Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)

Originally committed as revision 1147 to svn://svn.ffmpeg.org/ffmpeg/trunk

6 files changed, 972 insertions, 7 deletions

diff --git a/libavcodec/ppc/dsputil_altivec.c b/libavcodec/ppc/dsputil_altivec.c
index 18d9d27a48..8a50ccb900 100644
--- a/libavcodec/ppc/dsputil_altivec.c
+++ b/libavcodec/ppc/dsputil_altivec.c
@@ -1,15 +1,29 @@
+/*
+ * Copyright (c) 2002 Brian Foley
+ * Copyright (c) 2002 Dieter Shirley
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+ 
 #include "../dsputil.h"
+#include "dsputil_altivec.h"
 
 #if CONFIG_DARWIN
 #include <sys/sysctl.h>
 #endif
 
-int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
-int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
-int pix_sum_altivec(UINT8 * pix, int line_size);
-
-int has_altivec(void);
-
 int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size)
 {
     int i, s;
@@ -127,6 +141,105 @@ int pix_sum_altivec(UINT8 * pix, int line_size)
     return s;
 }
 
+void get_pixels_altivec(DCTELEM *restrict block, const UINT8 *pixels, int line_size)
+{
+    int i;
+    vector unsigned char perm, bytes, *pixv;
+    vector unsigned char zero = (vector unsigned char) (0);
+    vector signed short shorts;
+
+    for(i=0;i<8;i++)
+    {
+        // Read potentially unaligned pixels.
+        // We're reading 16 pixels, and actually only want 8,
+        // but we simply ignore the extras.
+        perm = vec_lvsl(0, pixels);
+        pixv = (vector unsigned char *) pixels;
+        bytes = vec_perm(pixv[0], pixv[1], perm);
+
+        // convert the bytes into shorts
+        shorts = (vector signed short)vec_mergeh(zero, bytes);
+
+        // save the data to the block, we assume the block is 16-byte aligned
+        vec_st(shorts, i*16, (vector signed short*)block);
+
+        pixels += line_size;
+    }
+}
+
+void diff_pixels_altivec(DCTELEM *restrict block, const UINT8 *s1,
+        const UINT8 *s2, int stride)
+{
+    int i;
+    vector unsigned char perm, bytes, *pixv;
+    vector unsigned char zero = (vector unsigned char) (0);
+    vector signed short shorts1, shorts2;
+
+    for(i=0;i<4;i++)
+    {
+        // Read potentially unaligned pixels
+        // We're reading 16 pixels, and actually only want 8,
+        // but we simply ignore the extras.
+        perm = vec_lvsl(0, s1);
+        pixv = (vector unsigned char *) s1;
+        bytes = vec_perm(pixv[0], pixv[1], perm);
+
+        // convert the bytes into shorts
+        shorts1 = (vector signed short)vec_mergeh(zero, bytes);
+
+        // Do the same for the second block of pixels
+        perm = vec_lvsl(0, s2);
+        pixv = (vector unsigned char *) s2;
+        bytes = vec_perm(pixv[0], pixv[1], perm);
+
+        // convert the bytes into shorts
+        shorts2 = (vector signed short)vec_mergeh(zero, bytes);
+
+        // Do the subtraction
+        shorts1 = vec_sub(shorts1, shorts2);
+
+        // save the data to the block, we assume the block is 16-byte aligned
+        vec_st(shorts1, 0, (vector signed short*)block);
+
+        s1 += stride;
+        s2 += stride;
+        block += 8;
+
+
+        // The code below is a copy of the code above... This is a manual
+        // unroll.
+
+        // Read potentially unaligned pixels
+        // We're reading 16 pixels, and actually only want 8,
+        // but we simply ignore the extras.
