summaryrefslogtreecommitdiffstats
path: root/mp3lib
diff options
context:
space:
mode:
authoralex <alex@b3059339-0415-0410-9bf9-f77b7e298cf2>2003-06-22 03:32:16 +0000
committeralex <alex@b3059339-0415-0410-9bf9-f77b7e298cf2>2003-06-22 03:32:16 +0000
commitd0cf347a6269992b0d52d4863c02be91730b1c78 (patch)
tree9b9a9454a2cbc3b6465bd200ab9b3f3fe5663d0c /mp3lib
parent0865846ecb7a6b4e5fbf5127d355892d2ff53d9c (diff)
downloadmpv-d0cf347a6269992b0d52d4863c02be91730b1c78.tar.bz2
mpv-d0cf347a6269992b0d52d4863c02be91730b1c78.tar.xz
moved 3dnow and 3dnowex dct36 optimisations into gcc inline assembly
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@10323 b3059339-0415-0410-9bf9-f77b7e298cf2
Diffstat (limited to 'mp3lib')
-rw-r--r--mp3lib/Makefile4
-rw-r--r--mp3lib/dct36_3dnow.c497
-rw-r--r--mp3lib/dct36_3dnow.s499
-rw-r--r--mp3lib/dct36_k7.c34
-rw-r--r--mp3lib/dct36_k7.s511
5 files changed, 533 insertions, 1012 deletions
diff --git a/mp3lib/Makefile b/mp3lib/Makefile
index 2a8bd4de27..4feaa4bd4d 100644
--- a/mp3lib/Makefile
+++ b/mp3lib/Makefile
@@ -17,9 +17,9 @@ OBJS += decode_MMX.o dct64_MMX.o tabinit_MMX.o
#SRCS += dct64_sse.s
#OBJS += dct64_sse.o
#endif
-SRCS += dct36_3dnow.s dct64_3dnow.c
+SRCS += dct36_3dnow.c dct64_3dnow.c
OBJS += dct36_3dnow.o dct64_3dnow.o
-SRCS += dct36_k7.s dct64_k7.c
+SRCS += dct36_k7.c dct64_k7.c
OBJS += dct36_k7.o dct64_k7.o
endif
ifeq ($(TARGET_ARCH_POWERPC),yes)
diff --git a/mp3lib/dct36_3dnow.c b/mp3lib/dct36_3dnow.c
new file mode 100644
index 0000000000..e62babe3b6
--- /dev/null
+++ b/mp3lib/dct36_3dnow.c
@@ -0,0 +1,497 @@
+/*
+ * dct36_3dnow.c - 3DNow! optimized dct36()
+ *
+ * This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
+ * <squash@mb.kcom.ne.jp>, only two types of changes have been made:
+ *
+ * - removed PREFETCH instruction for speedup
+ * - changed function name for support 3DNow! automatic detection
+ *
+ * You can find Kashiyama's original 3dnow! support patch
+ * (for mpg123-0.59o) at
+ * http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
+ *
+ * by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
+ * <kim@comtec.co.jp> - after 1.Apr.1999
+ *
+ * Original disclaimer:
+ * The author of this program disclaim whole expressed or implied
+ * warranties with regard to this program, and in no event shall the
+ * author of this program liable to whatever resulted from the use of
+ * this program. Use it at your own risk.
+ *
+ * 2003/06/21: Moved to GCC inline assembly - Alex Beregszaszi
+ */
+
+#define real float /* ugly - but only way */
+
+#include "../mangle.h"
+
+#ifdef __DCT36_OPTIMIZE_FOR_K7
+void dct36_3dnowex(real *inbuf, real *o1,
+ real *o2, real *wintab, real *tsbuf)
+#else
+void dct36_3dnow(real *inbuf, real *o1,
+ real *o2, real *wintab, real *tsbuf)
+#endif
+{
+ __asm__ __volatile__(
+ "movq (%%eax),%%mm0\n\t"
+ "movq 4(%%eax),%%mm1\n\t"
+ "pfadd %%mm1,%%mm0\n\t"
+ "movq %%mm0,4(%%eax)\n\t"
+ "psrlq $32,%%mm1\n\t"
+ "movq 12(%%eax),%%mm2\n\t"
+ "punpckldq %%mm2,%%mm1\n\t"
+ "pfadd %%mm2,%%mm1\n\t"
+ "movq %%mm1,12(%%eax)\n\t"
+ "psrlq $32,%%mm2\n\t"
+ "movq 20(%%eax),%%mm3\n\t"
+ "punpckldq %%mm3,%%mm2\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movq %%mm2,20(%%eax)\n\t"
+ "psrlq $32,%%mm3\n\t"
+ "movq 28(%%eax),%%mm4\n\t"
+ "punpckldq %%mm4,%%mm3\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movq %%mm3,28(%%eax)\n\t"
+ "psrlq $32,%%mm4\n\t"
+ "movq 36(%%eax),%%mm5\n\t"
+ "punpckldq %%mm5,%%mm4\n\t"
+ "pfadd %%mm5,%%mm4\n\t"
+ "movq %%mm4,36(%%eax)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movq 44(%%eax),%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movq %%mm5,44(%%eax)\n\t"
+ "psrlq $32,%%mm6\n\t"
+ "movq 52(%%eax),%%mm7\n\t"
+ "punpckldq %%mm7,%%mm6\n\t"
+ "pfadd %%mm7,%%mm6\n\t"
+ "movq %%mm6,52(%%eax)\n\t"
+ "psrlq $32,%%mm7\n\t"
+ "movq 60(%%eax),%%mm0\n\t"
+ "punpckldq %%mm0,%%mm7\n\t"
+ "pfadd %%mm0,%%mm7\n\t"
+ "movq %%mm7,60(%%eax)\n\t"
+ "psrlq $32,%%mm0\n\t"
+ "movd 68(%%eax),%%mm1\n\t"
+ "pfadd %%mm1,%%mm0\n\t"
+ "movd %%mm0,68(%%eax)\n\t"
+ "movd 4(%%eax),%%mm0\n\t"
+ "movd 12(%%eax),%%mm1\n\t"
+ "punpckldq %%mm1,%%mm0\n\t"
+ "punpckldq 20(%%eax),%%mm1\n\t"
+ "pfadd %%mm1,%%mm0\n\t"
+ "movd %%mm0,12(%%eax)\n\t"
+ "psrlq $32,%%mm0\n\t"
+ "movd %%mm0,20(%%eax)\n\t"
+ "psrlq $32,%%mm1\n\t"
+ "movd 28(%%eax),%%mm2\n\t"
