000000000 +0200
@@ -103,18 +107,34 @@
 #define DELTA_BIT_NONE (2)
 #define DELTA_BIT_RESERVED (3)

+#if ARCH_X86_64
+# define REG_a "rax"
+# define REG_d "rdx"
+# define REG_S "rsi"
+# define REG_D "rdi"
+# define REG_BP "rbp"
+#else
+# define REG_a "eax"
+# define REG_d "edx"
+# define REG_S "esi"
+# define REG_D "edi"
+# define REG_BP "ebp"
+#endif
+
 void a52_bit_allocate (a52_state_t * state, ba_t * ba, int bndstart,
 		       int start, int end, int fastleak, int slowleak,
 		       expbap_t * expbap);

 int a52_downmix_init (int input, int flags, sample_t * level,
 		      sample_t clev, sample_t slev);
+void downmix_accel_init(uint32_t mm_accel);
 int a52_downmix_coeff (sample_t * coeff, int acmod, int output, sample_t level,
 		       sample_t clev, sample_t slev);
-void a52_downmix (sample_t * samples, int acmod, int output, sample_t bias,
+extern void (*a52_downmix) (sample_t * samples, int acmod, int output, sample_t bias,
 		  sample_t clev, sample_t slev);
-void a52_upmix (sample_t * samples, int acmod, int output);
+extern void (*a52_upmix) (sample_t * samples, int acmod, int output);

 void a52_imdct_init (uint32_t mm_accel);
 void a52_imdct_256 (sample_t * data, sample_t * delay, sample_t bias);
-void a52_imdct_512 (sample_t * data, sample_t * delay, sample_t bias);
+extern void (*a52_imdct_512) (sample_t * data, sample_t * delay, sample_t bias);
+void imdct_do_512 (sample_t * data, sample_t * delay, sample_t bias);
--- liba52/bitstream.c	2006-06-12 15:05:07.000000000 +0200
+++ liba52/bitstream.c	2006-06-05 02:23:02.000000000 +0200
@@ -31,6 +35,10 @@

 #define BUFFER_SIZE 4096

+#ifdef ALT_BITSTREAM_READER
+int indx=0;
+#endif
+
 void a52_bitstream_set_ptr (a52_state_t * state, uint8_t * buf)
 {
     int align;
@@ -38,6 +46,9 @@
     align = (long)buf & 3;
     state->buffer_start = (uint32_t *) (buf - align);
     state->bits_left = 0;
+#ifdef ALT_BITSTREAM_READER
+    indx=0;
+#endif
     bitstream_get (state, align * 8);
 }

--- liba52/bitstream.h	2006-06-12 15:05:07.000000000 +0200
+++ liba52/bitstream.h	2006-06-05 02:23:02.000000000 +0200
@@ -21,6 +25,42 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */

+/* code from ffmpeg/libavcodec */
+#if defined(__sparc__) || defined(hpux)
+/*
+ * the alt bitstream reader performs unaligned memory accesses; that doesn't work
+ * on sparc/hpux.  For now, disable ALT_BITSTREAM_READER.
+ */
+#undef	ALT_BITSTREAM_READER
+#else
+// alternative (faster) bitstram reader (reades upto 3 bytes over the end of the input)
+#define ALT_BITSTREAM_READER
+
+/* used to avoid misaligned exceptions on some archs (alpha, ...) */
+#if ARCH_X86 || HAVE_ARMV6
+#    define unaligned32(a) (*(uint32_t*)(a))
+#else
+#    ifdef __GNUC__
+static inline uint32_t unaligned32(const void *v) {
+    struct Unaligned {
+	uint32_t i;
+    } __attribute__((packed));
+
+    return ((const struct Unaligned *) v)->i;
+}
+#    elif defined(__DECC)
+static inline uint32_t unaligned32(const void *v) {
+    return *(const __unaligned uint32_t *) v;
+}
+#    else
+static inline uint32_t unaligned32(const void *v) {
+    return *(const uint32_t *) v;
+}
+#    endif
+#endif //!ARCH_X86
+
+#endif
+
 /* (stolen from the kernel) */
 #if HAVE_BIGENDIAN

@@ -28,7 +68,7 @@

 #else

-#	if 0 && defined (__i386__)
+#	if defined (__i386__)

 #	define swab32(x) __i386_swab32(x)
 	static inline const uint32_t __i386_swab32(uint32_t x)
@@ -39,19 +79,34 @@

 #	else

-#	define swab32(x)\
-((((uint8_t*)&x)[0] << 24) | (((uint8_t*)&x)[1] << 16) |  \
- (((uint8_t*)&x)[2] << 8)  | (((uint8_t*)&x)[3]))
-
+#	define swab32(x) __generic_swab32(x)
+	static inline const uint32_t __generic_swab32(uint32_t x)
+	{
+		return ((((uint8_t*)&x)[0] << 24) | (((uint8_t*)&x)[1] << 16) |
+		 (((uint8_t*)&x)[2] << 8)  | (((uint8_t*)&x)[3]));
+	}
 #	endif
 #endif

+#ifdef ALT_BITSTREAM_READER
+extern int indx;
+#endif
+
 void a52_bitstream_set_ptr (a52_state_t * state, uint8_t * buf);
 uint32_t a52_bitstream_get_bh (a52_state_t * state, uint32_t num_bits);
 int32_t a52_bitstream_get_bh_2 (a52_state_t * state, uint32_t num_bits);

 static inline uint32_t bitstream_get (a52_state_t * state, uint32_t num_bits)
 {
+#ifdef ALT_BITSTREAM_READER
+    uint32_t result= swab32( unaligned32(((uint8_t *)state->buffer_start)+(indx>>3)) );
+
+    result<<= (indx&0x07);
+    result>>= 32 - num_bits;
+    indx+= num_bits;
+
+    return result;
+#else
     uint32_t result;

     if (num_bits < state->bits_left) {
@@ -61,10 +116,29 @@
     }

     return a52_bitstream_get_bh (state, num_bits);
+#endif
+}
+
+static inline void bitstream_skip(a52_state_t * state, int num_bits)
+{
+#ifdef ALT_BITSTREAM_READER
+	indx+= num_bits;
+#else
+	bitstream_get(state, num_bits);
+#endif
 }

 static inline int32_t bitstream_get_2 (a52_state_t * state, uint32_t num_bits)
 {
+#ifdef ALT_BITSTREAM_READER
+    int32_t result= swab32( unaligned32(((uint8_t *)state->buffer_start)+(indx>>3)) );
+
+    result<<= (indx&0x07);
+    result>>= 32 - num_bits;
+    indx+= num_bits;
+
+    return result;
+#else
     int32_t result;

     if (num_bits < state->bits_left) {
@@ -74,4 +148,5 @@
     }

     return a52_bitstream_get_bh_2 (state, num_bits);
+#endif
 }
--- liba52/downmix.c	2006-06-12 15:17:53.000000000 +0200
+++ liba52/downmix.c	2006-06-05 02:23:02.000000000 +0200
@@ -19,18 +23,46 @@
  * You should have received a copy of the GNU General Public License
  * along with this program; if not, write to the Free Software
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ * SSE optimizations from Michael Niedermayer ([e-mail])
  */

 #include "config.h"

 #include <string.h>
 #include <inttypes.h>

 #include "a52.h"
 #include "a52_internal.h"
+#include "mm_accel.h"

 #define CONVERT(acmod,output) (((output) << 3) + (acmod))

+
+void (*a52_downmix)(sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev)= NULL;
+void (*a52_upmix)(sample_t * samples, int acmod, int output)= NULL;
+
+static void downmix_SSE (sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev);
+static void downmix_3dnow (sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev);
+static void downmix_C (sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev);
+static void upmix_MMX (sample_t * samples, int acmod, int output);
+static void upmix_C (sample_t * samples, int acmod, int output);
+
+void downmix_accel_init(uint32_t mm_accel)
+{
+    a52_upmix= upmix_C;
+    a52_downmix= downmix_C;
+#if ARCH_X86 || ARCH_X86_64
+    if(mm_accel & MM_ACCEL_X86_MMX) a52_upmix= upmix_MMX;
+    if(mm_accel & MM_ACCEL_X86_SSE) a52_downmix= downmix_SSE;
+    if(mm_accel & MM_ACCEL_X86_3DNOW) a52_downmix= downmix_3dnow;
+#endif
+}
+
 int a52_downmix_init (int input, int flags, sample_t * level,
 		      sample_t clev, sample_t slev)
 {
@@ -447,7 +479,7 @@
 	samples[i] = 0;
 }

-void a52_downmix (sample_t * samples, int acmod, int output, sample_t bias,
+void downmix_C (sample_t * samples, int acmod, int output, sample_t bias,
 		  sample_t clev, sample_t slev)
 {
     switch (CONVERT (acmod, output & A52_CHANNEL_MASK)) {
@@ -559,7 +591,7 @@
 	break;

     case CONVERT (A52_3F2R, A52_2F1R):
-	mix3to2 (samples, bias);
+	mix3to2 (samples, bias); //FIXME possible bug? (output doesnt seem to be used)
 	move2to1 (samples + 768, samples + 512, bias);
 	break;

@@ -583,12 +615,12 @@
 	break;

     case CONVERT (A52_3F1R, A52_3F2R):
-	memcpy (samples + 1027, samples + 768, 256 * sizeof (sample_t));
+	memcpy (samples + 1024, samples + 768, 256 * sizeof (sample_t));
 	break;
     }
 }

-void a52_upmix (sample_t * samples, int acmod, int output)
+void upmix_C (sample_t * samples, int acmod, int output)
 {
     switch (CONVERT (acmod, output & A52_CHANNEL_MASK)) {