+        perm = vec_lvsl(0, s1);
+        pixv = (vector unsigned char *) s1;
+        bytes = vec_perm(pixv[0], pixv[1], perm);
+
+        // convert the bytes into shorts
+        shorts1 = (vector signed short)vec_mergeh(zero, bytes);
+
+        // Do the same for the second block of pixels
+        perm = vec_lvsl(0, s2);
+        pixv = (vector unsigned char *) s2;
+        bytes = vec_perm(pixv[0], pixv[1], perm);
+
+        // convert the bytes into shorts
+        shorts2 = (vector signed short)vec_mergeh(zero, bytes);
+
+        // Do the subtraction
+        shorts1 = vec_sub(shorts1, shorts2);
+
+        // save the data to the block, we assume the block is 16-byte aligned
+        vec_st(shorts1, 0, (vector signed short*)block);
+
+        s1 += stride;
+        s2 += stride;
+        block += 8;
+    }
+}
+
+
 int has_altivec(void)
 {
 #if CONFIG_DARWIN
@@ -141,3 +254,4 @@ int has_altivec(void)
 #endif
     return 0;
 }
+
diff --git a/libavcodec/ppc/dsputil_altivec.h b/libavcodec/ppc/dsputil_altivec.h
index 42c373e76b..bdf8f5ffee 100644
--- a/libavcodec/ppc/dsputil_altivec.h
+++ b/libavcodec/ppc/dsputil_altivec.h
@@ -1,5 +1,26 @@
+/*
+ * Copyright (c) 2002 Brian Foley
+ * Copyright (c) 2002 Dieter Shirley
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+ 
 extern int pix_abs16x16_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_abs8x8_altivec(uint8_t *pix1, uint8_t *pix2, int line_size);
 extern int pix_sum_altivec(UINT8 * pix, int line_size);
+extern void diff_pixels_altivec(DCTELEM* block, const UINT8* s1, const UINT8* s2, int stride);
+extern void get_pixels_altivec(DCTELEM* block, const UINT8 * pixels, int line_size);
 
 extern int has_altivec(void);
diff --git a/libavcodec/ppc/dsputil_ppc.c b/libavcodec/ppc/dsputil_ppc.c
index d9c8ba86e0..28c56ee94a 100644
--- a/libavcodec/ppc/dsputil_ppc.c
+++ b/libavcodec/ppc/dsputil_ppc.c
@@ -1,3 +1,22 @@
+/*
+ * Copyright (c) 2002 Brian Foley
+ * Copyright (c) 2002 Dieter Shirley
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+
 #include "../dsputil.h"
 
 #ifdef HAVE_ALTIVEC
@@ -6,14 +25,23 @@
 
 void dsputil_init_ppc(void)
 {
+    // Common optimisations whether Altivec or not
+
+    // ... pending ...
+
 #if HAVE_ALTIVEC
     if (has_altivec()) {
+        // Altivec specific optimisations
         pix_abs16x16 = pix_abs16x16_altivec;
         pix_abs8x8 = pix_abs8x8_altivec;
         pix_sum = pix_sum_altivec;
+        diff_pixels = diff_pixels_altivec;
+        get_pixels = get_pixels_altivec;
     } else
 #endif
     {
-        /* Non-AltiVec PPC optimisations here */
+        // Non-AltiVec PPC optimisations
+
+        // ... pending ...
     }
 }
diff --git a/libavcodec/ppc/idct_altivec.c b/libavcodec/ppc/idct_altivec.c
new file mode 100644
index 0000000000..8036d403fa
--- /dev/null
+++ b/libavcodec/ppc/idct_altivec.c
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2001 Michel Lespinasse
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *
+ */
+
+/*
+ * NOTE: This code is based on GPL code from the libmpeg2 project.  The
+ * author, Michel Lespinasses, has given explicit permission to release
+ * under LGPL as part of ffmpeg.
+ *
+ */
+
+/*
+ * FFMpeg integration by Dieter Shirley
+ *
+ * This file is a direct copy of the altivec idct module from the libmpeg2
+ * project.  I've deleted all of the libmpeg2 specific code, renamed the functions and
+ * re-ordered the function parameters.  The only change to the IDCT function
+ * itself was to factor out the partial transposition, and to perform a full
+ * transpose at the end of the function.
+ */
+
+
+#include <stdlib.h>                                      /* malloc(), free() */
+#include <string.h>
+#include "../dsputil.h"
+
+#define vector_s16_t vector signed short
+#define vector_u16_t vector unsigned short
+#define vector_s8_t vector signed char
+#define vector_u8_t vector unsigned char
+#define vector_s32_t vector signed int
+#define vector_u32_t vector unsigned int
+
+#define IDCT_HALF					\
+    /* 1st stage */					\
+    t1 = vec_mradds (a1, vx7, vx1 );			\
+    t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7));	\
+    t7 = vec_mradds (a2, vx5, vx3);			\
+    t3 = vec_mradds (ma2, vx3, vx5);			\
+							\
+    /* 2nd stage */					\
+    t5 = vec_adds (vx0, vx4);				\
+    t0 = vec_subs (vx0, vx4);				\
+    t2 = vec_mradds (a0, vx6, vx2);			\
+    t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6));	\
+    t6 = vec_adds (t8, t3);				\
+    t3 = vec_subs (t8, t3);				\
+    t8 = vec_subs (t1, t7);				\
+    t1 = vec_adds (t1, t7);				\
+							\