+ "punpckldq %%mm2,%%mm1\n\t"
+ "punpckldq 36(%%eax),%%mm2\n\t"
+ "pfadd %%mm2,%%mm1\n\t"
+ "movd %%mm1,28(%%eax)\n\t"
+ "psrlq $32,%%mm1\n\t"
+ "movd %%mm1,36(%%eax)\n\t"
+ "psrlq $32,%%mm2\n\t"
+ "movd 44(%%eax),%%mm3\n\t"
+ "punpckldq %%mm3,%%mm2\n\t"
+ "punpckldq 52(%%eax),%%mm3\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movd %%mm2,44(%%eax)\n\t"
+ "psrlq $32,%%mm2\n\t"
+ "movd %%mm2,52(%%eax)\n\t"
+ "psrlq $32,%%mm3\n\t"
+ "movd 60(%%eax),%%mm4\n\t"
+ "punpckldq %%mm4,%%mm3\n\t"
+ "punpckldq 68(%%eax),%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movd %%mm3,60(%%eax)\n\t"
+ "psrlq $32,%%mm3\n\t"
+ "movd %%mm3,68(%%eax)\n\t"
+
+ "movq 24(%%eax),%%mm0\n\t"
+ "movq 48(%%eax),%%mm1\n\t"
+ "movd "MANGLE(COS9)"+12,%%mm2\n\t"
+ "punpckldq %%mm2,%%mm2\n\t"
+ "movd "MANGLE(COS9)"+24,%%mm3\n\t"
+ "punpckldq %%mm3,%%mm3\n\t"
+ "pfmul %%mm2,%%mm0\n\t"
+ "pfmul %%mm3,%%mm1\n\t"
+ "pushl %%eax\n\t"
+ "movl $1,%%eax\n\t"
+ "movd %%eax,%%mm7\n\t"
+ "pi2fd %%mm7,%%mm7\n\t"
+ "popl %%eax\n\t"
+ "movq 8(%%eax),%%mm2\n\t"
+ "movd "MANGLE(COS9)"+4,%%mm3\n\t"
+ "punpckldq %%mm3,%%mm3\n\t"
+ "pfmul %%mm3,%%mm2\n\t"
+ "pfadd %%mm0,%%mm2\n\t"
+ "movq 40(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+20,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movq 56(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+28,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movq (%%eax),%%mm3\n\t"
+ "movq 16(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+8,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movq 32(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+16,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "pfadd %%mm1,%%mm3\n\t"
+ "movq 64(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+32,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movq %%mm2,%%mm4\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 108(%%edx),%%mm6\n\t"
+ "punpckldq 104(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+#ifdef __DCT36_OPTIMIZE_FOR_K7
+ "pswapd %%mm5,%%mm5\n\t"
+ "movq %%mm5,32(%%ecx)\n\t"
+#else
+ "movd %%mm5,36(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,32(%%ecx)\n\t"
+#endif
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 32(%%edx),%%mm6\n\t"
+ "punpckldq 36(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 32(%%esi),%%mm6\n\t"
+ "punpckldq 36(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,1024(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1152(%%ebx)\n\t"
+ "movq %%mm3,%%mm4\n\t"
+ "pfsub %%mm2,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 140(%%edx),%%mm6\n\t"
+ "punpckldq 72(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,68(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,0(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 0(%%edx),%%mm6\n\t"
+ "punpckldq 68(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 0(%%esi),%%mm6\n\t"
+ "punpckldq 68(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,0(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,2176(%%ebx)\n\t"
+ "movq 8(%%eax),%%mm2\n\t"
+ "movq 40(%%eax),%%mm3\n\t"
+ "pfsub %%mm3,%%mm2\n\t"
+ "movq 56(%%eax),%%mm3\n\t"
+ "pfsub %%mm3,%%mm2\n\t"
+ "movd "MANGLE(COS9)"+12,%%mm3\n\t"
+ "punpckldq %%mm3,%%mm3\n\t"
+ "pfmul %%mm3,%%mm2\n\t"
+ "movq 16(%%eax),%%mm3\n\t"
+ "movq 32(%%eax),%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movq 64(%%eax),%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movd "MANGLE(COS9)"+24,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "movq 48(%%eax),%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movq (%%eax),%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movq %%mm2,%%mm4\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 112(%%edx),%%mm6\n\t"
+ "punpckldq 100(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,40(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,28(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 28(%%edx),%%mm6\n\t"
+ "punpckldq 40(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 28(%%esi),%%mm6\n\t"
+ "punpckldq 40(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,896(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1280(%%ebx)\n\t"
+ "movq %%mm3,%%mm4\n\t"
+ "pfsub %%mm2,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 136(%%edx),%%mm6\n\t"