@@ -653,3 +685,1104 @@
 	goto mix_31to21;
     }
 }
+
+#if ARCH_X86 || ARCH_X86_64
+static void mix2to1_SSE (sample_t * dest, sample_t * src, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %2, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps (%0, %%"REG_S"), %%xmm0	\n\t"
+	"movaps 16(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps (%1, %%"REG_S"), %%xmm0	\n\t"
+	"addps 16(%1, %%"REG_S"), %%xmm1\n\t"
+	"addps %%xmm7, %%xmm0		\n\t"
+	"addps %%xmm7, %%xmm1		\n\t"
+	"movaps %%xmm0, (%1, %%"REG_S")	\n\t"
+	"movaps %%xmm1, 16(%1, %%"REG_S")\n\t"
+	"add $32, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (src+256), "r" (dest+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix3to1_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps (%0, %%"REG_S"), %%xmm0	\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps 2048(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps %%xmm7, %%xmm1		\n\t"
+	"addps %%xmm1, %%xmm0		\n\t"
+	"movaps %%xmm0, (%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix4to1_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps (%0, %%"REG_S"), %%xmm0	\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps 2048(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps 3072(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps %%xmm7, %%xmm0		\n\t"
+	"addps %%xmm1, %%xmm0		\n\t"
+	"movaps %%xmm0, (%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix5to1_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps (%0, %%"REG_S"), %%xmm0	\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps 2048(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps 3072(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps %%xmm7, %%xmm0		\n\t"
+	"addps 4096(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps %%xmm1, %%xmm0		\n\t"
+	"movaps %%xmm0, (%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix3to2_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps %%xmm7, %%xmm0		\n\t" //common
+	"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+	"movaps 2048(%0, %%"REG_S"), %%xmm2\n\t"
+	"addps %%xmm0, %%xmm1		\n\t"
+	"addps %%xmm0, %%xmm2		\n\t"
+	"movaps %%xmm1, (%0, %%"REG_S")	\n\t"
+	"movaps %%xmm2, 1024(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix21to2_SSE (sample_t * left, sample_t * right, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %2, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 1024(%1, %%"REG_S"), %%xmm0\n\t"
+		"addps %%xmm7, %%xmm0		\n\t" //common
+		"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+		"movaps (%1, %%"REG_S"), %%xmm2	\n\t"
+		"addps %%xmm0, %%xmm1		\n\t"
+		"addps %%xmm0, %%xmm2		\n\t"
+		"movaps %%xmm1, (%0, %%"REG_S")	\n\t"
+		"movaps %%xmm2, (%1, %%"REG_S")	\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (left+256), "r" (right+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix21toS_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 2048(%0, %%"REG_S"), %%xmm0\n\t"  // surround
+		"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+		"movaps 1024(%0, %%"REG_S"), %%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, %%"REG_S")	\n\t"
+		"movaps %%xmm2, 1024(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix31to2_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+		"addps 3072(%0, %%"REG_S"), %%xmm0\n\t"
+		"addps %%xmm7, %%xmm0		\n\t" // common
+		"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+		"movaps 2048(%0, %%"REG_S"), %%xmm2\n\t"
+		"addps %%xmm0, %%xmm1		\n\t"
+		"addps %%xmm0, %%xmm2		\n\t"
+		"movaps %%xmm1, (%0, %%"REG_S")	\n\t"
+		"movaps %%xmm2, 1024(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix31toS_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+		"movaps 3072(%0, %%"REG_S"), %%xmm3\n\t" // surround
+		"addps %%xmm7, %%xmm0		\n\t" // common
+		"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+		"movaps 2048(%0, %%"REG_S"), %%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, %%"REG_S")	\n\t"
+		"movaps %%xmm2, 1024(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix22toS_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %1, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 2048(%0, %%"REG_S"), %%xmm0\n\t"
+		"addps 3072(%0, %%"REG_S"), %%xmm0\n\t" // surround
+		"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+		"movaps 1024(%0, %%"REG_S"), %%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, %%"REG_S")	\n\t"
+		"movaps %%xmm2, 1024(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix32to2_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps %%xmm7, %%xmm0		\n\t" // common
+	"movaps %%xmm0, %%xmm1		\n\t" // common
+	"addps (%0, %%"REG_S"), %%xmm0	\n\t"
+	"addps 2048(%0, %%"REG_S"), %%xmm1\n\t"
+	"addps 3072(%0, %%"REG_S"), %%xmm0\n\t"
+	"addps 4096(%0, %%"REG_S"), %%xmm1\n\t"
+	"movaps %%xmm0, (%0, %%"REG_S")	\n\t"
+	"movaps %%xmm1, 1024(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix32toS_SSE (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movlps %1, %%xmm7		\n\t"
+	"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:				\n\t"
+	"movaps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+	"movaps 3072(%0, %%"REG_S"), %%xmm2\n\t"
+	"addps %%xmm7, %%xmm0		\n\t" // common
+	"addps 4096(%0, %%"REG_S"), %%xmm2\n\t" // surround
+	"movaps (%0, %%"REG_S"), %%xmm1	\n\t"
+	"movaps 2048(%0, %%"REG_S"), %%xmm3\n\t"
+	"subps %%xmm2, %%xmm1		\n\t"
+	"addps %%xmm2, %%xmm3		\n\t"
+	"addps %%xmm0, %%xmm1		\n\t"
+	"addps %%xmm0, %%xmm3		\n\t"
+	"movaps %%xmm1, (%0, %%"REG_S")	\n\t"
+	"movaps %%xmm3, 1024(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void move2to1_SSE (sample_t * src, sample_t * dest, sample_t bias)
+{
+	__asm__ volatile(
+		"movlps %2, %%xmm7		\n\t"
+		"shufps $0x00, %%xmm7, %%xmm7	\n\t"
+		"mov $-1024, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps (%0, %%"REG_S"), %%xmm0	\n\t"
+		"movaps 16(%0, %%"REG_S"), %%xmm1\n\t"
+		"addps 1024(%0, %%"REG_S"), %%xmm0\n\t"
+		"addps 1040(%0, %%"REG_S"), %%xmm1\n\t"
+		"addps %%xmm7, %%xmm0		\n\t"
+		"addps %%xmm7, %%xmm1		\n\t"
+		"movaps %%xmm0, (%1, %%"REG_S")	\n\t"
+		"movaps %%xmm1, 16(%1, %%"REG_S")\n\t"
+		"add $32, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (src+256), "r" (dest+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void zero_MMX(sample_t * samples)
+{
+	__asm__ volatile(
+		"mov $-1024, %%"REG_S"		\n\t"
+		"pxor %%mm0, %%mm0		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movq %%mm0, (%0, %%"REG_S")	\n\t"
+		"movq %%mm0, 8(%0, %%"REG_S")	\n\t"
+		"movq %%mm0, 16(%0, %%"REG_S")	\n\t"
+		"movq %%mm0, 24(%0, %%"REG_S")	\n\t"
+		"add $32, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+		"emms"
+	:: "r" (samples+256)
+	: "%"REG_S
+	);
+}
+
+static void downmix_SSE (sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev)
+{
+    switch (CONVERT (acmod, output & A52_CHANNEL_MASK)) {
+
+    case CONVERT (A52_CHANNEL, A52_CHANNEL2):
+	memcpy (samples, samples + 256, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_CHANNEL, A52_MONO):
+    case CONVERT (A52_STEREO, A52_MONO):
+    mix_2to1_SSE:
+	mix2to1_SSE (samples, samples + 256, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_MONO):
+	if (slev == 0)
+	    goto mix_2to1_SSE;
+    case CONVERT (A52_3F, A52_MONO):
+    mix_3to1_SSE:
+	mix3to1_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_MONO):
+	if (slev == 0)
+	    goto mix_3to1_SSE;
+    case CONVERT (A52_2F2R, A52_MONO):
+	if (slev == 0)
+	    goto mix_2to1_SSE;
+	mix4to1_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_MONO):
+	if (slev == 0)
+	    goto mix_3to1_SSE;
+	mix5to1_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_MONO, A52_DOLBY):
+	memcpy (samples + 256, samples, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F, A52_STEREO):
+    case CONVERT (A52_3F, A52_DOLBY):
+    mix_3to2_SSE:
+	mix3to2_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_STEREO):
+	if (slev == 0)
+	    break;
+	mix21to2_SSE (samples, samples + 256, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_DOLBY):
+	mix21toS_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_STEREO):
+	if (slev == 0)
+	    goto mix_3to2_SSE;
+	mix31to2_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_DOLBY):
+	mix31toS_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_2F2R, A52_STEREO):
+	if (slev == 0)
+	    break;
+	mix2to1_SSE (samples, samples + 512, bias);
+	mix2to1_SSE (samples + 256, samples + 768, bias);
+	break;
+
+    case CONVERT (A52_2F2R, A52_DOLBY):
+	mix22toS_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_STEREO):
+	if (slev == 0)
+	    goto mix_3to2_SSE;
+	mix32to2_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_DOLBY):
+	mix32toS_SSE (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_3F):
+	if (slev == 0)
+	    break;
+	mix21to2_SSE (samples, samples + 512, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F):
+	if (slev == 0)
+	    break;
+	mix2to1_SSE (samples, samples + 768, bias);
+	mix2to1_SSE (samples + 512, samples + 1024, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_2F1R):
+	mix3to2_SSE (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_2F2R, A52_2F1R):
+	mix2to1_SSE (samples + 512, samples + 768, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_2F1R):
+	mix3to2_SSE (samples, bias); //FIXME possible bug? (output doesnt seem to be used)
+	move2to1_SSE (samples + 768, samples + 512, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F1R):
+	mix2to1_SSE (samples + 768, samples + 1024, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_2F2R):
+	memcpy (samples + 768, samples + 512, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F1R, A52_2F2R):
+	mix3to2_SSE (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F2R, A52_2F2R):
+	mix3to2_SSE (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	memcpy (samples + 768, samples + 1024, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F1R, A52_3F2R):
+	memcpy (samples + 1024, samples + 768, 256 * sizeof (sample_t));
+	break;
+    }
+}
+
+static void upmix_MMX (sample_t * samples, int acmod, int output)
+{
+    switch (CONVERT (acmod, output & A52_CHANNEL_MASK)) {
+
+    case CONVERT (A52_CHANNEL, A52_CHANNEL2):
+	memcpy (samples + 256, samples, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F2R, A52_MONO):
+	zero_MMX (samples + 1024);
+    case CONVERT (A52_3F1R, A52_MONO):
+    case CONVERT (A52_2F2R, A52_MONO):
+	zero_MMX (samples + 768);
+    case CONVERT (A52_3F, A52_MONO):
+    case CONVERT (A52_2F1R, A52_MONO):
+	zero_MMX (samples + 512);
+    case CONVERT (A52_CHANNEL, A52_MONO):
+    case CONVERT (A52_STEREO, A52_MONO):
+	zero_MMX (samples + 256);
+	break;
+
+    case CONVERT (A52_3F2R, A52_STEREO):
+    case CONVERT (A52_3F2R, A52_DOLBY):
+	zero_MMX (samples + 1024);
+    case CONVERT (A52_3F1R, A52_STEREO):
+    case CONVERT (A52_3F1R, A52_DOLBY):
+	zero_MMX (samples + 768);
+    case CONVERT (A52_3F, A52_STEREO):
+    case CONVERT (A52_3F, A52_DOLBY):
+    mix_3to2_MMX:
+	memcpy (samples + 512, samples + 256, 256 * sizeof (sample_t));
+	zero_MMX (samples + 256);
+	break;
+
+    case CONVERT (A52_2F2R, A52_STEREO):
+    case CONVERT (A52_2F2R, A52_DOLBY):
+	zero_MMX (samples + 768);
+    case CONVERT (A52_2F1R, A52_STEREO):
+    case CONVERT (A52_2F1R, A52_DOLBY):
+	zero_MMX (samples + 512);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F):
+	zero_MMX (samples + 1024);
+    case CONVERT (A52_3F1R, A52_3F):
+    case CONVERT (A52_2F2R, A52_2F1R):
+	zero_MMX (samples + 768);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F1R):
+	zero_MMX (samples + 1024);
+	break;
+
+    case CONVERT (A52_3F2R, A52_2F1R):
+	zero_MMX (samples + 1024);
+    case CONVERT (A52_3F1R, A52_2F1R):
+    mix_31to21_MMX:
+	memcpy (samples + 768, samples + 512, 256 * sizeof (sample_t));
+	goto mix_3to2_MMX;
+
+    case CONVERT (A52_3F2R, A52_2F2R):
+	memcpy (samples + 1024, samples + 768, 256 * sizeof (sample_t));
+	goto mix_31to21_MMX;
+    }
+}
+
+static void mix2to1_3dnow (sample_t * dest, sample_t * src, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %2, %%mm7	\n\t"
+	"punpckldq %2, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq  (%0, %%"REG_S"), %%mm0	\n\t"
+	"movq  8(%0, %%"REG_S"), %%mm1	\n\t"
+	"movq  16(%0, %%"REG_S"), %%mm2	\n\t"
+	"movq  24(%0, %%"REG_S"), %%mm3	\n\t"
+	"pfadd (%1, %%"REG_S"), %%mm0	\n\t"
+	"pfadd 8(%1, %%"REG_S"), %%mm1	\n\t"
+	"pfadd 16(%1, %%"REG_S"), %%mm2	\n\t"
+	"pfadd 24(%1, %%"REG_S"), %%mm3	\n\t"
+	"pfadd %%mm7, %%mm0		\n\t"
+	"pfadd %%mm7, %%mm1		\n\t"
+	"pfadd %%mm7, %%mm2		\n\t"
+	"pfadd %%mm7, %%mm3		\n\t"
+	"movq  %%mm0, (%1, %%"REG_S")	\n\t"
+	"movq  %%mm1, 8(%1, %%"REG_S")	\n\t"
+	"movq  %%mm2, 16(%1, %%"REG_S")	\n\t"
+	"movq  %%mm3, 24(%1, %%"REG_S")	\n\t"
+	"add $32, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (src+256), "r" (dest+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix3to1_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %1, %%mm7	\n\t"
+	"punpckldq %1, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq  (%0, %%"REG_S"), %%mm0	\n\t"
+	"movq  8(%0, %%"REG_S"), %%mm1	\n\t"
+	"movq  1024(%0, %%"REG_S"), %%mm2\n\t"
+	"movq  1032(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd 2048(%0, %%"REG_S"), %%mm0\n\t"
+	"pfadd 2056(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd %%mm7, %%mm0		\n\t"
+	"pfadd %%mm7, %%mm1		\n\t"
+	"pfadd %%mm2, %%mm0		\n\t"
+	"pfadd %%mm3, %%mm1		\n\t"
+	"movq  %%mm0, (%0, %%"REG_S")	\n\t"
+	"movq  %%mm1, 8(%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix4to1_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %1, %%mm7	\n\t"
+	"punpckldq %1, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq  (%0, %%"REG_S"), %%mm0	\n\t"
+	"movq  8(%0, %%"REG_S"), %%mm1	\n\t"
+	"movq  1024(%0, %%"REG_S"), %%mm2\n\t"
+	"movq  1032(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd 2048(%0, %%"REG_S"), %%mm0\n\t"