+    /* 3rd stage */					\
+    t7 = vec_adds (t5, t2);				\
+    t2 = vec_subs (t5, t2);				\
+    t5 = vec_adds (t0, t4);				\
+    t0 = vec_subs (t0, t4);				\
+    t4 = vec_subs (t8, t3);				\
+    t3 = vec_adds (t8, t3);				\
+							\
+    /* 4th stage */					\
+    vy0 = vec_adds (t7, t1);				\
+    vy7 = vec_subs (t7, t1);				\
+    vy1 = vec_mradds (c4, t3, t5);			\
+    vy6 = vec_mradds (mc4, t3, t5);			\
+    vy2 = vec_mradds (c4, t4, t0);			\
+    vy5 = vec_mradds (mc4, t4, t0);			\
+    vy3 = vec_adds (t2, t6);				\
+    vy4 = vec_subs (t2, t6);
+
+	
+#define IDCT								\
+    vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7;		\
+    vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7;		\
+    vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias;			\
+    vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8;			\
+    vector_u16_t shift;							\
+									\
+    c4 = vec_splat (constants[0], 0);					\
+    a0 = vec_splat (constants[0], 1);					\
+    a1 = vec_splat (constants[0], 2);					\
+    a2 = vec_splat (constants[0], 3);					\
+    mc4 = vec_splat (constants[0], 4);					\
+    ma2 = vec_splat (constants[0], 5);					\
+    bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3);	\
+									\
+    zero = vec_splat_s16 (0);						\
+    shift = vec_splat_u16 (4);						\
+									\
+    vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero);	\
+    vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero);	\
+    vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero);	\
+    vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero);	\
+    vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero);	\
+    vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero);	\
+    vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero);	\
+    vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero);	\
+									\
+    IDCT_HALF								\
+									\
+    vx0 = vec_mergeh (vy0, vy4);					\
+    vx1 = vec_mergel (vy0, vy4);					\
+    vx2 = vec_mergeh (vy1, vy5);					\
+    vx3 = vec_mergel (vy1, vy5);					\
+    vx4 = vec_mergeh (vy2, vy6);					\
+    vx5 = vec_mergel (vy2, vy6);					\
+    vx6 = vec_mergeh (vy3, vy7);					\
+    vx7 = vec_mergel (vy3, vy7);					\
+									\
+    vy0 = vec_mergeh (vx0, vx4);					\
+    vy1 = vec_mergel (vx0, vx4);					\
+    vy2 = vec_mergeh (vx1, vx5);					\
+    vy3 = vec_mergel (vx1, vx5);					\
+    vy4 = vec_mergeh (vx2, vx6);					\
+    vy5 = vec_mergel (vx2, vx6);					\
+    vy6 = vec_mergeh (vx3, vx7);					\
+    vy7 = vec_mergel (vx3, vx7);					\
+									\
+    vx0 = vec_adds (vec_mergeh (vy0, vy4), bias);			\
+    vx1 = vec_mergel (vy0, vy4);					\
+    vx2 = vec_mergeh (vy1, vy5);					\
+    vx3 = vec_mergel (vy1, vy5);					\
+    vx4 = vec_mergeh (vy2, vy6);					\
+    vx5 = vec_mergel (vy2, vy6);					\
+    vx6 = vec_mergeh (vy3, vy7);					\
+    vx7 = vec_mergel (vy3, vy7);					\
+									\
+    IDCT_HALF								\
+									\
+    shift = vec_splat_u16 (6);						\
+    vx0 = vec_sra (vy0, shift);						\
+    vx1 = vec_sra (vy1, shift);						\
+    vx2 = vec_sra (vy2, shift);						\
+    vx3 = vec_sra (vy3, shift);						\
+    vx4 = vec_sra (vy4, shift);						\
+    vx5 = vec_sra (vy5, shift);						\
+    vx6 = vec_sra (vy6, shift);						\
+    vx7 = vec_sra (vy7, shift);
+
+static const vector_s16_t constants[5] = {
+    (vector_s16_t)(23170, 13573, 6518, 21895, -23170, -21895, 32, 31),
+    (vector_s16_t)(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725),
+    (vector_s16_t)(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521),
+    (vector_s16_t)(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692),
+    (vector_s16_t)(19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722)
+};
+
+void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block)
+{
+    vector_u8_t tmp;
+
+    IDCT
+
+#define COPY(dest,src)						\
+    tmp = vec_packsu (src, src);				\
+    vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest);	\
+    vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
+
+    COPY (dest, vx0)	dest += stride;
+    COPY (dest, vx1)	dest += stride;
+    COPY (dest, vx2)	dest += stride;
+    COPY (dest, vx3)	dest += stride;
+    COPY (dest, vx4)	dest += stride;
+    COPY (dest, vx5)	dest += stride;
+    COPY (dest, vx6)	dest += stride;
+    COPY (dest, vx7)
+}
+
+void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block)
+{
+    vector_u8_t tmp;
+    vector_s16_t tmp2, tmp3;
+    vector_u8_t perm0;
+    vector_u8_t perm1;
+    vector_u8_t p0, p1, p;
+
+    IDCT
+
+    p0 = vec_lvsl (0, dest);