+ "punpckldq 76(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,64(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,4(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 4(%%edx),%%mm6\n\t"
+ "punpckldq 64(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 4(%%esi),%%mm6\n\t"
+ "punpckldq 64(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,128(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,2048(%%ebx)\n\t"
+
+ "movq 8(%%eax),%%mm2\n\t"
+ "movd "MANGLE(COS9)"+20,%%mm3\n\t"
+ "punpckldq %%mm3,%%mm3\n\t"
+ "pfmul %%mm3,%%mm2\n\t"
+ "pfsub %%mm0,%%mm2\n\t"
+ "movq 40(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+28,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfsub %%mm3,%%mm2\n\t"
+ "movq 56(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+4,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movq (%%eax),%%mm3\n\t"
+ "movq 16(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+32,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movq 32(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+8,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "pfadd %%mm1,%%mm3\n\t"
+ "movq 64(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+16,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "movq %%mm2,%%mm4\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 116(%%edx),%%mm6\n\t"
+ "punpckldq 96(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,44(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,24(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 24(%%edx),%%mm6\n\t"
+ "punpckldq 44(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 24(%%esi),%%mm6\n\t"
+ "punpckldq 44(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,768(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1408(%%ebx)\n\t"
+ "movq %%mm3,%%mm4\n\t"
+ "pfsub %%mm2,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 132(%%edx),%%mm6\n\t"
+ "punpckldq 80(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,60(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,8(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 8(%%edx),%%mm6\n\t"
+ "punpckldq 60(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 8(%%esi),%%mm6\n\t"
+ "punpckldq 60(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,256(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1920(%%ebx)\n\t"
+ "movq 8(%%eax),%%mm2\n\t"
+ "movd "MANGLE(COS9)"+28,%%mm3\n\t"
+ "punpckldq %%mm3,%%mm3\n\t"
+ "pfmul %%mm3,%%mm2\n\t"
+ "pfsub %%mm0,%%mm2\n\t"
+ "movq 40(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+4,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfadd %%mm3,%%mm2\n\t"
+ "movq 56(%%eax),%%mm3\n\t"
+ "movd "MANGLE(COS9)"+20,%%mm4\n\t"
+ "punpckldq %%mm4,%%mm4\n\t"
+ "pfmul %%mm4,%%mm3\n\t"
+ "pfsub %%mm3,%%mm2\n\t"
+ "movq (%%eax),%%mm3\n\t"
+ "movq 16(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+16,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movq 32(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+32,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfadd %%mm4,%%mm3\n\t"
+ "pfadd %%mm1,%%mm3\n\t"
+ "movq 64(%%eax),%%mm4\n\t"
+ "movd "MANGLE(COS9)"+8,%%mm5\n\t"
+ "punpckldq %%mm5,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "pfsub %%mm4,%%mm3\n\t"
+ "movq %%mm2,%%mm4\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 120(%%edx),%%mm6\n\t"
+ "punpckldq 92(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,48(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,20(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 20(%%edx),%%mm6\n\t"
+ "punpckldq 48(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 20(%%esi),%%mm6\n\t"
+ "punpckldq 48(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,640(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1536(%%ebx)\n\t"
+ "movq %%mm3,%%mm4\n\t"
+ "pfsub %%mm2,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 128(%%edx),%%mm6\n\t"
+ "punpckldq 84(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,56(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,12(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 12(%%edx),%%mm6\n\t"
+ "punpckldq 56(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 12(%%esi),%%mm6\n\t"
+ "punpckldq 56(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,384(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1792(%%ebx)\n\t"