+	"pfadd 2056(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd 3072(%0, %%"REG_S"), %%mm2\n\t"
+	"pfadd 3080(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd %%mm7, %%mm0		\n\t"
+	"pfadd %%mm7, %%mm1		\n\t"
+	"pfadd %%mm2, %%mm0		\n\t"
+	"pfadd %%mm3, %%mm1		\n\t"
+	"movq  %%mm0, (%0, %%"REG_S")	\n\t"
+	"movq  %%mm1, 8(%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix5to1_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %1, %%mm7	\n\t"
+	"punpckldq %1, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq  (%0, %%"REG_S"), %%mm0	\n\t"
+	"movq  8(%0, %%"REG_S"), %%mm1	\n\t"
+	"movq  1024(%0, %%"REG_S"), %%mm2\n\t"
+	"movq  1032(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd 2048(%0, %%"REG_S"), %%mm0\n\t"
+	"pfadd 2056(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd 3072(%0, %%"REG_S"), %%mm2\n\t"
+	"pfadd 3080(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd %%mm7, %%mm0		\n\t"
+	"pfadd %%mm7, %%mm1		\n\t"
+	"pfadd 4096(%0, %%"REG_S"), %%mm2\n\t"
+	"pfadd 4104(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd %%mm2, %%mm0		\n\t"
+	"pfadd %%mm3, %%mm1		\n\t"
+	"movq  %%mm0, (%0, %%"REG_S")	\n\t"
+	"movq  %%mm1, 8(%0, %%"REG_S")	\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix3to2_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %1, %%mm7	\n\t"
+	"punpckldq %1, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq   1024(%0, %%"REG_S"), %%mm0\n\t"
+	"movq   1032(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd  %%mm7, %%mm0		\n\t" //common
+	"pfadd  %%mm7, %%mm1		\n\t" //common
+	"movq   (%0, %%"REG_S"), %%mm2	\n\t"
+	"movq   8(%0, %%"REG_S"), %%mm3	\n\t"
+	"movq   2048(%0, %%"REG_S"), %%mm4\n\t"
+	"movq   2056(%0, %%"REG_S"), %%mm5\n\t"
+	"pfadd  %%mm0, %%mm2		\n\t"
+	"pfadd  %%mm1, %%mm3		\n\t"
+	"pfadd  %%mm0, %%mm4		\n\t"
+	"pfadd  %%mm1, %%mm5		\n\t"
+	"movq   %%mm2, (%0, %%"REG_S")	\n\t"
+	"movq   %%mm3, 8(%0, %%"REG_S")	\n\t"
+	"movq   %%mm4, 1024(%0, %%"REG_S")\n\t"
+	"movq   %%mm5, 1032(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix21to2_3dnow (sample_t * left, sample_t * right, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %2, %%mm7	\n\t"
+		"punpckldq %2, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq  1024(%1, %%"REG_S"), %%mm0\n\t"
+		"movq  1032(%1, %%"REG_S"), %%mm1\n\t"
+		"pfadd %%mm7, %%mm0		\n\t" //common
+		"pfadd %%mm7, %%mm1		\n\t" //common
+		"movq  (%0, %%"REG_S"), %%mm2	\n\t"
+		"movq  8(%0, %%"REG_S"), %%mm3	\n\t"
+		"movq  (%1, %%"REG_S"), %%mm4	\n\t"
+		"movq  8(%1, %%"REG_S"), %%mm5	\n\t"
+		"pfadd %%mm0, %%mm2		\n\t"
+		"pfadd %%mm1, %%mm3		\n\t"
+		"pfadd %%mm0, %%mm4		\n\t"
+		"pfadd %%mm1, %%mm5		\n\t"
+		"movq  %%mm2, (%0, %%"REG_S")	\n\t"
+		"movq  %%mm3, 8(%0, %%"REG_S")	\n\t"
+		"movq  %%mm4, (%1, %%"REG_S")	\n\t"
+		"movq  %%mm5, 8(%1, %%"REG_S")	\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (left+256), "r" (right+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix21toS_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %1, %%mm7	\n\t"
+		"punpckldq %1, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq  2048(%0, %%"REG_S"), %%mm0\n\t"  // surround
+		"movq  2056(%0, %%"REG_S"), %%mm1\n\t"  // surround
+		"movq  (%0, %%"REG_S"), %%mm2	\n\t"
+		"movq  8(%0, %%"REG_S"), %%mm3	\n\t"
+		"movq  1024(%0, %%"REG_S"), %%mm4\n\t"
+		"movq  1032(%0, %%"REG_S"), %%mm5\n\t"
+		"pfadd %%mm7, %%mm2		\n\t"
+		"pfadd %%mm7, %%mm3		\n\t"
+		"pfadd %%mm7, %%mm4		\n\t"
+		"pfadd %%mm7, %%mm5		\n\t"
+		"pfsub %%mm0, %%mm2		\n\t"
+		"pfsub %%mm1, %%mm3		\n\t"
+		"pfadd %%mm0, %%mm4		\n\t"
+		"pfadd %%mm1, %%mm5		\n\t"
+		"movq  %%mm2, (%0, %%"REG_S")	\n\t"
+		"movq  %%mm3, 8(%0, %%"REG_S")	\n\t"
+		"movq  %%mm4, 1024(%0, %%"REG_S")\n\t"
+		"movq  %%mm5, 1032(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix31to2_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %1, %%mm7	\n\t"
+		"punpckldq %1, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq  1024(%0, %%"REG_S"), %%mm0\n\t"
+		"movq  1032(%0, %%"REG_S"), %%mm1\n\t"
+		"pfadd 3072(%0, %%"REG_S"), %%mm0\n\t"
+		"pfadd 3080(%0, %%"REG_S"), %%mm1\n\t"
+		"pfadd %%mm7, %%mm0		\n\t" // common
+		"pfadd %%mm7, %%mm1		\n\t" // common
+		"movq  (%0, %%"REG_S"), %%mm2	\n\t"
+		"movq  8(%0, %%"REG_S"), %%mm3	\n\t"
+		"movq  2048(%0, %%"REG_S"), %%mm4\n\t"
+		"movq  2056(%0, %%"REG_S"), %%mm5\n\t"
+		"pfadd %%mm0, %%mm2		\n\t"
+		"pfadd %%mm1, %%mm3		\n\t"
+		"pfadd %%mm0, %%mm4		\n\t"
+		"pfadd %%mm1, %%mm5		\n\t"
+		"movq  %%mm2, (%0, %%"REG_S")	\n\t"
+		"movq  %%mm3, 8(%0, %%"REG_S")	\n\t"
+		"movq  %%mm4, 1024(%0, %%"REG_S")\n\t"
+		"movq  %%mm5, 1032(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix31toS_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %1, %%mm7	\n\t"
+		"punpckldq %1, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq   1024(%0, %%"REG_S"), %%mm0\n\t"
+		"movq   1032(%0, %%"REG_S"), %%mm1\n\t"
+		"pfadd  %%mm7, %%mm0		\n\t" // common
+		"pfadd  %%mm7, %%mm1		\n\t" // common
+		"movq   (%0, %%"REG_S"), %%mm2	\n\t"
+		"movq   8(%0, %%"REG_S"), %%mm3	\n\t"
+		"movq   2048(%0, %%"REG_S"), %%mm4\n\t"
+		"movq   2056(%0, %%"REG_S"), %%mm5\n\t"
+		"pfadd  %%mm0, %%mm2		\n\t"
+		"pfadd  %%mm1, %%mm3		\n\t"
+		"pfadd  %%mm0, %%mm4		\n\t"
+		"pfadd  %%mm1, %%mm5		\n\t"
+		"movq   3072(%0, %%"REG_S"), %%mm0\n\t" // surround
+		"movq   3080(%0, %%"REG_S"), %%mm1\n\t" // surround
+		"pfsub  %%mm0, %%mm2		\n\t"
+		"pfsub  %%mm1, %%mm3		\n\t"
+		"pfadd  %%mm0, %%mm4		\n\t"
+		"pfadd  %%mm1, %%mm5		\n\t"
+		"movq   %%mm2, (%0, %%"REG_S")	\n\t"
+		"movq   %%mm3, 8(%0, %%"REG_S")	\n\t"
+		"movq   %%mm4, 1024(%0, %%"REG_S")\n\t"
+		"movq   %%mm5, 1032(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix22toS_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %1, %%mm7	\n\t"
+		"punpckldq %1, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq  2048(%0, %%"REG_S"), %%mm0\n\t"
+		"movq  2056(%0, %%"REG_S"), %%mm1\n\t"
+		"pfadd 3072(%0, %%"REG_S"), %%mm0\n\t" // surround
+		"pfadd 3080(%0, %%"REG_S"), %%mm1\n\t" // surround
+		"movq  (%0, %%"REG_S"), %%mm2	\n\t"
+		"movq  8(%0, %%"REG_S"), %%mm3	\n\t"
+		"movq  1024(%0, %%"REG_S"), %%mm4\n\t"
+		"movq  1032(%0, %%"REG_S"), %%mm5\n\t"
+		"pfadd %%mm7, %%mm2		\n\t"
+		"pfadd %%mm7, %%mm3		\n\t"
+		"pfadd %%mm7, %%mm4		\n\t"
+		"pfadd %%mm7, %%mm5		\n\t"
+		"pfsub %%mm0, %%mm2		\n\t"
+		"pfsub %%mm1, %%mm3		\n\t"
+		"pfadd %%mm0, %%mm4		\n\t"
+		"pfadd %%mm1, %%mm5		\n\t"
+		"movq  %%mm2, (%0, %%"REG_S")	\n\t"
+		"movq  %%mm3, 8(%0, %%"REG_S")	\n\t"
+		"movq  %%mm4, 1024(%0, %%"REG_S")\n\t"
+		"movq  %%mm5, 1032(%0, %%"REG_S")\n\t"
+		"add $16, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void mix32to2_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"movd  %1, %%mm7	\n\t"
+	"punpckldq %1, %%mm7	\n\t"
+	"mov $-1024, %%"REG_S"	\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movq   1024(%0, %%"REG_S"), %%mm0\n\t"
+	"movq   1032(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd  %%mm7, %%mm0		\n\t" // common
+	"pfadd  %%mm7, %%mm1		\n\t" // common
+	"movq   %%mm0, %%mm2		\n\t" // common
+	"movq   %%mm1, %%mm3		\n\t" // common
+	"pfadd  (%0, %%"REG_S"), %%mm0	\n\t"
+	"pfadd  8(%0, %%"REG_S"), %%mm1	\n\t"
+	"pfadd  2048(%0, %%"REG_S"), %%mm2\n\t"
+	"pfadd  2056(%0, %%"REG_S"), %%mm3\n\t"
+	"pfadd  3072(%0, %%"REG_S"), %%mm0\n\t"
+	"pfadd  3080(%0, %%"REG_S"), %%mm1\n\t"
+	"pfadd  4096(%0, %%"REG_S"), %%mm2\n\t"
+	"pfadd  4104(%0, %%"REG_S"), %%mm3\n\t"
+	"movq   %%mm0, (%0, %%"REG_S")	\n\t"
+	"movq   %%mm1, 8(%0, %%"REG_S")	\n\t"
+	"movq   %%mm2, 1024(%0, %%"REG_S")\n\t"
+	"movq   %%mm3, 1032(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+/* todo: should be optimized better */
+static void mix32toS_3dnow (sample_t * samples, sample_t bias)
+{
+	__asm__ volatile(
+	"mov $-1024, %%"REG_S"		\n\t"
+	ASMALIGN(4)
+	"1:			\n\t"
+	"movd  %1, %%mm7		\n\t"
+	"punpckldq %1, %%mm7		\n\t"
+	"movq  1024(%0, %%"REG_S"), %%mm0\n\t"
+	"movq  1032(%0, %%"REG_S"), %%mm1\n\t"
+	"movq  3072(%0, %%"REG_S"), %%mm4\n\t"
+	"movq  3080(%0, %%"REG_S"), %%mm5\n\t"
+	"pfadd %%mm7, %%mm0		\n\t" // common
+	"pfadd %%mm7, %%mm1		\n\t" // common
+	"pfadd 4096(%0, %%"REG_S"), %%mm4\n\t" // surround
+	"pfadd 4104(%0, %%"REG_S"), %%mm5\n\t" // surround
+	"movq  (%0, %%"REG_S"), %%mm2	\n\t"
+	"movq  8(%0, %%"REG_S"), %%mm3	\n\t"
+	"movq  2048(%0, %%"REG_S"), %%mm6\n\t"
+	"movq  2056(%0, %%"REG_S"), %%mm7\n\t"
+	"pfsub %%mm4, %%mm2		\n\t"
+	"pfsub %%mm5, %%mm3		\n\t"
+	"pfadd %%mm4, %%mm6		\n\t"
+	"pfadd %%mm5, %%mm7		\n\t"
+	"pfadd %%mm0, %%mm2		\n\t"
+	"pfadd %%mm1, %%mm3		\n\t"
+	"pfadd %%mm0, %%mm6		\n\t"
+	"pfadd %%mm1, %%mm7		\n\t"
+	"movq  %%mm2, (%0, %%"REG_S")	\n\t"
+	"movq  %%mm3, 8(%0, %%"REG_S")	\n\t"
+	"movq  %%mm6, 1024(%0, %%"REG_S")\n\t"
+	"movq  %%mm7, 1032(%0, %%"REG_S")\n\t"
+	"add $16, %%"REG_S"		\n\t"
+	" jnz 1b			\n\t"
+	:: "r" (samples+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void move2to1_3dnow (sample_t * src, sample_t * dest, sample_t bias)
+{
+	__asm__ volatile(
+		"movd  %2, %%mm7	\n\t"
+		"punpckldq %2, %%mm7	\n\t"
+		"mov $-1024, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movq  (%0, %%"REG_S"), %%mm0	\n\t"
+		"movq  8(%0, %%"REG_S"), %%mm1	\n\t"
+		"movq  16(%0, %%"REG_S"), %%mm2	\n\t"
+		"movq  24(%0, %%"REG_S"), %%mm3	\n\t"
+		"pfadd 1024(%0, %%"REG_S"), %%mm0\n\t"
+		"pfadd 1032(%0, %%"REG_S"), %%mm1\n\t"
+		"pfadd 1040(%0, %%"REG_S"), %%mm2\n\t"
+		"pfadd 1048(%0, %%"REG_S"), %%mm3\n\t"
+		"pfadd %%mm7, %%mm0		\n\t"
+		"pfadd %%mm7, %%mm1		\n\t"
+		"pfadd %%mm7, %%mm2		\n\t"
+		"pfadd %%mm7, %%mm3		\n\t"
+		"movq  %%mm0, (%1, %%"REG_S")	\n\t"
+		"movq  %%mm1, 8(%1, %%"REG_S")	\n\t"
+		"movq  %%mm2, 16(%1, %%"REG_S")	\n\t"
+		"movq  %%mm3, 24(%1, %%"REG_S")	\n\t"
+		"add $32, %%"REG_S"		\n\t"
+		" jnz 1b			\n\t"
+	:: "r" (src+256), "r" (dest+256), "m" (bias)
+	: "%"REG_S
+	);
+}
+
+static void downmix_3dnow (sample_t * samples, int acmod, int output, sample_t bias,
+	      sample_t clev, sample_t slev)
+{
+    switch (CONVERT (acmod, output & A52_CHANNEL_MASK)) {
+
+    case CONVERT (A52_CHANNEL, A52_CHANNEL2):
+	memcpy (samples, samples + 256, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_CHANNEL, A52_MONO):
+    case CONVERT (A52_STEREO, A52_MONO):
+    mix_2to1_3dnow:
+	mix2to1_3dnow (samples, samples + 256, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_MONO):
+	if (slev == 0)
+	    goto mix_2to1_3dnow;
+    case CONVERT (A52_3F, A52_MONO):
+    mix_3to1_3dnow:
+	mix3to1_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_MONO):
+	if (slev == 0)
+	    goto mix_3to1_3dnow;
+    case CONVERT (A52_2F2R, A52_MONO):
+	if (slev == 0)
+	    goto mix_2to1_3dnow;
+	mix4to1_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_MONO):
+	if (slev == 0)
+	    goto mix_3to1_3dnow;
+	mix5to1_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_MONO, A52_DOLBY):
+	memcpy (samples + 256, samples, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F, A52_STEREO):
+    case CONVERT (A52_3F, A52_DOLBY):
+    mix_3to2_3dnow:
+	mix3to2_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_STEREO):
+	if (slev == 0)
+	    break;
+	mix21to2_3dnow (samples, samples + 256, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_DOLBY):
+	mix21toS_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_STEREO):
+	if (slev == 0)
+	    goto mix_3to2_3dnow;
+	mix31to2_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_DOLBY):
+	mix31toS_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_2F2R, A52_STEREO):
+	if (slev == 0)
+	    break;
+	mix2to1_3dnow (samples, samples + 512, bias);
+	mix2to1_3dnow (samples + 256, samples + 768, bias);
+	break;
+
+    case CONVERT (A52_2F2R, A52_DOLBY):
+	mix22toS_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_STEREO):
+	if (slev == 0)
+	    goto mix_3to2_3dnow;
+	mix32to2_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_DOLBY):
+	mix32toS_3dnow (samples, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_3F):
+	if (slev == 0)
+	    break;
+	mix21to2_3dnow (samples, samples + 512, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F):
+	if (slev == 0)
+	    break;
+	mix2to1_3dnow (samples, samples + 768, bias);
+	mix2to1_3dnow (samples + 512, samples + 1024, bias);
+	break;
+
+    case CONVERT (A52_3F1R, A52_2F1R):
+	mix3to2_3dnow (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_2F2R, A52_2F1R):
+	mix2to1_3dnow (samples + 512, samples + 768, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_2F1R):
+	mix3to2_3dnow (samples, bias); //FIXME possible bug? (output doesnt seem to be used)
+	move2to1_3dnow (samples + 768, samples + 512, bias);
+	break;
+
+    case CONVERT (A52_3F2R, A52_3F1R):
+	mix2to1_3dnow (samples + 768, samples + 1024, bias);
+	break;
+
+    case CONVERT (A52_2F1R, A52_2F2R):
+	memcpy (samples + 768, samples + 512, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F1R, A52_2F2R):
+	mix3to2_3dnow (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F2R, A52_2F2R):
+	mix3to2_3dnow (samples, bias);
+	memcpy (samples + 512, samples + 768, 256 * sizeof (sample_t));
+	memcpy (samples + 768, samples + 1024, 256 * sizeof (sample_t));
+	break;
+
+    case CONVERT (A52_3F1R, A52_3F2R):
+	memcpy (samples + 1024, samples + 768, 256 * sizeof (sample_t));
+	break;
+    }
+    __asm__ volatile("femms":::"memory");
+}
+
+#endif // ARCH_X86 || ARCH_X86_64
--- liba52/imdct.c	2008-02-19 00:18:33.000000000 +0100
+++ liba52/imdct.c	2008-02-19 00:16:40.000000000 +0100
@@ -22,6 +26,11 @@
  * You should have received a copy of the GNU General Public License
  * along with this program; if not, write to the Free Software
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ * SSE optimizations from Michael Niedermayer ([e-mail])
+ * 3DNOW optimizations from Nick Kurshev <[e-mail]>
+ *   michael did port them from libac3 (untested, perhaps totally broken)
+ * AltiVec optimizations from Romain Dolbeau ([e-mail])
  */