+    p1 = vec_lvsl (stride, dest);
+    p = vec_splat_u8 (-1);
+    perm0 = vec_mergeh (p, p0);
+    perm1 = vec_mergeh (p, p1);
+
+#define ADD(dest,src,perm)						\
+    /* *(uint64_t *)&tmp = *(uint64_t *)dest; */			\
+    tmp = vec_ld (0, dest);						\
+    tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm);	\
+    tmp3 = vec_adds (tmp2, src);					\
+    tmp = vec_packsu (tmp3, tmp3);					\
+    vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest);		\
+    vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
+
+    ADD (dest, vx0, perm0)	dest += stride;
+    ADD (dest, vx1, perm1)	dest += stride;
+    ADD (dest, vx2, perm0)	dest += stride;
+    ADD (dest, vx3, perm1)	dest += stride;
+    ADD (dest, vx4, perm0)	dest += stride;
+    ADD (dest, vx5, perm1)	dest += stride;
+    ADD (dest, vx6, perm0)	dest += stride;
+    ADD (dest, vx7, perm1)
+}
+
diff --git a/libavcodec/ppc/mpegvideo_altivec.c b/libavcodec/ppc/mpegvideo_altivec.c
new file mode 100644
index 0000000000..bcbc1e6baf
--- /dev/null
+++ b/libavcodec/ppc/mpegvideo_altivec.c
@@ -0,0 +1,509 @@
+/*
+ * Copyright (c) 2002 Dieter Shirley
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+ 
+#include <stdlib.h>
+#include <stdio.h>
+#include "../dsputil.h"
+#include "../mpegvideo.h"
+
+
+// Used when initializing constant vectors
+#define FOUR_INSTANCES(x) x,x,x,x
+
+// Swaps two variables (used for altivec registers)
+#define SWAP(a,b) \
+do { \
+    __typeof__(a) swap_temp=a; \
+    a=b; \
+    b=swap_temp; \
+} while (0)
+
+// transposes a matrix consisting of four vectors with four elements each
+#define TRANSPOSE4(a,b,c,d) \
+do { \
+  __typeof__(a) _trans_ach = vec_mergeh(a, c); \
+  __typeof__(a) _trans_acl = vec_mergel(a, c); \
+  __typeof__(a) _trans_bdh = vec_mergeh(b, d); \
+  __typeof__(a) _trans_bdl = vec_mergel(b, d); \
+ \
+  a = vec_mergeh(_trans_ach, _trans_bdh); \
+  b = vec_mergel(_trans_ach, _trans_bdh); \
+  c = vec_mergeh(_trans_acl, _trans_bdl); \
+  d = vec_mergel(_trans_acl, _trans_bdl); \
+} while (0)
+
+#define TRANSPOSE8(a,b,c,d,e,f,g,h) \
+do { \
+    __typeof__(a)  _A1, _B1, _C1, _D1, _E1, _F1, _G1, _H1; \
+    __typeof__(a)  _A2, _B2, _C2, _D2, _E2, _F2, _G2, _H2; \
+ \
+    _A1 = vec_mergeh (a, e); \
+    _B1 = vec_mergel (a, e); \
+    _C1 = vec_mergeh (b, f); \
+    _D1 = vec_mergel (b, f); \
+    _E1 = vec_mergeh (c, g); \
+    _F1 = vec_mergel (c, g); \
+    _G1 = vec_mergeh (d, h); \
+    _H1 = vec_mergel (d, h); \
+ \
+    _A2 = vec_mergeh (_A1, _E1); \
+    _B2 = vec_mergel (_A1, _E1); \
+    _C2 = vec_mergeh (_B1, _F1); \
+    _D2 = vec_mergel (_B1, _F1); \
+    _E2 = vec_mergeh (_C1, _G1); \
+    _F2 = vec_mergel (_C1, _G1); \
+    _G2 = vec_mergeh (_D1, _H1); \
+    _H2 = vec_mergel (_D1, _H1); \
+ \
+    a = vec_mergeh (_A2, _E2); \
+    b = vec_mergel (_A2, _E2); \
+    c = vec_mergeh (_B2, _F2); \
+    d = vec_mergel (_B2, _F2); \
+    e = vec_mergeh (_C2, _G2); \
+    f = vec_mergel (_C2, _G2); \
+    g = vec_mergeh (_D2, _H2); \
+    h = vec_mergel (_D2, _H2); \
+} while (0)
+
+
+// Loads a four-byte value (int or float) from the target address
+// into every element in the target vector.  Only works if the
+// target address is four-byte aligned (which should be always).
+#define LOAD4(vec, address) \
+{ \
+    __typeof__(vec)* _load_addr = (__typeof__(vec)*)(address); \
+    vector unsigned char _perm_vec = vec_lvsl(0,(address)); \
+    vec = vec_ld(0, _load_addr); \
+    vec = vec_perm(vec, vec, _perm_vec); \
+    vec = vec_splat(vec, 0); \
+}
+
+int dct_quantize_altivec(MpegEncContext* s, 
+                        DCTELEM* data, int n,
+                        int qscale, int* overflow)
+{
+    int lastNonZero;
+    vector float row0, row1, row2, row3, row4, row5, row6, row7;
+    vector float alt0, alt1, alt2, alt3, alt4, alt5, alt6, alt7;
+    const vector float zero = {FOUR_INSTANCES(0.0f)};
+
+    // Load the data into the row/alt vectors
+    {
+        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
+
+        data0 = vec_ld(0, data);
+        data1 = vec_ld(16, data);
+        data2 = vec_ld(32, data);
+        data3 = vec_ld(48, data);
+        data4 = vec_ld(64, data);
+        data5 = vec_ld(80, data);
+        data6 = vec_ld(96, data);
+        data7 = vec_ld(112, data);
+
+        // Transpose the data before we start
+        TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
+
+        // load the data into floating point vectors.  We load
+        // the high half of each row into the main row vectors
+        // and the low half into the alt vectors.