+
+ "movq (%%eax),%%mm4\n\t"
+ "movq 16(%%eax),%%mm3\n\t"
+ "pfsub %%mm3,%%mm4\n\t"
+ "movq 32(%%eax),%%mm3\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq 48(%%eax),%%mm3\n\t"
+ "pfsub %%mm3,%%mm4\n\t"
+ "movq 64(%%eax),%%mm3\n\t"
+ "pfadd %%mm3,%%mm4\n\t"
+ "movq %%mm7,%%mm5\n\t"
+ "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t"
+ "pfmul %%mm5,%%mm4\n\t"
+ "movq %%mm4,%%mm5\n\t"
+ "pfacc %%mm5,%%mm5\n\t"
+ "movd 124(%%edx),%%mm6\n\t"
+ "punpckldq 88(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd %%mm5,52(%%ecx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,16(%%ecx)\n\t"
+ "movq %%mm4,%%mm6\n\t"
+ "punpckldq %%mm6,%%mm5\n\t"
+ "pfsub %%mm6,%%mm5\n\t"
+ "punpckhdq %%mm5,%%mm5\n\t"
+ "movd 16(%%edx),%%mm6\n\t"
+ "punpckldq 52(%%edx),%%mm6\n\t"
+ "pfmul %%mm6,%%mm5\n\t"
+ "movd 16(%%esi),%%mm6\n\t"
+ "punpckldq 52(%%esi),%%mm6\n\t"
+ "pfadd %%mm6,%%mm5\n\t"
+ "movd %%mm5,512(%%ebx)\n\t"
+ "psrlq $32,%%mm5\n\t"
+ "movd %%mm5,1664(%%ebx)\n\t"
+
+ "femms\n\t"
+ :
+ : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf)
+ : "memory");
+}
diff --git a/mp3lib/dct36_3dnow.s b/mp3lib/dct36_3dnow.s
deleted file mode 100644
index a729bb4646..0000000000
--- a/mp3lib/dct36_3dnow.s
+++ /dev/null
@@ -1,499 +0,0 @@
-/
-/ dct36_3dnow.s - 3DNow! optimized dct36()
-/
-/ This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
-/ <squash@mb.kcom.ne.jp>,only two types of changes have been made:
-/
-/ - remove PREFETCH instruction for speedup
-/ - change function name for support 3DNow! automatic detect
-/
-/ You can find Kashiyama's original 3dnow! support patch
-/ (for mpg123-0.59o) at
-/ http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
-/
-/ by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
-/ <kim@comtec.co.jp> - after 1.Apr.1999
-/
-
-///
-/// Replacement of dct36() with AMD's 3DNow! SIMD operations support
-///
-/// Syuuhei Kashiyama <squash@mb.kcom.ne.jp>
-///
-/// The author of this program disclaim whole expressed or implied
-/// warranties with regard to this program, and in no event shall the
-/// author of this program liable to whatever resulted from the use of
-/// this program. Use it at your own risk.
-///
-
- .globl dct36_3dnow
- .type dct36_3dnow,@function
-dct36_3dnow:
- pushl %ebp
- movl %esp,%ebp
- subl $120,%esp
- pushl %esi
- pushl %ebx
- movl 8(%ebp),%eax
- movl 12(%ebp),%esi
- movl 16(%ebp),%ecx
- movl 20(%ebp),%edx
- movl 24(%ebp),%ebx
- leal -128(%ebp),%esp
-
- femms
- movq (%eax),%mm0
- movq 4(%eax),%mm1
- pfadd %mm1,%mm0
- movq %mm0,4(%eax)
- psrlq $32,%mm1
- movq 12(%eax),%mm2
- punpckldq %mm2,%mm1
- pfadd %mm2,%mm1
- movq %mm1,12(%eax)
- psrlq $32,%mm2
- movq 20(%eax),%mm3
- punpckldq %mm3,%mm2
- pfadd %mm3,%mm2
- movq %mm2,20(%eax)
- psrlq $32,%mm3
- movq 28(%eax),%mm4
- punpckldq %mm4,%mm3
- pfadd %mm4,%mm3
- movq %mm3,28(%eax)
- psrlq $32,%mm4
- movq 36(%eax),%mm5
- punpckldq %mm5,%mm4
- pfadd %mm5,%mm4
- movq %mm4,36(%eax)
- psrlq $32,%mm5
- movq 44(%eax),%mm6
- punpckldq %mm6,%mm5
- pfadd %mm6,%mm5
- movq %mm5,44(%eax)
- psrlq $32,%mm6
- movq 52(%eax),%mm7
- punpckldq %mm7,%mm6
- pfadd %mm7,%mm6
- movq %mm6,52(%eax)
- psrlq $32,%mm7
- movq 60(%eax),%mm0
- punpckldq %mm0,%mm7
- pfadd %mm0,%mm7
- movq %mm7,60(%eax)
- psrlq $32,%mm0
- movd 68(%eax),%mm1
- pfadd %mm1,%mm0
- movd %mm0,68(%eax)
- movd 4(%eax),%mm0
- movd 12(%eax),%mm1
- punpckldq %mm1,%mm0
- punpckldq 20(%eax),%mm1
- pfadd %mm1,%mm0
- movd %mm0,12(%eax)
- psrlq $32,%mm0
- movd %mm0,20(%eax)
- psrlq $32,%mm1
- movd 28(%eax),%mm2
- punpckldq %mm2,%mm1
- punpckldq 36(%eax),%mm2
- pfadd %mm2,%mm1
- movd %mm1,28(%eax)
- psrlq $32,%mm1
- movd %mm1,36(%eax)
- psrlq $32,%mm2
- movd 44(%eax),%mm3
- punpckldq %mm3,%mm2
- punpckldq 52(%eax),%mm3
- pfadd %mm3,%mm2
- movd %mm2,44(%eax)
- psrlq $32,%mm2
- movd %mm2,52(%eax)
- psrlq $32,%mm3
- movd 60(%eax),%mm4
- punpckldq %mm4,%mm3
- punpckldq 68(%eax),%mm4
- pfadd %mm4,%mm3
- movd %mm3,60(%eax)
- psrlq $32,%mm3
- movd %mm3,68(%eax)
-
- movq 24(%eax),%mm0
- movq 48(%eax),%mm1
- movd COS9+12,%mm2
- punpckldq %mm2,%mm2
- movd COS9+24,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm2,%mm0
- pfmul %mm3,%mm1
- pushl %eax
- movl $1,%eax
- movd %eax,%mm7
- pi2fd %mm7,%mm7
- popl %eax
- movq 8(%eax),%mm2