 #include "config.h"
@@ -39,12 +48,50 @@
 #include "a52.h"
 #include "a52_internal.h"
 #include "mm_accel.h"
+#include "mangle.h"
+
+void (*a52_imdct_512) (sample_t * data, sample_t * delay, sample_t bias);
+
+#if CONFIG_RUNTIME_CPUDETECT
+#undef HAVE_AMD3DNOWEXT
+#define HAVE_AMD3DNOWEXT 0
+#endif

 typedef struct complex_s {
     sample_t real;
     sample_t imag;
 } complex_t;

+static const int pm128[128] attribute_used __attribute__((aligned(16))) =
+{
+	0, 16, 32, 48, 64, 80,  96, 112,  8, 40, 72, 104, 24, 56,  88, 120,
+	4, 20, 36, 52, 68, 84, 100, 116, 12, 28, 44,  60, 76, 92, 108, 124,
+	2, 18, 34, 50, 66, 82,  98, 114, 10, 42, 74, 106, 26, 58,  90, 122,
+	6, 22, 38, 54, 70, 86, 102, 118, 14, 46, 78, 110, 30, 62,  94, 126,
+	1, 17, 33, 49, 65, 81,  97, 113,  9, 41, 73, 105, 25, 57,  89, 121,
+	5, 21, 37, 53, 69, 85, 101, 117, 13, 29, 45,  61, 77, 93, 109, 125,
+	3, 19, 35, 51, 67, 83,  99, 115, 11, 43, 75, 107, 27, 59,  91, 123,
+	7, 23, 39, 55, 71, 87, 103, 119, 15, 31, 47,  63, 79, 95, 111, 127
+};
+
+static uint8_t attribute_used bit_reverse_512[] = {
+	0x00, 0x40, 0x20, 0x60, 0x10, 0x50, 0x30, 0x70,
+	0x08, 0x48, 0x28, 0x68, 0x18, 0x58, 0x38, 0x78,
+	0x04, 0x44, 0x24, 0x64, 0x14, 0x54, 0x34, 0x74,
+	0x0c, 0x4c, 0x2c, 0x6c, 0x1c, 0x5c, 0x3c, 0x7c,
+	0x02, 0x42, 0x22, 0x62, 0x12, 0x52, 0x32, 0x72,
+	0x0a, 0x4a, 0x2a, 0x6a, 0x1a, 0x5a, 0x3a, 0x7a,
+	0x06, 0x46, 0x26, 0x66, 0x16, 0x56, 0x36, 0x76,
+	0x0e, 0x4e, 0x2e, 0x6e, 0x1e, 0x5e, 0x3e, 0x7e,
+	0x01, 0x41, 0x21, 0x61, 0x11, 0x51, 0x31, 0x71,
+	0x09, 0x49, 0x29, 0x69, 0x19, 0x59, 0x39, 0x79,
+	0x05, 0x45, 0x25, 0x65, 0x15, 0x55, 0x35, 0x75,
+	0x0d, 0x4d, 0x2d, 0x6d, 0x1d, 0x5d, 0x3d, 0x7d,
+	0x03, 0x43, 0x23, 0x63, 0x13, 0x53, 0x33, 0x73,
+	0x0b, 0x4b, 0x2b, 0x6b, 0x1b, 0x5b, 0x3b, 0x7b,
+	0x07, 0x47, 0x27, 0x67, 0x17, 0x57, 0x37, 0x77,
+	0x0f, 0x4f, 0x2f, 0x6f, 0x1f, 0x5f, 0x3f, 0x7f};
+
 static uint8_t fftorder[] = {
       0,128, 64,192, 32,160,224, 96, 16,144, 80,208,240,112, 48,176,
       8,136, 72,200, 40,168,232,104,248,120, 56,184, 24,152,216, 88,
@@ -56,6 +103,40 @@
       6,134, 70,198, 38,166,230,102,246,118, 54,182, 22,150,214, 86
 };