+        row0 = vec_ctf(vec_unpackh(data0), 0);
+        alt0 = vec_ctf(vec_unpackl(data0), 0);
+        row1 = vec_ctf(vec_unpackh(data1), 0);
+        alt1 = vec_ctf(vec_unpackl(data1), 0);
+        row2 = vec_ctf(vec_unpackh(data2), 0);
+        alt2 = vec_ctf(vec_unpackl(data2), 0);
+        row3 = vec_ctf(vec_unpackh(data3), 0);
+        alt3 = vec_ctf(vec_unpackl(data3), 0);
+        row4 = vec_ctf(vec_unpackh(data4), 0);
+        alt4 = vec_ctf(vec_unpackl(data4), 0);
+        row5 = vec_ctf(vec_unpackh(data5), 0);
+        alt5 = vec_ctf(vec_unpackl(data5), 0);
+        row6 = vec_ctf(vec_unpackh(data6), 0);
+        alt6 = vec_ctf(vec_unpackl(data6), 0);
+        row7 = vec_ctf(vec_unpackh(data7), 0);
+        alt7 = vec_ctf(vec_unpackl(data7), 0);
+    }
+
+    // The following block could exist as a separate an altivec dct
+		// function.  However, if we put it inline, the DCT data can remain
+		// in the vector local variables, as floats, which we'll use during the
+		// quantize step...
+    {
+        const vector float vec_0_298631336 = {FOUR_INSTANCES(0.298631336f)};
+        const vector float vec_0_390180644 = {FOUR_INSTANCES(-0.390180644f)};
+        const vector float vec_0_541196100 = {FOUR_INSTANCES(0.541196100f)};
+        const vector float vec_0_765366865 = {FOUR_INSTANCES(0.765366865f)};
+        const vector float vec_0_899976223 = {FOUR_INSTANCES(-0.899976223f)};
+        const vector float vec_1_175875602 = {FOUR_INSTANCES(1.175875602f)};
+        const vector float vec_1_501321110 = {FOUR_INSTANCES(1.501321110f)};
+        const vector float vec_1_847759065 = {FOUR_INSTANCES(-1.847759065f)};
+        const vector float vec_1_961570560 = {FOUR_INSTANCES(-1.961570560f)};
+        const vector float vec_2_053119869 = {FOUR_INSTANCES(2.053119869f)};
+        const vector float vec_2_562915447 = {FOUR_INSTANCES(-2.562915447f)};
+        const vector float vec_3_072711026 = {FOUR_INSTANCES(3.072711026f)};
+
+
+        int whichPass, whichHalf;
+
+        for(whichPass = 1; whichPass<=2; whichPass++)
+        {
+            for(whichHalf = 1; whichHalf<=2; whichHalf++)
+            {
+                vector float tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+                vector float tmp10, tmp11, tmp12, tmp13;
+                vector float z1, z2, z3, z4, z5;
+
+                tmp0 = vec_add(row0, row7); // tmp0 = dataptr[0] + dataptr[7];
+                tmp7 = vec_sub(row0, row7); // tmp7 = dataptr[0] - dataptr[7];
+                tmp3 = vec_add(row3, row4); // tmp3 = dataptr[3] + dataptr[4];
+                tmp4 = vec_sub(row3, row4); // tmp4 = dataptr[3] - dataptr[4];
+                tmp1 = vec_add(row1, row6); // tmp1 = dataptr[1] + dataptr[6];
+                tmp6 = vec_sub(row1, row6); // tmp6 = dataptr[1] - dataptr[6];
+                tmp2 = vec_add(row2, row5); // tmp2 = dataptr[2] + dataptr[5];
+                tmp5 = vec_sub(row2, row5); // tmp5 = dataptr[2] - dataptr[5];
+
+                tmp10 = vec_add(tmp0, tmp3); // tmp10 = tmp0 + tmp3;
+                tmp13 = vec_sub(tmp0, tmp3); // tmp13 = tmp0 - tmp3;
+                tmp11 = vec_add(tmp1, tmp2); // tmp11 = tmp1 + tmp2;
+                tmp12 = vec_sub(tmp1, tmp2); // tmp12 = tmp1 - tmp2;
+
+
+                // dataptr[0] = (DCTELEM) ((tmp10 + tmp11) << PASS1_BITS);
+                row0 = vec_add(tmp10, tmp11);
+
+                // dataptr[4] = (DCTELEM) ((tmp10 - tmp11) << PASS1_BITS);
+                row4 = vec_sub(tmp10, tmp11);
+
+
+                // z1 = MULTIPLY(tmp12 + tmp13, FIX_0_541196100);
+                z1 = vec_madd(vec_add(tmp12, tmp13), vec_0_541196100, (vector float)zero);
+
+                // dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, FIX_0_765366865),
+                //		   CONST_BITS-PASS1_BITS);
+                row2 = vec_madd(tmp13, vec_0_765366865, z1);
+
+                // dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, - FIX_1_847759065),
+                //		   CONST_BITS-PASS1_BITS);
+                row6 = vec_madd(tmp12, vec_1_847759065, z1);
+
+                z1 = vec_add(tmp4, tmp7); // z1 = tmp4 + tmp7;