- movd COS9+4,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfadd %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+20,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+28,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+0,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 108(%edx),%mm6
- punpckldq 104(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,36(%ecx)
- psrlq $32,%mm5
- movd %mm5,32(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 32(%edx),%mm6
- punpckldq 36(%edx),%mm6
- pfmul %mm6,%mm5
- movd 32(%esi),%mm6
- punpckldq 36(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,1024(%ebx)
- psrlq $32,%mm5
- movd %mm5,1152(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+32,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 140(%edx),%mm6
- punpckldq 72(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,68(%ecx)
- psrlq $32,%mm5
- movd %mm5,0(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 0(%edx),%mm6
- punpckldq 68(%edx),%mm6
- pfmul %mm6,%mm5
- movd 0(%esi),%mm6
- punpckldq 68(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,0(%ebx)
- psrlq $32,%mm5
- movd %mm5,2176(%ebx)
- movq 8(%eax),%mm2
- movq 40(%eax),%mm3
- pfsub %mm3,%mm2
- movq 56(%eax),%mm3
- pfsub %mm3,%mm2
- movd COS9+12,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- movq 16(%eax),%mm3
- movq 32(%eax),%mm4
- pfsub %mm4,%mm3
- movq 64(%eax),%mm4
- pfsub %mm4,%mm3
- movd COS9+24,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- movq 48(%eax),%mm4
- pfsub %mm4,%mm3
- movq (%eax),%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+4,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 112(%edx),%mm6
- punpckldq 100(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,40(%ecx)
- psrlq $32,%mm5
- movd %mm5,28(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 28(%edx),%mm6
- punpckldq 40(%edx),%mm6
- pfmul %mm6,%mm5
- movd 28(%esi),%mm6
- punpckldq 40(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,896(%ebx)
- psrlq $32,%mm5
- movd %mm5,1280(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+28,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 136(%edx),%mm6
- punpckldq 76(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,64(%ecx)
- psrlq $32,%mm5
- movd %mm5,4(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 4(%edx),%mm6
- punpckldq 64(%edx),%mm6
- pfmul %mm6,%mm5
- movd 4(%esi),%mm6
- punpckldq 64(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,128(%ebx)
- psrlq $32,%mm5
- movd %mm5,2048(%ebx)
-
- movq 8(%eax),%mm2
- movd COS9+20,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfsub %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+28,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfsub %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+4,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+8,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 116(%edx),%mm6
- punpckldq 96(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,44(%ecx)
- psrlq $32,%mm5
- movd %mm5,24(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 24(%edx),%mm6
- punpckldq 44(%edx),%mm6
- pfmul %mm6,%mm5
- movd 24(%esi),%mm6
- punpckldq 44(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,768(%ebx)
- psrlq $32,%mm5
- movd %mm5,1408(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+24,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 132(%edx),%mm6
- punpckldq 80(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,60(%ecx)
- psrlq $32,%mm5
- movd %mm5,8(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 8(%edx),%mm6
- punpckldq 60(%edx),%mm6
- pfmul %mm6,%mm5
- movd 8(%esi),%mm6
- punpckldq 60(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,256(%ebx)
- psrlq $32,%mm5
- movd %mm5,1920(%ebx)
- movq 8(%eax),%mm2
- movd COS9+28,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfsub %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+4,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+20,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfsub %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+12,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 120(%edx),%mm6
- punpckldq 92(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,48(%ecx)
- psrlq $32,%mm5
- movd %mm5,20(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 20(%edx),%mm6
- punpckldq 48(%edx),%mm6
- pfmul %mm6,%mm5
- movd 20(%esi),%mm6