+static complex_t __attribute__((aligned(16))) buf[128];
+
+/* Twiddle factor LUT */
+static complex_t __attribute__((aligned(16))) w_1[1];
+static complex_t __attribute__((aligned(16))) w_2[2];
+static complex_t __attribute__((aligned(16))) w_4[4];
+static complex_t __attribute__((aligned(16))) w_8[8];
+static complex_t __attribute__((aligned(16))) w_16[16];
+static complex_t __attribute__((aligned(16))) w_32[32];
+static complex_t __attribute__((aligned(16))) w_64[64];
+static complex_t __attribute__((aligned(16))) * w[7] = {w_1, w_2, w_4, w_8, w_16, w_32, w_64};
+
+/* Twiddle factors for IMDCT */
+static sample_t __attribute__((aligned(16))) xcos1[128];
+static sample_t __attribute__((aligned(16))) xsin1[128];
+
+#if ARCH_X86 || ARCH_X86_64
+// NOTE: SSE needs 16byte alignment or it will segfault
+//
+static float __attribute__((aligned(16))) sseSinCos1c[256];
+static float __attribute__((aligned(16))) sseSinCos1d[256];
+static float attribute_used __attribute__((aligned(16))) ps111_1[4]={1,1,1,-1};
+//static float __attribute__((aligned(16))) sseW0[4];
+static float __attribute__((aligned(16))) sseW1[8];
+static float __attribute__((aligned(16))) sseW2[16];
+static float __attribute__((aligned(16))) sseW3[32];
+static float __attribute__((aligned(16))) sseW4[64];
+static float __attribute__((aligned(16))) sseW5[128];
+static float __attribute__((aligned(16))) sseW6[256];
+static float __attribute__((aligned(16))) *sseW[7]=
+	{NULL /*sseW0*/,sseW1,sseW2,sseW3,sseW4,sseW5,sseW6};
+static float __attribute__((aligned(16))) sseWindow[512];
+#endif
+
 /* Root values for IFFT */
 static sample_t roots16[3];
 static sample_t roots32[7];
@@ -241,7 +322,7 @@
     ifft_pass (buf, roots128 - 32, 32);
 }

-void a52_imdct_512 (sample_t * data, sample_t * delay, sample_t bias)
+void imdct_do_512 (sample_t * data, sample_t * delay, sample_t bias)
 {
     int i, k;
     sample_t t_r, t_i, a_r, a_i, b_r, b_i, w_1, w_2;
@@ -285,6 +366,704 @@
     }
 }