+                z2 = vec_add(tmp5, tmp6); // z2 = tmp5 + tmp6;
+                z3 = vec_add(tmp4, tmp6); // z3 = tmp4 + tmp6;
+                z4 = vec_add(tmp5, tmp7); // z4 = tmp5 + tmp7;
+
+                // z5 = MULTIPLY(z3 + z4, FIX_1_175875602); /* sqrt(2) * c3 */
+                z5 = vec_madd(vec_add(z3, z4), vec_1_175875602, (vector float)zero);
+
+                // z3 = MULTIPLY(z3, - FIX_1_961570560); /* sqrt(2) * (-c3-c5) */
+                z3 = vec_madd(z3, vec_1_961570560, z5);
+
+                // z4 = MULTIPLY(z4, - FIX_0_390180644); /* sqrt(2) * (c5-c3) */
+                z4 = vec_madd(z4, vec_0_390180644, z5);
+
+                // The following adds are rolled into the multiplies above
+                // z3 = vec_add(z3, z5);  // z3 += z5;
+                // z4 = vec_add(z4, z5);  // z4 += z5;
+
+                // z2 = MULTIPLY(z2, - FIX_2_562915447); /* sqrt(2) * (-c1-c3) */
+                // Wow!  It's actually more effecient to roll this multiply
+                // into the adds below, even thought the multiply gets done twice!
+                // z2 = vec_madd(z2, vec_2_562915447, (vector float)zero);
+
+                // z1 = MULTIPLY(z1, - FIX_0_899976223); /* sqrt(2) * (c7-c3) */
+                // Same with this one...
+                // z1 = vec_madd(z1, vec_0_899976223, (vector float)zero);
+
+                // tmp4 = MULTIPLY(tmp4, FIX_0_298631336); /* sqrt(2) * (-c1+c3+c5-c7) */
+                // dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS);
+                row7 = vec_madd(tmp4, vec_0_298631336, vec_madd(z1, vec_0_899976223, z3));
+
+                // tmp5 = MULTIPLY(tmp5, FIX_2_053119869); /* sqrt(2) * ( c1+c3-c5+c7) */
+                // dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS);
+                row5 = vec_madd(tmp5, vec_2_053119869, vec_madd(z2, vec_2_562915447, z4));
+
+                // tmp6 = MULTIPLY(tmp6, FIX_3_072711026); /* sqrt(2) * ( c1+c3+c5-c7) */
+                // dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS);
+                row3 = vec_madd(tmp6, vec_3_072711026, vec_madd(z2, vec_2_562915447, z3));
+
+                // tmp7 = MULTIPLY(tmp7, FIX_1_501321110); /* sqrt(2) * ( c1+c3-c5-c7) */
+                // dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS);
+                row1 = vec_madd(z1, vec_0_899976223, vec_madd(tmp7, vec_1_501321110, z4));
+
+                // Swap the row values with the alts.  If this is the first half,
+                // this sets up the low values to be acted on in the second half.
+                // If this is the second half, it puts the high values back in
+                // the row values where they are expected to be when we're done.
+                SWAP(row0, alt0);
+                SWAP(row1, alt1);
+                SWAP(row2, alt2);
+                SWAP(row3, alt3);
+                SWAP(row4, alt4);
+                SWAP(row5, alt5);
+                SWAP(row6, alt6);
+                SWAP(row7, alt7);
+            }
+
+            if (whichPass == 1)
+            {
+                // transpose the data for the second pass
+                 
+                // First, block transpose the upper right with lower left.
+                SWAP(row4, alt0);
+                SWAP(row5, alt1);
+                SWAP(row6, alt2);
+                SWAP(row7, alt3);
+
+                // Now, transpose each block of four
+                TRANSPOSE4(row0, row1, row2, row3);
+                TRANSPOSE4(row4, row5, row6, row7);
+                TRANSPOSE4(alt0, alt1, alt2, alt3);
+                TRANSPOSE4(alt4, alt5, alt6, alt7);
+            }
+        }
+    }
+
+    // used after quantise step
+    int oldBaseValue = 0;
+
+    // perform the quantise step, using the floating point data
+    // still in the row/alt registers
+    {
+        const int* biasAddr;
+        const vector signed int* qmat;
+        vector float bias, negBias;
+
+        if (s->mb_intra)
+        {
+            vector signed int baseVector;
+
+            // We must cache element 0 in the intra case
+            // (it needs special handling).