- punpckldq 48(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,640(%ebx)
- psrlq $32,%mm5
- movd %mm5,1536(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+20,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 128(%edx),%mm6
- punpckldq 84(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,56(%ecx)
- psrlq $32,%mm5
- movd %mm5,12(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 12(%edx),%mm6
- punpckldq 56(%edx),%mm6
- pfmul %mm6,%mm5
- movd 12(%esi),%mm6
- punpckldq 56(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,384(%ebx)
- psrlq $32,%mm5
- movd %mm5,1792(%ebx)
-
- movq (%eax),%mm4
- movq 16(%eax),%mm3
- pfsub %mm3,%mm4
- movq 32(%eax),%mm3
- pfadd %mm3,%mm4
- movq 48(%eax),%mm3
- pfsub %mm3,%mm4
- movq 64(%eax),%mm3
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+16,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 124(%edx),%mm6
- punpckldq 88(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,52(%ecx)
- psrlq $32,%mm5
- movd %mm5,16(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 16(%edx),%mm6
- punpckldq 52(%edx),%mm6
- pfmul %mm6,%mm5
- movd 16(%esi),%mm6
- punpckldq 52(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,512(%ebx)
- psrlq $32,%mm5
- movd %mm5,1664(%ebx)
-
- femms
- popl %ebx
- popl %esi
- movl %ebp,%esp
- popl %ebp
- ret
diff --git a/mp3lib/dct36_k7.c b/mp3lib/dct36_k7.c
new file mode 100644
index 0000000000..c1f19e7c28
--- /dev/null
+++ b/mp3lib/dct36_k7.c
@@ -0,0 +1,34 @@
+/*
+ * dct36_k7.c - 3DNowEx(DSP)! optimized dct36()
+ *
+ * This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
+ * <squash@mb.kcom.ne.jp>, only two types of changes have been made:
+ *
+ * - added new opcode PSWAPD
+ * - removed PREFETCH instruction for speedup
+ * - changed function name for support 3DNowEx! automatic detection
+ *
+ * note: because K7 processors are an aggresive out-of-order three-way
+ * superscalar ones instruction order is not significand for them.
+ *
+ * You can find Kashiyama's original 3dnow! support patch
+ * (for mpg123-0.59o) at
+ * http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
+ *
+ * by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
+ * <kim@comtec.co.jp> - after 1.Apr.1999
+ *
+ * Original disclaimer:
+ * The author of this program disclaim whole expressed or implied
+ * warranties with regard to this program, and in no event shall the
+ * author of this program liable to whatever resulted from the use of
+ * this program. Use it at your own risk.
+ *
+ * Modified by Nick Kurshev <nickols_k@mail.ru>
+ *
+ * 2003/06/21: Moved to GCC inline assembly - Alex Beregszaszi
+ */
+
+#define __DCT36_OPTIMIZE_FOR_K7
+
+#include "dct36_3dnow.c"
diff --git a/mp3lib/dct36_k7.s b/mp3lib/dct36_k7.s
deleted file mode 100644
index 9c6096cec3..0000000000
--- a/mp3lib/dct36_k7.s
+++ /dev/null
@@ -1,511 +0,0 @@
-///
-/// Replacement of dct36() with AMD's 3DNowEx(DSP)! SIMD operations support
-///
-/// This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
-/// <squash@mb.kcom.ne.jp>,only some types of changes have been made:
-///
-/// - added new opcode PSWAPD
-/// - change function name for support 3DNowEx! automatic detect
-///
-/// note: because K7 processors are an aggresive out-of-order three-way
-/// superscalar ones instruction order is not significand for them.
-///
-/// Modified by Nick Kurshev <nickols_k@mail.ru>
-///
-/
-/ dct36_3dnow.s - 3DNow! optimized dct36()
-/
-/ This code based 'dct36_3dnow.s' by Syuuhei Kashiyama
-/ <squash@mb.kcom.ne.jp>,only two types of changes have been made:
-/
-/ - remove PREFETCH instruction for speedup
-/ - change function name for support 3DNow! automatic detect
-/
-/ You can find Kashiyama's original 3dnow! support patch
-/ (for mpg123-0.59o) at
-/ http://user.ecc.u-tokyo.ac.jp/~g810370/linux-simd/ (Japanese).
-/
-/ by KIMURA Takuhiro <kim@hannah.ipc.miyakyo-u.ac.jp> - until 31.Mar.1999
-/ <kim@comtec.co.jp> - after 1.Apr.1999
-/
-
-///
-/// Replacement of dct36() with AMD's 3DNow! SIMD operations support
-///
-/// Syuuhei Kashiyama <squash@mb.kcom.ne.jp>
-///
-/// The author of this program disclaim whole expressed or implied
-/// warranties with regard to this program, and in no event shall the
-/// author of this program liable to whatever resulted from the use of
-/// this program. Use it at your own risk.