+#if HAVE_ALTIVEC
+
+#ifdef HAVE_ALTIVEC_H
+#include <altivec.h>
+#endif
+
+// used to build registers permutation vectors (vcprm)
+// the 's' are for words in the _s_econd vector
+#define WORD_0 0x00,0x01,0x02,0x03
+#define WORD_1 0x04,0x05,0x06,0x07
+#define WORD_2 0x08,0x09,0x0a,0x0b
+#define WORD_3 0x0c,0x0d,0x0e,0x0f
+#define WORD_s0 0x10,0x11,0x12,0x13
+#define WORD_s1 0x14,0x15,0x16,0x17
+#define WORD_s2 0x18,0x19,0x1a,0x1b
+#define WORD_s3 0x1c,0x1d,0x1e,0x1f
+
+#define vcprm(a,b,c,d) (const vector unsigned char){WORD_ ## a, WORD_ ## b, WORD_ ## c, WORD_ ## d}
+#define vcii(a,b,c,d) (const vector float){FLOAT_ ## a, FLOAT_ ## b, FLOAT_ ## c, FLOAT_ ## d}
+
+#define FOUROF(a) {a,a,a,a}
+
+// vcprmle is used to keep the same index as in the SSE version.
+// it's the same as vcprm, with the index inversed
+// ('le' is Little Endian)
+#define vcprmle(a,b,c,d) vcprm(d,c,b,a)
+
+// used to build inverse/identity vectors (vcii)
+// n is _n_egative, p is _p_ositive
+#define FLOAT_n -1.
+#define FLOAT_p 1.
+
+
+void
+imdct_do_512_altivec(sample_t data[],sample_t delay[], sample_t bias)
+{
+  int i;
+  int k;
+  int p,q;
+  int m;
+  long two_m;
+  long two_m_plus_one;
+
+  sample_t tmp_b_i;
+  sample_t tmp_b_r;
+  sample_t tmp_a_i;
+  sample_t tmp_a_r;
+
+  sample_t *data_ptr;
+  sample_t *delay_ptr;
+  sample_t *window_ptr;
+
+  /* 512 IMDCT with source and dest data in 'data' */
+
+  /* Pre IFFT complex multiply plus IFFT cmplx conjugate & reordering*/
+  for( i=0; i < 128; i++) {
+    /* z[i] = (X[256-2*i-1] + j * X[2*i]) * (xcos1[i] + j * xsin1[i]) ; */
+    int j= bit_reverse_512[i];
+    buf[i].real =         (data[256-2*j-1] * xcos1[j])  -  (data[2*j]       * xsin1[j]);
+    buf[i].imag = -1.0 * ((data[2*j]       * xcos1[j])  +  (data[256-2*j-1] * xsin1[j]));
+  }
+
+  /* 1. iteration */
+  for(i = 0; i < 128; i += 2) {
+#if 0
+    tmp_a_r = buf[i].real;
+    tmp_a_i = buf[i].imag;
+    tmp_b_r = buf[i+1].real;
+    tmp_b_i = buf[i+1].imag;
+    buf[i].real = tmp_a_r + tmp_b_r;
+    buf[i].imag =  tmp_a_i + tmp_b_i;
+    buf[i+1].real = tmp_a_r - tmp_b_r;
+    buf[i+1].imag =  tmp_a_i - tmp_b_i;
+#else
+    vector float temp, bufv;
+
+    bufv = vec_ld(i << 3, (float*)buf);
+    temp = vec_perm(bufv, bufv, vcprm(2,3,0,1));
+    bufv = vec_madd(bufv, vcii(p,p,n,n), temp);
+    vec_st(bufv, i << 3, (float*)buf);
+#endif
+  }
+
+  /* 2. iteration */
+  // Note w[1]={{1,0}, {0,-1}}
+  for(i = 0; i < 128; i += 4) {
+#if 0
+    tmp_a_r = buf[i].real;
+    tmp_a_i = buf[i].imag;
+    tmp_b_r = buf[i+2].real;
+    tmp_b_i = buf[i+2].imag;
+    buf[i].real = tmp_a_r + tmp_b_r;
+    buf[i].imag =  tmp_a_i + tmp_b_i;
+    buf[i+2].real = tmp_a_r - tmp_b_r;
+    buf[i+2].imag =  tmp_a_i - tmp_b_i;
+    tmp_a_r = buf[i+1].real;
+    tmp_a_i = buf[i+1].imag;
+    /* WARNING: im <-> re here ! */
+    tmp_b_r = buf[i+3].imag;
+    tmp_b_i = buf[i+3].real;
+    buf[i+1].real = tmp_a_r + tmp_b_r;
+    buf[i+1].imag =  tmp_a_i - tmp_b_i;
+    buf[i+3].real = tmp_a_r - tmp_b_r;
+    buf[i+3].imag =  tmp_a_i + tmp_b_i;
+#else
+    vector float buf01, buf23, temp1, temp2;
+
+    buf01 = vec_ld((i + 0) << 3, (float*)buf);
+    buf23 = vec_ld((i + 2) << 3, (float*)buf);
+    buf23 = vec_perm(buf23,buf23,vcprm(0,1,3,2));
+
+    temp1 = vec_madd(buf23, vcii(p,p,p,n), buf01);
+    temp2 = vec_madd(buf23, vcii(n,n,n,p), buf01);
+
+    vec_st(temp1, (i + 0) << 3, (float*)buf);
+    vec_st(temp2, (i + 2) << 3, (float*)buf);
+#endif
+  }
+
+  /* 3. iteration */
+  for(i = 0; i < 128; i += 8) {
+#if 0
+    tmp_a_r = buf[i].real;
+    tmp_a_i = buf[i].imag;
+    tmp_b_r = buf[i+4].real;
+    tmp_b_i = buf[i+4].imag;
+    buf[i].real = tmp_a_r + tmp_b_r;
+    buf[i].imag =  tmp_a_i + tmp_b_i;
+    buf[i+4].real = tmp_a_r - tmp_b_r;
+    buf[i+4].imag =  tmp_a_i - tmp_b_i;
+    tmp_a_r = buf[1+i].real;
+    tmp_a_i = buf[1+i].imag;
+    tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
+    tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
+    buf[1+i].real = tmp_a_r + tmp_b_r;
+    buf[1+i].imag =  tmp_a_i + tmp_b_i;
+    buf[i+5].real = tmp_a_r - tmp_b_r;
+    buf[i+5].imag =  tmp_a_i - tmp_b_i;
+    tmp_a_r = buf[i+2].real;
+    tmp_a_i = buf[i+2].imag;
+    /* WARNING re <-> im & sign */
+    tmp_b_r = buf[i+6].imag;
+    tmp_b_i = - buf[i+6].real;
+    buf[i+2].real = tmp_a_r + tmp_b_r;
+    buf[i+2].imag =  tmp_a_i + tmp_b_i;
+    buf[i+6].real = tmp_a_r - tmp_b_r;
+    buf[i+6].imag =  tmp_a_i - tmp_b_i;
+    tmp_a_r = buf[i+3].real;
+    tmp_a_i = buf[i+3].imag;
+    tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
+    tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
+    buf[i+3].real = tmp_a_r + tmp_b_r;
+    buf[i+3].imag =  tmp_a_i + tmp_b_i;
+    buf[i+7].real = tmp_a_r - tmp_b_r;
+    buf[i+7].imag =  tmp_a_i - tmp_b_i;
+#else
+    vector float buf01, buf23, buf45, buf67;
+
+    buf01 = vec_ld((i + 0) << 3, (float*)buf);
+    buf23 = vec_ld((i + 2) << 3, (float*)buf);
+
+    tmp_b_r = (buf[i+5].real + buf[i+5].imag) * w[2][1].real;
+    tmp_b_i = (buf[i+5].imag - buf[i+5].real) * w[2][1].real;
+    buf[i+5].real = tmp_b_r;
+    buf[i+5].imag = tmp_b_i;
+    tmp_b_r = (buf[i+7].real - buf[i+7].imag) * w[2][3].imag;
+    tmp_b_i = (buf[i+7].imag + buf[i+7].real) * w[2][3].imag;
+    buf[i+7].real = tmp_b_r;
+    buf[i+7].imag = tmp_b_i;
+
+    buf23 = vec_ld((i + 2) << 3, (float*)buf);
+    buf45 = vec_ld((i + 4) << 3, (float*)buf);
+    buf67 = vec_ld((i + 6) << 3, (float*)buf);
+    buf67 = vec_perm(buf67, buf67, vcprm(1,0,2,3));
+
+    vec_st(vec_add(buf01, buf45), (i + 0) << 3, (float*)buf);
+    vec_st(vec_madd(buf67, vcii(p,n,p,p), buf23), (i + 2) << 3, (float*)buf);
+    vec_st(vec_sub(buf01, buf45), (i + 4) << 3, (float*)buf);
+    vec_st(vec_nmsub(buf67, vcii(p,n,p,p), buf23), (i + 6) << 3, (float*)buf);
+#endif
+  }
+
+  /* 4-7. iterations */
+  for (m=3; m < 7; m++) {
+    two_m = (1 << m);
+
+    two_m_plus_one = two_m<<1;
+
+    for(i = 0; i < 128; i += two_m_plus_one) {
+      for(k = 0; k < two_m; k+=2) {
+#if 0
+        int p = k + i;
+        int q = p + two_m;
+        tmp_a_r = buf[p].real;
+        tmp_a_i = buf[p].imag;
+        tmp_b_r =
+          buf[q].real * w[m][k].real -
+          buf[q].imag * w[m][k].imag;
+        tmp_b_i =
+          buf[q].imag * w[m][k].real +
+          buf[q].real * w[m][k].imag;
+        buf[p].real = tmp_a_r + tmp_b_r;
+        buf[p].imag =  tmp_a_i + tmp_b_i;
+        buf[q].real = tmp_a_r - tmp_b_r;
+        buf[q].imag =  tmp_a_i - tmp_b_i;
+
+        tmp_a_r = buf[(p + 1)].real;
+        tmp_a_i = buf[(p + 1)].imag;
+        tmp_b_r =
+          buf[(q + 1)].real * w[m][(k + 1)].real -
+          buf[(q + 1)].imag * w[m][(k + 1)].imag;
+        tmp_b_i =
+          buf[(q + 1)].imag * w[m][(k + 1)].real +
+          buf[(q + 1)].real * w[m][(k + 1)].imag;
+        buf[(p + 1)].real = tmp_a_r + tmp_b_r;
+        buf[(p + 1)].imag =  tmp_a_i + tmp_b_i;
+        buf[(q + 1)].real = tmp_a_r - tmp_b_r;
+        buf[(q + 1)].imag =  tmp_a_i - tmp_b_i;
+#else
+        int p = k + i;
+        int q = p + two_m;
+        vector float vecp, vecq, vecw, temp1, temp2, temp3, temp4;
+        const vector float vczero = (const vector float)FOUROF(0.);
+        // first compute buf[q] and buf[q+1]
+        vecq = vec_ld(q << 3, (float*)buf);
+        vecw = vec_ld(0, (float*)&(w[m][k]));
+        temp1 = vec_madd(vecq, vecw, vczero);
+        temp2 = vec_perm(vecq, vecq, vcprm(1,0,3,2));
+        temp2 = vec_madd(temp2, vecw, vczero);
+        temp3 = vec_perm(temp1, temp2, vcprm(0,s0,2,s2));
+        temp4 = vec_perm(temp1, temp2, vcprm(1,s1,3,s3));
+        vecq = vec_madd(temp4, vcii(n,p,n,p), temp3);
+        // then butterfly with buf[p] and buf[p+1]
+        vecp = vec_ld(p << 3, (float*)buf);
+
+        temp1 = vec_add(vecp, vecq);
+        temp2 = vec_sub(vecp, vecq);
+
+        vec_st(temp1, p << 3, (float*)buf);
+        vec_st(temp2, q << 3, (float*)buf);
+#endif
+      }
+    }
+  }
+
+  /* Post IFFT complex multiply  plus IFFT complex conjugate*/
+  for( i=0; i < 128; i+=4) {
+    /* y[n] = z[n] * (xcos1[n] + j * xsin1[n]) ; */
+#if 0
+    tmp_a_r =        buf[(i + 0)].real;
+    tmp_a_i = -1.0 * buf[(i + 0)].imag;
+    buf[(i + 0)].real =
+      (tmp_a_r * xcos1[(i + 0)])  -  (tmp_a_i  * xsin1[(i + 0)]);
+    buf[(i + 0)].imag =
+      (tmp_a_r * xsin1[(i + 0)])  +  (tmp_a_i  * xcos1[(i + 0)]);
+
+    tmp_a_r =        buf[(i + 1)].real;
+    tmp_a_i = -1.0 * buf[(i + 1)].imag;
+    buf[(i + 1)].real =
+      (tmp_a_r * xcos1[(i + 1)])  -  (tmp_a_i  * xsin1[(i + 1)]);
+    buf[(i + 1)].imag =
+      (tmp_a_r * xsin1[(i + 1)])  +  (tmp_a_i  * xcos1[(i + 1)]);
+
+    tmp_a_r =        buf[(i + 2)].real;
+    tmp_a_i = -1.0 * buf[(i + 2)].imag;
+    buf[(i + 2)].real =
+      (tmp_a_r * xcos1[(i + 2)])  -  (tmp_a_i  * xsin1[(i + 2)]);
+    buf[(i + 2)].imag =
+      (tmp_a_r * xsin1[(i + 2)])  +  (tmp_a_i  * xcos1[(i + 2)]);
+
+    tmp_a_r =        buf[(i + 3)].real;
+    tmp_a_i = -1.0 * buf[(i + 3)].imag;
+    buf[(i + 3)].real =
+      (tmp_a_r * xcos1[(i + 3)])  -  (tmp_a_i  * xsin1[(i + 3)]);
+    buf[(i + 3)].imag =
+      (tmp_a_r * xsin1[(i + 3)])  +  (tmp_a_i  * xcos1[(i + 3)]);
+#else
+    vector float bufv_0, bufv_2, cosv, sinv, temp1, temp2;
+    vector float temp0022, temp1133, tempCS01;
+    const vector float vczero = (const vector float)FOUROF(0.);
+
+    bufv_0 = vec_ld((i + 0) << 3, (float*)buf);
+    bufv_2 = vec_ld((i + 2) << 3, (float*)buf);
+
+    cosv = vec_ld(i << 2, xcos1);
+    sinv = vec_ld(i << 2, xsin1);
+
+    temp0022 = vec_perm(bufv_0, bufv_0, vcprm(0,0,2,2));
+    temp1133 = vec_perm(bufv_0, bufv_0, vcprm(1,1,3,3));
+    tempCS01 = vec_perm(cosv, sinv, vcprm(0,s0,1,s1));
+    temp1 = vec_madd(temp0022, tempCS01, vczero);
+    tempCS01 = vec_perm(cosv, sinv, vcprm(s0,0,s1,1));
+    temp2 = vec_madd(temp1133, tempCS01, vczero);
+    bufv_0 = vec_madd(temp2, vcii(p,n,p,n), temp1);
+
+    vec_st(bufv_0, (i + 0) << 3, (float*)buf);
+
+    /* idem with bufv_2 and high-order cosv/sinv */
+
+    temp0022 = vec_perm(bufv_2, bufv_2, vcprm(0,0,2,2));
+    temp1133 = vec_perm(bufv_2, bufv_2, vcprm(1,1,3,3));
+    tempCS01 = vec_perm(cosv, sinv, vcprm(2,s2,3,s3));
+    temp1 = vec_madd(temp0022, tempCS01, vczero);
+    tempCS01 = vec_perm(cosv, sinv, vcprm(s2,2,s3,3));
+    temp2 = vec_madd(temp1133, tempCS01, vczero);
+    bufv_2 = vec_madd(temp2, vcii(p,n,p,n), temp1);
+
+    vec_st(bufv_2, (i + 2) << 3, (float*)buf);
+
+#endif
+  }
+
+  data_ptr = data;
+  delay_ptr = delay;
+  window_ptr = a52_imdct_window;
+
+  /* Window and convert to real valued signal */
+  for(i=0; i< 64; i++) {
+    *data_ptr++   = -buf[64+i].imag   * *window_ptr++ + *delay_ptr++ + bias;
+    *data_ptr++   =  buf[64-i-1].real * *window_ptr++ + *delay_ptr++ + bias;
+  }
+
+  for(i=0; i< 64; i++) {
+    *data_ptr++  = -buf[i].real       * *window_ptr++ + *delay_ptr++ + bias;
+    *data_ptr++  =  buf[128-i-1].imag * *window_ptr++ + *delay_ptr++ + bias;
+  }
+
+  /* The trailing edge of the window goes into the delay line */
+  delay_ptr = delay;
+
+  for(i=0; i< 64; i++) {
+    *delay_ptr++  = -buf[64+i].real   * *--window_ptr;
+    *delay_ptr++  =  buf[64-i-1].imag * *--window_ptr;
+  }
+
+  for(i=0; i<64; i++) {
+    *delay_ptr++  =  buf[i].imag       * *--window_ptr;
+    *delay_ptr++  = -buf[128-i-1].real * *--window_ptr;
+  }
+}
+#endif
+
+
+// Stuff below this line is borrowed from libac3