+            baseVector = vec_cts(vec_splat(row0, 0), 0);
+            vec_ste(baseVector, 0, &oldBaseValue);
+
+            qmat = (vector signed int*)s->q_intra_matrix[qscale];
+            biasAddr = &(s->intra_quant_bias);
+        }
+        else
+        {
+            qmat = (vector signed int*)s->q_inter_matrix[qscale];
+            biasAddr = &(s->inter_quant_bias);
+        }
+
+        // Load the bias vector (We add 0.5 to the bias so that we're
+				// rounding when we convert to int, instead of flooring.)
+        {
+            vector signed int biasInt;
+            const vector float negOneFloat = (vector float)(FOUR_INSTANCES(-1.0f));
+            LOAD4(biasInt, biasAddr);
+            bias = vec_ctf(biasInt, QUANT_BIAS_SHIFT);
+            negBias = vec_madd(bias, negOneFloat, zero);
+        }
+
+        {
+            vector float q0, q1, q2, q3, q4, q5, q6, q7;
+
+            q0 = vec_ctf(qmat[0], QMAT_SHIFT);
+            q1 = vec_ctf(qmat[2], QMAT_SHIFT);
+            q2 = vec_ctf(qmat[4], QMAT_SHIFT);
+            q3 = vec_ctf(qmat[6], QMAT_SHIFT);
+            q4 = vec_ctf(qmat[8], QMAT_SHIFT);
+            q5 = vec_ctf(qmat[10], QMAT_SHIFT);
+            q6 = vec_ctf(qmat[12], QMAT_SHIFT);
+            q7 = vec_ctf(qmat[14], QMAT_SHIFT);
+
+            row0 = vec_sel(vec_madd(row0, q0, negBias), vec_madd(row0, q0, bias),
+                    vec_cmpgt(row0, zero));
+            row1 = vec_sel(vec_madd(row1, q1, negBias), vec_madd(row1, q1, bias),
+                    vec_cmpgt(row1, zero));
+            row2 = vec_sel(vec_madd(row2, q2, negBias), vec_madd(row2, q2, bias),
+                    vec_cmpgt(row2, zero));
+            row3 = vec_sel(vec_madd(row3, q3, negBias), vec_madd(row3, q3, bias),
+                    vec_cmpgt(row3, zero));
+            row4 = vec_sel(vec_madd(row4, q4, negBias), vec_madd(row4, q4, bias),
+                    vec_cmpgt(row4, zero));
+            row5 = vec_sel(vec_madd(row5, q5, negBias), vec_madd(row5, q5, bias),
+                    vec_cmpgt(row5, zero));
+            row6 = vec_sel(vec_madd(row6, q6, negBias), vec_madd(row6, q6, bias),
+                    vec_cmpgt(row6, zero));
+            row7 = vec_sel(vec_madd(row7, q7, negBias), vec_madd(row7, q7, bias),
+                    vec_cmpgt(row7, zero));
+
+            q0 = vec_ctf(qmat[1], QMAT_SHIFT);
+            q1 = vec_ctf(qmat[3], QMAT_SHIFT);
+            q2 = vec_ctf(qmat[5], QMAT_SHIFT);
+            q3 = vec_ctf(qmat[7], QMAT_SHIFT);
+            q4 = vec_ctf(qmat[9], QMAT_SHIFT);
+            q5 = vec_ctf(qmat[11], QMAT_SHIFT);
+            q6 = vec_ctf(qmat[13], QMAT_SHIFT);
+            q7 = vec_ctf(qmat[15], QMAT_SHIFT);
+
+            alt0 = vec_sel(vec_madd(alt0, q0, negBias), vec_madd(alt0, q0, bias),
+                    vec_cmpgt(alt0, zero));
+            alt1 = vec_sel(vec_madd(alt1, q1, negBias), vec_madd(alt1, q1, bias),
+                    vec_cmpgt(alt1, zero));
+            alt2 = vec_sel(vec_madd(alt2, q2, negBias), vec_madd(alt2, q2, bias),
+                    vec_cmpgt(alt2, zero));
+            alt3 = vec_sel(vec_madd(alt3, q3, negBias), vec_madd(alt3, q3, bias),
+                    vec_cmpgt(alt3, zero));
+            alt4 = vec_sel(vec_madd(alt4, q4, negBias), vec_madd(alt4, q4, bias),
+                    vec_cmpgt(alt4, zero));
+            alt5 = vec_sel(vec_madd(alt5, q5, negBias), vec_madd(alt5, q5, bias),
+                    vec_cmpgt(alt5, zero));
+            alt6 = vec_sel(vec_madd(alt6, q6, negBias), vec_madd(alt6, q6, bias),
+                    vec_cmpgt(alt6, zero));
+            alt7 = vec_sel(vec_madd(alt7, q7, negBias), vec_madd(alt7, q7, bias),
+                    vec_cmpgt(alt7, zero));
+        }
+
+ 
+    }
+
+    // Store the data back into the original block
+    {
+        vector signed short data0, data1, data2, data3, data4, data5, data6, data7;
+
+        data0 = vec_pack(vec_cts(row0, 0), vec_cts(alt0, 0));
+        data1 = vec_pack(vec_cts(row1, 0), vec_cts(alt1, 0));
+        data2 = vec_pack(vec_cts(row2, 0), vec_cts(alt2, 0));
+        data3 = vec_pack(vec_cts(row3, 0), vec_cts(alt3, 0));
+        data4 = vec_pack(vec_cts(row4, 0), vec_cts(alt4, 0));