-///
-
- .globl dct36_3dnowex
- .type dct36_3dnowex,@function
-dct36_3dnowex:
- pushl %ebp
- movl %esp,%ebp
- subl $120,%esp
- pushl %esi
- pushl %ebx
- movl 8(%ebp),%eax
- movl 12(%ebp),%esi
- movl 16(%ebp),%ecx
- movl 20(%ebp),%edx
- movl 24(%ebp),%ebx
- leal -128(%ebp),%esp
-
- femms
- movq (%eax),%mm0
- movq 4(%eax),%mm1
- pfadd %mm1,%mm0
- movq %mm0,4(%eax)
- psrlq $32,%mm1
- movq 12(%eax),%mm2
- punpckldq %mm2,%mm1
- pfadd %mm2,%mm1
- movq %mm1,12(%eax)
- psrlq $32,%mm2
- movq 20(%eax),%mm3
- punpckldq %mm3,%mm2
- pfadd %mm3,%mm2
- movq %mm2,20(%eax)
- psrlq $32,%mm3
- movq 28(%eax),%mm4
- punpckldq %mm4,%mm3
- pfadd %mm4,%mm3
- movq %mm3,28(%eax)
- psrlq $32,%mm4
- movq 36(%eax),%mm5
- punpckldq %mm5,%mm4
- pfadd %mm5,%mm4
- movq %mm4,36(%eax)
- psrlq $32,%mm5
- movq 44(%eax),%mm6
- punpckldq %mm6,%mm5
- pfadd %mm6,%mm5
- movq %mm5,44(%eax)
- psrlq $32,%mm6
- movq 52(%eax),%mm7
- punpckldq %mm7,%mm6
- pfadd %mm7,%mm6
- movq %mm6,52(%eax)
- psrlq $32,%mm7
- movq 60(%eax),%mm0
- punpckldq %mm0,%mm7
- pfadd %mm0,%mm7
- movq %mm7,60(%eax)
- psrlq $32,%mm0
- movd 68(%eax),%mm1
- pfadd %mm1,%mm0
- movd %mm0,68(%eax)
- movd 4(%eax),%mm0
- movd 12(%eax),%mm1
- punpckldq %mm1,%mm0
- punpckldq 20(%eax),%mm1
- pfadd %mm1,%mm0
- movd %mm0,12(%eax)
- psrlq $32,%mm0
- movd %mm0,20(%eax)
- psrlq $32,%mm1
- movd 28(%eax),%mm2
- punpckldq %mm2,%mm1
- punpckldq 36(%eax),%mm2
- pfadd %mm2,%mm1
- movd %mm1,28(%eax)
- psrlq $32,%mm1
- movd %mm1,36(%eax)
- psrlq $32,%mm2
- movd 44(%eax),%mm3
- punpckldq %mm3,%mm2
- punpckldq 52(%eax),%mm3
- pfadd %mm3,%mm2
- movd %mm2,44(%eax)
- psrlq $32,%mm2
- movd %mm2,52(%eax)
- psrlq $32,%mm3
- movd 60(%eax),%mm4
- punpckldq %mm4,%mm3
- punpckldq 68(%eax),%mm4
- pfadd %mm4,%mm3
- movd %mm3,60(%eax)
- psrlq $32,%mm3
- movd %mm3,68(%eax)
- movq 24(%eax),%mm0
- movq 48(%eax),%mm1
- movd COS9+12,%mm2
- punpckldq %mm2,%mm2
- movd COS9+24,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm2,%mm0
- pfmul %mm3,%mm1
- pushl %eax
- movl $1,%eax
- movd %eax,%mm7
- pi2fd %mm7,%mm7
- popl %eax
- movq 8(%eax),%mm2
- movd COS9+4,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfadd %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+20,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+28,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+0,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 108(%edx),%mm6
- punpckldq 104(%edx),%mm6
- pfmul %mm6,%mm5
- pswapd %mm5, %mm5
- movq %mm5, 32(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 32(%edx),%mm6
- punpckldq 36(%edx),%mm6
- pfmul %mm6,%mm5
- movd 32(%esi),%mm6
- punpckldq 36(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,1024(%ebx)
- psrlq $32,%mm5
- movd %mm5,1152(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+32,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 140(%edx),%mm6
- punpckldq 72(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,68(%ecx)
- psrlq $32,%mm5
- movd %mm5,0(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 0(%edx),%mm6
- punpckldq 68(%edx),%mm6
- pfmul %mm6,%mm5
- movd 0(%esi),%mm6
- punpckldq 68(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,0(%ebx)
- psrlq $32,%mm5
- movd %mm5,2176(%ebx)
- movq 8(%eax),%mm2
- movq 40(%eax),%mm3
- pfsub %mm3,%mm2
- movq 56(%eax),%mm3
- pfsub %mm3,%mm2
- movd COS9+12,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- movq 16(%eax),%mm3
- movq 32(%eax),%mm4
- pfsub %mm4,%mm3
- movq 64(%eax),%mm4
- pfsub %mm4,%mm3
- movd COS9+24,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- movq 48(%eax),%mm4
- pfsub %mm4,%mm3
- movq (%eax),%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+4,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 112(%edx),%mm6
- punpckldq 100(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,40(%ecx)
- psrlq $32,%mm5
- movd %mm5,28(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 28(%edx),%mm6
- punpckldq 40(%edx),%mm6
- pfmul %mm6,%mm5
- movd 28(%esi),%mm6
- punpckldq 40(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,896(%ebx)
- psrlq $32,%mm5