+#include "srfftp.h"
+#if ARCH_X86 || ARCH_X86_64
+#undef HAVE_AMD3DNOW
+#define HAVE_AMD3DNOW 1
+#include "srfftp_3dnow.h"
+
+const i_cmplx_t x_plus_minus_3dnow __attribute__ ((aligned (8))) = {{ 0x00000000UL, 0x80000000UL }};
+const i_cmplx_t x_minus_plus_3dnow __attribute__ ((aligned (8))) = {{ 0x80000000UL, 0x00000000UL }};
+const complex_t HSQRT2_3DNOW __attribute__ ((aligned (8))) = { 0.707106781188, 0.707106781188 };
+
+#undef HAVE_AMD3DNOWEXT
+#define HAVE_AMD3DNOWEXT 0
+#include "imdct_3dnow.h"
+#undef HAVE_AMD3DNOWEXT
+#define HAVE_AMD3DNOWEXT 1
+#include "imdct_3dnow.h"
+
+#if !ARCH_X86_64 || !defined(PIC)
+void
+imdct_do_512_sse(sample_t data[],sample_t delay[], sample_t bias)
+{
+/*	int i,k;
+    int p,q;*/
+    int m;
+    long two_m;
+    long two_m_plus_one;
+    long two_m_plus_one_shl3;
+    complex_t *buf_offset;
+
+/*  sample_t tmp_a_i;
+    sample_t tmp_a_r;
+    sample_t tmp_b_i;
+    sample_t tmp_b_r;*/
+
+    sample_t *data_ptr;
+    sample_t *delay_ptr;
+    sample_t *window_ptr;
+
+    /* 512 IMDCT with source and dest data in 'data' */
+    /* see the c version (dct_do_512()), its allmost identical, just in C */
+
+    /* Pre IFFT complex multiply plus IFFT cmplx conjugate */
+    /* Bit reversed shuffling */
+	__asm__ volatile(
+		"xor %%"REG_S", %%"REG_S"		\n\t"
+		"lea "MANGLE(bit_reverse_512)", %%"REG_a"\n\t"
+		"mov $1008, %%"REG_D"			\n\t"
+		"push %%"REG_BP"			\n\t" //use ebp without telling gcc
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movlps (%0, %%"REG_S"), %%xmm0	\n\t" // XXXI
+		"movhps 8(%0, %%"REG_D"), %%xmm0	\n\t" // RXXI
+		"movlps 8(%0, %%"REG_S"), %%xmm1	\n\t" // XXXi
+		"movhps (%0, %%"REG_D"), %%xmm1	\n\t" // rXXi
+		"shufps $0x33, %%xmm1, %%xmm0		\n\t" // irIR
+		"movaps "MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm2\n\t"
+		"mulps %%xmm0, %%xmm2			\n\t"
+		"shufps $0xB1, %%xmm0, %%xmm0		\n\t" // riRI
+		"mulps "MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
+		"subps %%xmm0, %%xmm2			\n\t"
+		"movzb (%%"REG_a"), %%"REG_d"		\n\t"
+		"movzb 1(%%"REG_a"), %%"REG_BP"		\n\t"
+		"movlps %%xmm2, (%1, %%"REG_d", 8)	\n\t"
+		"movhps %%xmm2, (%1, %%"REG_BP", 8)	\n\t"
+		"add $16, %%"REG_S"			\n\t"
+		"add $2, %%"REG_a"			\n\t" // avoid complex addressing for P4 crap
+		"sub $16, %%"REG_D"			\n\t"
+		"jnc 1b				 	\n\t"
+		"pop %%"REG_BP"				\n\t"//no we didnt touch ebp *g*
+		:: "b" (data), "c" (buf)
+		: "%"REG_S, "%"REG_D, "%"REG_a, "%"REG_d
+	);
+
+
+    /* FFT Merge */
+/* unoptimized variant
+    for (m=1; m < 7; m++) {
+	if(m)
+	    two_m = (1 << m);
+	else
+	    two_m = 1;
+
+	two_m_plus_one = (1 << (m+1));
+
+	for(i = 0; i < 128; i += two_m_plus_one) {
+	    for(k = 0; k < two_m; k++) {
+		p = k + i;
+		q = p + two_m;
+		tmp_a_r = buf[p].real;
+		tmp_a_i = buf[p].imag;
+		tmp_b_r = buf[q].real * w[m][k].real - buf[q].imag * w[m][k].imag;
+		tmp_b_i = buf[q].imag * w[m][k].real + buf[q].real * w[m][k].imag;
+		buf[p].real = tmp_a_r + tmp_b_r;
+		buf[p].imag =  tmp_a_i + tmp_b_i;
+		buf[q].real = tmp_a_r - tmp_b_r;
+		buf[q].imag =  tmp_a_i - tmp_b_i;
+	    }
+	}
+    }
+*/
+
+    /* 1. iteration */
+	// Note w[0][0]={1,0}
+	__asm__ volatile(
+		"xorps %%xmm1, %%xmm1	\n\t"
+		"xorps %%xmm2, %%xmm2	\n\t"
+		"mov %0, %%"REG_S"	\n\t"
+		ASMALIGN(4)
+		"1:			\n\t"
+		"movlps (%%"REG_S"), %%xmm0\n\t" //buf[p]
+		"movlps 8(%%"REG_S"), %%xmm1\n\t" //buf[q]
+		"movhps (%%"REG_S"), %%xmm0\n\t" //buf[p]
+		"movhps 8(%%"REG_S"), %%xmm2\n\t" //buf[q]
+		"addps %%xmm1, %%xmm0	\n\t"
+		"subps %%xmm2, %%xmm0	\n\t"
+		"movaps %%xmm0, (%%"REG_S")\n\t"
+		"add $16, %%"REG_S"	\n\t"
+		"cmp %1, %%"REG_S"	\n\t"
+		" jb 1b			\n\t"
+		:: "g" (buf), "r" (buf + 128)
+		: "%"REG_S
+	);
+
+    /* 2. iteration */
+	// Note w[1]={{1,0}, {0,-1}}
+	__asm__ volatile(
+		"movaps "MANGLE(ps111_1)", %%xmm7\n\t" // 1,1,1,-1
+		"mov %0, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 16(%%"REG_S"), %%xmm2	\n\t" //r2,i2,r3,i3
+		"shufps $0xB4, %%xmm2, %%xmm2	\n\t" //r2,i2,i3,r3
+		"mulps %%xmm7, %%xmm2		\n\t" //r2,i2,i3,-r3
+		"movaps (%%"REG_S"), %%xmm0	\n\t" //r0,i0,r1,i1
+		"movaps (%%"REG_S"), %%xmm1	\n\t" //r0,i0,r1,i1
+		"addps %%xmm2, %%xmm0		\n\t"
+		"subps %%xmm2, %%xmm1		\n\t"
+		"movaps %%xmm0, (%%"REG_S")	\n\t"
+		"movaps %%xmm1, 16(%%"REG_S")	\n\t"
+		"add $32, %%"REG_S"	\n\t"
+		"cmp %1, %%"REG_S"	\n\t"
+		" jb 1b			\n\t"
+		:: "g" (buf), "r" (buf + 128)
+		: "%"REG_S
+	);
+
+    /* 3. iteration */
+/*
+ Note sseW2+0={1,1,sqrt(2),sqrt(2))
+ Note sseW2+16={0,0,sqrt(2),-sqrt(2))
+ Note sseW2+32={0,0,-sqrt(2),-sqrt(2))
+ Note sseW2+48={1,-1,sqrt(2),-sqrt(2))
+*/
+	__asm__ volatile(
+		"movaps 48+"MANGLE(sseW2)", %%xmm6\n\t"
+		"movaps 16+"MANGLE(sseW2)", %%xmm7\n\t"
+		"xorps %%xmm5, %%xmm5		\n\t"
+		"xorps %%xmm2, %%xmm2		\n\t"
+		"mov %0, %%"REG_S"		\n\t"
+		ASMALIGN(4)
+		"1:				\n\t"
+		"movaps 32(%%"REG_S"), %%xmm2	\n\t" //r4,i4,r5,i5
+		"movaps 48(%%"REG_S"), %%xmm3	\n\t" //r6,i6,r7,i7
+		"movaps "MANGLE(sseW2)", %%xmm4	\n\t" //r4,i4,r5,i5
+		"movaps 32+"MANGLE(sseW2)", %%xmm5\n\t" //r6,i6,r7,i7
+		"mulps %%xmm2, %%xmm4		\n\t"
+		"mulps %%xmm3, %%xmm5		\n\t"
+		"shufps $0xB1, %%xmm2, %%xmm2	\n\t" //i4,r4,i5,r5
+		"shufps $0xB1, %%xmm3, %%xmm3	\n\t" //i6,r6,i7,r7
+		"mulps %%xmm6, %%xmm3		\n\t"
+		"mulps %%xmm7, %%xmm2		\n\t"
+		"movaps (%%"REG_S"), %%xmm0	\n\t" //r0,i0,r1,i1
+		"movaps 16(%%"REG_S"), %%xmm1	\n\t" //r2,i2,r3,i3
+		"addps %%xmm4, %%xmm2		\n\t"
+		"addps %%xmm5, %%xmm3		\n\t"
+		"movaps %%xmm2, %%xmm4		\n\t"
+		"movaps %%xmm3, %%xmm5		\n\t"
+		"addps %%xmm0, %%xmm2		\n\t"
+		"addps %%xmm1, %%xmm3		\n\t"
+		"subps %%xmm4, %%xmm0		\n\t"
+		"subps %%xmm5, %%xmm1		\n\t"
+		"movaps %%xmm2, (%%"REG_S")	\n\t"
+		"movaps %%xmm3, 16(%%"REG_S")	\n\t"
+		"movaps %%xmm0, 32(%%"REG_S")	\n\t"
+		"movaps %%xmm1, 48(%%"REG_S")	\n\t"
+		"add $64, %%"REG_S"	\n\t"
+		"cmp %1, %%"REG_S"	\n\t"
+		" jb 1b			\n\t"
+		:: "g" (buf), "r" (buf + 128)
+		: "%"REG_S
+	);
+
+    /* 4-7. iterations */
+    for (m=3; m < 7; m++) {
+	two_m = (1 << m);
+	two_m_plus_one = two_m<<1;
+	two_m_plus_one_shl3 = (two_m_plus_one<<3);
+	buf_offset = buf+128;
+	__asm__ volatile(
+		"mov %0, %%"REG_S"			\n\t"
+		ASMALIGN(4)
+		"1:					\n\t"
+		"xor %%"REG_D", %%"REG_D"		\n\t" // k
+		"lea (%%"REG_S", %3), %%"REG_d"		\n\t"
+		"2:					\n\t"
+		"movaps (%%"REG_d", %%"REG_D"), %%xmm1	\n\t"
+		"movaps (%4, %%"REG_D", 2), %%xmm2	\n\t"
+		"mulps %%xmm1, %%xmm2			\n\t"
+		"shufps $0xB1, %%xmm1, %%xmm1		\n\t"
+		"mulps 16(%4, %%"REG_D", 2), %%xmm1	\n\t"
+		"movaps (%%"REG_S", %%"REG_D"), %%xmm0	\n\t"
+		"addps %%xmm2, %%xmm1			\n\t"
+		"movaps %%xmm1, %%xmm2			\n\t"
+		"addps %%xmm0, %%xmm1			\n\t"
+		"subps %%xmm2, %%xmm0			\n\t"
+		"movaps %%xmm1, (%%"REG_S", %%"REG_D")	\n\t"
+		"movaps %%xmm0, (%%"REG_d", %%"REG_D")	\n\t"
+		"add $16, %%"REG_D"			\n\t"
+		"cmp %3, %%"REG_D"			\n\t" //FIXME (opt) count against 0
+		"jb 2b					\n\t"
+		"add %2, %%"REG_S"			\n\t"
+		"cmp %1, %%"REG_S"			\n\t"
+		" jb 1b					\n\t"
+		:: "g" (buf), "m" (buf_offset), "m" (two_m_plus_one_shl3), "r" (two_m<<3),
+		   "r" (sseW[m])
+		: "%"REG_S, "%"REG_D, "%"REG_d
+	);
+    }
+
+    /* Post IFFT complex multiply  plus IFFT complex conjugate*/
+	__asm__ volatile(
+		"mov $-1024, %%"REG_S"			\n\t"
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movaps (%0, %%"REG_S"), %%xmm0		\n\t"
+		"movaps (%0, %%"REG_S"), %%xmm1		\n\t"
+		"shufps $0xB1, %%xmm0, %%xmm0		\n\t"
+		"mulps 1024+"MANGLE(sseSinCos1c)"(%%"REG_S"), %%xmm1\n\t"
+		"mulps 1024+"MANGLE(sseSinCos1d)"(%%"REG_S"), %%xmm0\n\t"
+		"addps %%xmm1, %%xmm0			\n\t"
+		"movaps %%xmm0, (%0, %%"REG_S")		\n\t"
+		"add $16, %%"REG_S"			\n\t"
+		" jnz 1b				\n\t"
+		:: "r" (buf+128)
+		: "%"REG_S
+	);
+
+
+    data_ptr = data;
+    delay_ptr = delay;
+    window_ptr = a52_imdct_window;
+
+    /* Window and convert to real valued signal */
+	__asm__ volatile(
+		"xor %%"REG_D", %%"REG_D"		\n\t"  // 0
+		"xor %%"REG_S", %%"REG_S"		\n\t"  // 0
+		"movss %3, %%xmm2			\n\t"  // bias
+		"shufps $0x00, %%xmm2, %%xmm2		\n\t"  // bias, bias, ...
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movlps (%0, %%"REG_S"), %%xmm0		\n\t" // ? ? A ?
+		"movlps 8(%0, %%"REG_S"), %%xmm1	\n\t" // ? ? C ?
+		"movhps -16(%0, %%"REG_D"), %%xmm1	\n\t" // ? D C ?
+		"movhps -8(%0, %%"REG_D"), %%xmm0	\n\t" // ? B A ?
+		"shufps $0x99, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps "MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
+		"addps (%2, %%"REG_S"), %%xmm0		\n\t"
+		"addps %%xmm2, %%xmm0			\n\t"
+		"movaps %%xmm0, (%1, %%"REG_S")		\n\t"
+		"add  $16, %%"REG_S"			\n\t"
+		"sub  $16, %%"REG_D"			\n\t"
+		"cmp  $512, %%"REG_S"			\n\t"
+		" jb 1b					\n\t"
+		:: "r" (buf+64), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
+		: "%"REG_S, "%"REG_D
+	);
+	data_ptr+=128;
+	delay_ptr+=128;
+//	window_ptr+=128;
+
+	__asm__ volatile(
+		"mov $1024, %%"REG_D"			\n\t"  // 512
+		"xor %%"REG_S", %%"REG_S"		\n\t"  // 0
+		"movss %3, %%xmm2			\n\t"  // bias
+		"shufps $0x00, %%xmm2, %%xmm2		\n\t"  // bias, bias, ...
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movlps (%0, %%"REG_S"), %%xmm0		\n\t" // ? ? ? A
+		"movlps 8(%0, %%"REG_S"), %%xmm1	\n\t" // ? ? ? C
+		"movhps -16(%0, %%"REG_D"), %%xmm1	\n\t" // D ? ? C
+		"movhps -8(%0, %%"REG_D"), %%xmm0	\n\t" // B ? ? A
+		"shufps $0xCC, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps 512+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
+		"addps (%2, %%"REG_S"), %%xmm0		\n\t"
+		"addps %%xmm2, %%xmm0			\n\t"
+		"movaps %%xmm0, (%1, %%"REG_S")		\n\t"
+		"add $16, %%"REG_S"			\n\t"
+		"sub $16, %%"REG_D"			\n\t"
+		"cmp $512, %%"REG_S"			\n\t"
+		" jb 1b					\n\t"
+		:: "r" (buf), "r" (data_ptr), "r" (delay_ptr), "m" (bias)
+		: "%"REG_S, "%"REG_D
+	);
+	data_ptr+=128;
+//	window_ptr+=128;
+
+    /* The trailing edge of the window goes into the delay line */
+    delay_ptr = delay;
+
+	__asm__ volatile(
+		"xor %%"REG_D", %%"REG_D"		\n\t"  // 0
+		"xor %%"REG_S", %%"REG_S"		\n\t"  // 0
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movlps (%0, %%"REG_S"), %%xmm0		\n\t" // ? ? ? A
+		"movlps 8(%0, %%"REG_S"), %%xmm1	\n\t" // ? ? ? C
+		"movhps -16(%0, %%"REG_D"), %%xmm1	\n\t" // D ? ? C
+		"movhps -8(%0, %%"REG_D"), %%xmm0	\n\t" // B ? ? A
+		"shufps $0xCC, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps 1024+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
+		"movaps %%xmm0, (%1, %%"REG_S")		\n\t"
+		"add $16, %%"REG_S"			\n\t"
+		"sub $16, %%"REG_D"			\n\t"
+		"cmp $512, %%"REG_S"			\n\t"
+		" jb 1b					\n\t"
+		:: "r" (buf+64), "r" (delay_ptr)
+		: "%"REG_S, "%"REG_D
+	);
+	delay_ptr+=128;
+//	window_ptr-=128;
+
+	__asm__ volatile(
+		"mov $1024, %%"REG_D"			\n\t"  // 1024
+		"xor %%"REG_S", %%"REG_S"		\n\t"  // 0
+		ASMALIGN(4)
+		"1:					\n\t"
+		"movlps (%0, %%"REG_S"), %%xmm0	\n\t" // ? ? A ?
+		"movlps 8(%0, %%"REG_S"), %%xmm1	\n\t" // ? ? C ?
+		"movhps -16(%0, %%"REG_D"), %%xmm1	\n\t" // ? D C ?
+		"movhps -8(%0, %%"REG_D"), %%xmm0	\n\t" // ? B A ?
+		"shufps $0x99, %%xmm1, %%xmm0		\n\t" // D C B A
+		"mulps 1536+"MANGLE(sseWindow)"(%%"REG_S"), %%xmm0\n\t"
+		"movaps %%xmm0, (%1, %%"REG_S")		\n\t"
+		"add $16, %%"REG_S"			\n\t"
+		"sub $16, %%"REG_D"			\n\t"
+		"cmp $512, %%"REG_S"			\n\t"
+		" jb 1b					\n\t"
+		:: "r" (buf), "r" (delay_ptr)
+		: "%"REG_S, "%"REG_D
+	);
+}
+#endif
+#endif // ARCH_X86 || ARCH_X86_64
+
 void a52_imdct_256(sample_t * data, sample_t * delay, sample_t bias)
 {
     int i, k;
@@ -364,7 +1143,7 @@

 void a52_imdct_init (uint32_t mm_accel)
 {
-    int i, k;
+    int i, j, k;
     double sum;

     /* compute imdct window - kaiser-bessel derived window, alpha = 5.0 */
@@ -416,6 +1195,101 @@
 	post2[i].real = cos ((M_PI / 128) * (i + 0.5));
 	post2[i].imag = sin ((M_PI / 128) * (i + 0.5));
     }
+    for (i = 0; i < 128; i++) {
+	xcos1[i] = -cos ((M_PI / 2048) * (8 * i + 1));
+	xsin1[i] = -sin ((M_PI / 2048) * (8 * i + 1));
+    }
+    for (i = 0; i < 7; i++) {
+	j = 1 << i;
+	for (k = 0; k < j; k++) {
+	    w[i][k].real = cos (-M_PI * k / j);
+	    w[i][k].imag = sin (-M_PI * k / j);
+	}
+    }
+#if ARCH_X86 || ARCH_X86_64
+	for (i = 0; i < 128; i++) {
+	    sseSinCos1c[2*i+0]= xcos1[i];
+	    sseSinCos1c[2*i+1]= -xcos1[i];
+	    sseSinCos1d[2*i+0]= xsin1[i];
+	    sseSinCos1d[2*i+1]= xsin1[i];
+	}
+	for (i = 1; i < 7; i++) {
+	    j = 1 << i;
+	    for (k = 0; k < j; k+=2) {
+
+	    	sseW[i][4*k + 0] = w[i][k+0].real;
+	    	sseW[i][4*k + 1] = w[i][k+0].real;
+	    	sseW[i][4*k + 2] = w[i][k+1].real;
+	    	sseW[i][4*k + 3] = w[i][k+1].real;
+
+	    	sseW[i][4*k + 4] = -w[i][k+0].imag;
+	    	sseW[i][4*k + 5] = w[i][k+0].imag;
+	    	sseW[i][4*k + 6] = -w[i][k+1].imag;
+	    	sseW[i][4*k + 7] = w[i][k+1].imag;
+
+	//we multiply more or less uninitalized numbers so we need to use exactly 0.0
+		if(k==0)
+		{
+//			sseW[i][4*k + 0]= sseW[i][4*k + 1]= 1.0;
+			sseW[i][4*k + 4]= sseW[i][4*k + 5]= 0.0;
+		}
+
+		if(2*k == j)
+		{
+			sseW[i][4*k + 0]= sseW[i][4*k + 1]= 0.0;
+//			sseW[i][4*k + 4]= -(sseW[i][4*k + 5]= -1.0);
+		}
+	    }
+	}
+
+	for(i=0; i<128; i++)
+	{
+		sseWindow[2*i+0]= -a52_imdct_window[2*i+0];
+		sseWindow[2*i+1]=  a52_imdct_window[2*i+1];
+	}
+
+	for(i=0; i<64; i++)
+	{
+		sseWindow[256 + 2*i+0]= -a52_imdct_window[254 - 2*i+1];
+		sseWindow[256 + 2*i+1]=  a52_imdct_window[254 - 2*i+0];
+		sseWindow[384 + 2*i+0]=  a52_imdct_window[126 - 2*i+1];
+		sseWindow[384 + 2*i+1]= -a52_imdct_window[126 - 2*i+0];
+	}
+#endif
+	a52_imdct_512 = imdct_do_512;
+	ifft128 = ifft128_c;
+	ifft64 = ifft64_c;
+
+#if ARCH_X86 || ARCH_X86_64
+#if !ARCH_X86_64 || !defined(PIC)
+	if(mm_accel & MM_ACCEL_X86_SSE)
+	{
+	  fprintf (stderr, "Using SSE optimized IMDCT transform\n");
+	  a52_imdct_512 = imdct_do_512_sse;
+	}
+	else
+#endif
+	if(mm_accel & MM_ACCEL_X86_3DNOWEXT)
+	{
+	  fprintf (stderr, "Using 3DNowEx optimized IMDCT transform\n");
+	  a52_imdct_512 = imdct_do_512_3dnowex;
+	}
+	else
+	if(mm_accel & MM_ACCEL_X86_3DNOW)
+	{
+	  fprintf (stderr, "Using 3DNow optimized IMDCT transform\n");
+	  a52_imdct_512 = imdct_do_512_3dnow;
+	}
+	else
+#endif // ARCH_X86 || ARCH_X86_64
+#if HAVE_ALTIVEC
+        if (mm_accel & MM_ACCEL_PPC_ALTIVEC)
+	{
+	  fprintf(stderr, "Using AltiVec optimized IMDCT transform\n");
+          a52_imdct_512 = imdct_do_512_altivec;
+	}
+	else
+#endif

 #ifdef LIBA52_DJBFFT
     if (mm_accel & MM_ACCEL_DJBFFT) {
@@ -426,7 +1300,5 @@
 #endif
     {
 	fprintf (stderr, "No accelerated IMDCT transform found\n");
-	ifft128 = ifft128_c;
-	ifft64 = ifft64_c;
     }
 }
--- include/mm_accel.h	2006-06-12 15:05:00.000000000 +0200
+++ liba52/mm_accel.h	2006-06-05 02:23:04.000000000 +0200
@@ -30,7 +34,12 @@
 /* x86 accelerations */
 #define MM_ACCEL_X86_MMX	0x80000000
 #define MM_ACCEL_X86_3DNOW	0x40000000
+#define MM_ACCEL_X86_3DNOWEXT	0x08000000
 #define MM_ACCEL_X86_MMXEXT	0x20000000
+#define MM_ACCEL_X86_SSE	0x10000000
+
+/* PPC accelerations */
+#define MM_ACCEL_PPC_ALTIVEC	0x00010000

 uint32_t mm_accel (void);

--- liba52/parse.c	2006-12-05 08:08:01.000000000 +0100
+++ liba52/parse.c	2006-12-05 08:08:44.000000000 +0100
@@ -24,6 +28,7 @@
 #include "config.h"

 #include <stdlib.h>
+#include <stdio.h>
 #include <string.h>
 #include <inttypes.h>

@@ -31,13 +36,12 @@
 #include "a52_internal.h"
 #include "bitstream.h"
 #include "tables.h"
+#include "mm_accel.h"
+#include "libavutil/avutil.h"

-#ifdef HAVE_MEMALIGN
+#if HAVE_MEMALIGN
 /* some systems have memalign() but no declaration for it */
 void * memalign (size_t align, size_t size);
-#else
-/* assume malloc alignment is sufficient */
-#define memalign(align,size) malloc (size)
 #endif

 typedef struct {
@@ -60,7 +64,16 @@
     if (state == NULL)
 	return NULL;

+#if defined(__MINGW32__) && defined(HAVE_SSE)
+    state->samples = av_malloc(256 * 12 * sizeof (sample_t));
+#else
     state->samples = memalign (16, 256 * 12 * sizeof (sample_t));
+#endif
+    if(((int)state->samples%16) && (mm_accel&MM_ACCEL_X86_SSE)){
+      mm_accel &=~MM_ACCEL_X86_SSE;
+      fprintf(stderr, "liba52: unable to get 16 byte aligned memory disabling usage of SSE instructions\n");
+    }
+
     if (state->samples == NULL) {
 	free (state);
 	return NULL;
@@ -74,6 +87,7 @@
     state->lfsr_state = 1;

     a52_imdct_init (mm_accel);
+    downmix_accel_init(mm_accel);

     return state;
 }
@@ -141,7 +155,7 @@
     state->acmod = acmod = buf[6] >> 5;

     a52_bitstream_set_ptr (state, buf + 6);
-    bitstream_get (state, 3);	/* skip acmod we already parsed */
+    bitstream_skip (state, 3);	/* skip acmod we already parsed */

     if ((acmod == 2) && (bitstream_get (state, 2) == 2))	/* dsurmod */
 	acmod = A52_DOLBY;
@@ -172,28 +186,28 @@

     chaninfo = !acmod;
     do {
-	bitstream_get (state, 5);	/* dialnorm */
+	bitstream_skip (state, 5);	/* dialnorm */
 	if (bitstream_get (state, 1))	/* compre */
-	    bitstream_get (state, 8);	/* compr */
+	    bitstream_skip (state, 8);	/* compr */
 	if (bitstream_get (state, 1))	/* langcode */
-	    bitstream_get (state, 8);	/* langcod */
+	    bitstream_skip (state, 8);	/* langcod */
 	if (bitstream_get (state, 1))	/* audprodie */
-	    bitstream_get (state, 7);	/* mixlevel + roomtyp */
+	    bitstream_skip (state, 7);	/* mixlevel + roomtyp */
     } while (chaninfo--);

-    bitstream_get (state, 2);		/* copyrightb + origbs */
+    bitstream_skip (state, 2);		/* copyrightb + origbs */

     if (bitstream_get (state, 1))	/* timecod1e */
-	bitstream_get (state, 14);	/* timecod1 */
+	bitstream_skip (state, 14);	/* timecod1 */
     if (bitstream_get (state, 1))	/* timecod2e */
-	bitstream_get (state, 14);	/* timecod2 */
+	bitstream_skip (state, 14);	/* timecod2 */

     if (bitstream_get (state, 1)) {	/* addbsie */
 	int addbsil;

 	addbsil = bitstream_get (state, 6);
 	do {
-	    bitstream_get (state, 8);	/* addbsi */
+	    bitstream_skip (state, 8);	/* addbsi */
 	} while (addbsil--);
     }

@@ -680,7 +694,7 @@
 				 state->fbw_expbap[i].exp[0],
 				 state->fbw_expbap[i].exp + 1))
 		return 1;
-	    bitstream_get (state, 2);	/* gainrng */
+	    bitstream_skip (state, 2);	/* gainrng */
 	}
     if (lfeexpstr != EXP_REUSE) {
 	do_bit_alloc |= 32;
@@ -755,7 +769,7 @@
     if (bitstream_get (state, 1)) {	/* skiple */
 	i = bitstream_get (state, 9);	/* skipl */
 	while (i--)
-	    bitstream_get (state, 8);
+	    bitstream_skip (state, 8);
     }

     samples = state->samples;
@@ -896,6 +910,10 @@

 void a52_free (a52_state_t * state)
 {
-    free (state->samples);
+#if defined(__MINGW32__) && defined(HAVE_SSE)
+    av_free (state->samples);
+#else
+     free (state->samples);
+#endif
     free (state);
 }