- movd %mm5,1280(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+28,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 136(%edx),%mm6
- punpckldq 76(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,64(%ecx)
- psrlq $32,%mm5
- movd %mm5,4(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 4(%edx),%mm6
- punpckldq 64(%edx),%mm6
- pfmul %mm6,%mm5
- movd 4(%esi),%mm6
- punpckldq 64(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,128(%ebx)
- psrlq $32,%mm5
- movd %mm5,2048(%ebx)
-
- movq 8(%eax),%mm2
- movd COS9+20,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfsub %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+28,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfsub %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+4,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+8,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 116(%edx),%mm6
- punpckldq 96(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,44(%ecx)
- psrlq $32,%mm5
- movd %mm5,24(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 24(%edx),%mm6
- punpckldq 44(%edx),%mm6
- pfmul %mm6,%mm5
- movd 24(%esi),%mm6
- punpckldq 44(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,768(%ebx)
- psrlq $32,%mm5
- movd %mm5,1408(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+24,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 132(%edx),%mm6
- punpckldq 80(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,60(%ecx)
- psrlq $32,%mm5
- movd %mm5,8(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 8(%edx),%mm6
- punpckldq 60(%edx),%mm6
- pfmul %mm6,%mm5
- movd 8(%esi),%mm6
- punpckldq 60(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,256(%ebx)
- psrlq $32,%mm5
- movd %mm5,1920(%ebx)
- movq 8(%eax),%mm2
- movd COS9+28,%mm3
- punpckldq %mm3,%mm3
- pfmul %mm3,%mm2
- pfsub %mm0,%mm2
- movq 40(%eax),%mm3
- movd COS9+4,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfadd %mm3,%mm2
- movq 56(%eax),%mm3
- movd COS9+20,%mm4
- punpckldq %mm4,%mm4
- pfmul %mm4,%mm3
- pfsub %mm3,%mm2
- movq (%eax),%mm3
- movq 16(%eax),%mm4
- movd COS9+16,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq 32(%eax),%mm4
- movd COS9+32,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfadd %mm4,%mm3
- pfadd %mm1,%mm3
- movq 64(%eax),%mm4
- movd COS9+8,%mm5
- punpckldq %mm5,%mm5
- pfmul %mm5,%mm4
- pfsub %mm4,%mm3
- movq %mm2,%mm4
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+12,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 120(%edx),%mm6
- punpckldq 92(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,48(%ecx)
- psrlq $32,%mm5
- movd %mm5,20(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 20(%edx),%mm6
- punpckldq 48(%edx),%mm6
- pfmul %mm6,%mm5
- movd 20(%esi),%mm6
- punpckldq 48(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,640(%ebx)
- psrlq $32,%mm5
- movd %mm5,1536(%ebx)
- movq %mm3,%mm4
- pfsub %mm2,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+20,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 128(%edx),%mm6
- punpckldq 84(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,56(%ecx)
- psrlq $32,%mm5
- movd %mm5,12(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 12(%edx),%mm6
- punpckldq 56(%edx),%mm6
- pfmul %mm6,%mm5
- movd 12(%esi),%mm6
- punpckldq 56(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,384(%ebx)
- psrlq $32,%mm5
- movd %mm5,1792(%ebx)
-
- movq (%eax),%mm4
- movq 16(%eax),%mm3
- pfsub %mm3,%mm4
- movq 32(%eax),%mm3
- pfadd %mm3,%mm4
- movq 48(%eax),%mm3
- pfsub %mm3,%mm4
- movq 64(%eax),%mm3
- pfadd %mm3,%mm4
- movq %mm7,%mm5
- punpckldq tfcos36+16,%mm5
- pfmul %mm5,%mm4
- movq %mm4,%mm5
- pfacc %mm5,%mm5
- movd 124(%edx),%mm6
- punpckldq 88(%edx),%mm6
- pfmul %mm6,%mm5
- movd %mm5,52(%ecx)
- psrlq $32,%mm5
- movd %mm5,16(%ecx)
- movq %mm4,%mm6
- punpckldq %mm6,%mm5
- pfsub %mm6,%mm5
- punpckhdq %mm5,%mm5
- movd 16(%edx),%mm6
- punpckldq 52(%edx),%mm6
- pfmul %mm6,%mm5
- movd 16(%esi),%mm6
- punpckldq 52(%esi),%mm6
- pfadd %mm6,%mm5
- movd %mm5,512(%ebx)
- psrlq $32,%mm5
- movd %mm5,1664(%ebx)
-
- femms
- popl %ebx
- popl %esi
- movl %ebp,%esp
- popl %ebp
- ret