+        data5 = vec_pack(vec_cts(row5, 0), vec_cts(alt5, 0));
+        data6 = vec_pack(vec_cts(row6, 0), vec_cts(alt6, 0));
+        data7 = vec_pack(vec_cts(row7, 0), vec_cts(alt7, 0));
+
+        {
+            // Clamp for overflow
+            vector signed int max_q_int, min_q_int;
+            vector signed short max_q, min_q;
+
+            LOAD4(max_q_int, &(s->max_qcoeff));
+            LOAD4(min_q_int, &(s->min_qcoeff));
+
+            max_q = vec_pack(max_q_int, max_q_int);
+            min_q = vec_pack(min_q_int, min_q_int);
+
+            data0 = vec_max(vec_min(data0, max_q), min_q);
+            data1 = vec_max(vec_min(data1, max_q), min_q);
+            data2 = vec_max(vec_min(data2, max_q), min_q);
+            data4 = vec_max(vec_min(data4, max_q), min_q);
+            data5 = vec_max(vec_min(data5, max_q), min_q);
+            data6 = vec_max(vec_min(data6, max_q), min_q);
+            data7 = vec_max(vec_min(data7, max_q), min_q);
+        }
+
+        vector bool char zero_01, zero_23, zero_45, zero_67;
+        vector signed char scanIndices_01, scanIndices_23, scanIndices_45, scanIndices_67;
+        vector signed char negOne = vec_splat_s8(-1);
+        vector signed char* scanPtr =
+                (vector signed char*)(s->intra_scantable.inverse);
+
+        // Determine the largest non-zero index.
+        zero_01 = vec_pack(vec_cmpeq(data0, (vector short)zero),
+                vec_cmpeq(data1, (vector short)zero));
+        zero_23 = vec_pack(vec_cmpeq(data2, (vector short)zero),
+                vec_cmpeq(data3, (vector short)zero));
+        zero_45 = vec_pack(vec_cmpeq(data4, (vector short)zero),
+                vec_cmpeq(data5, (vector short)zero));
+        zero_67 = vec_pack(vec_cmpeq(data6, (vector short)zero),
+                vec_cmpeq(data7, (vector short)zero));
+
+        // 64 biggest values
+        scanIndices_01 = vec_sel(scanPtr[0], negOne, zero_01);
+        scanIndices_23 = vec_sel(scanPtr[1], negOne, zero_23);
+        scanIndices_45 = vec_sel(scanPtr[2], negOne, zero_45);
+        scanIndices_67 = vec_sel(scanPtr[3], negOne, zero_67);
+
+        // 32 largest values
+        scanIndices_01 = vec_max(scanIndices_01, scanIndices_23);
+        scanIndices_45 = vec_max(scanIndices_45, scanIndices_67);
+
+        // 16 largest values
+        scanIndices_01 = vec_max(scanIndices_01, scanIndices_45);
+
+        // 8 largest values
+        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
+                vec_mergel(scanIndices_01, negOne));
+
+        // 4 largest values
+        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
+                vec_mergel(scanIndices_01, negOne));
+
+        // 2 largest values
+        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
+                vec_mergel(scanIndices_01, negOne));
+
+        // largest value
+        scanIndices_01 = vec_max(vec_mergeh(scanIndices_01, negOne),
+                vec_mergel(scanIndices_01, negOne));
+
+        scanIndices_01 = vec_splat(scanIndices_01, 0);
+
+        signed char lastNonZeroChar;
+
+        vec_ste(scanIndices_01, 0, &lastNonZeroChar);
+
+        lastNonZero = lastNonZeroChar;
+        
+        // While the data is still in vectors we check for the transpose IDCT permute
+        // and handle it using the vector unit if we can.  This is the permute used
+        // by the altivec idct, so it is common when using the altivec dct.
+
+        if ((lastNonZero > 0) && (s->idct_permutation_type == FF_TRANSPOSE_IDCT_PERM))
+        {
+            TRANSPOSE8(data0, data1, data2, data3, data4, data5, data6, data7);
+        }
+
+        vec_st(data0, 0, data);
+        vec_st(data1, 16, data);
+        vec_st(data2, 32, data);
+        vec_st(data3, 48, data);
+        vec_st(data4, 64, data);
+        vec_st(data5, 80, data);
+        vec_st(data6, 96, data);
+        vec_st(data7, 112, data);
+    }
+
+    // special handling of block[0]
+    if (s->mb_intra)
+    {
+        if (!s->h263_aic)
+        {
+            if (n < 4)
+                oldBaseValue /= s->y_dc_scale;
+            else
+                oldBaseValue /= s->c_dc_scale;
+        }
+
+        // Divide by 8, rounding the result
+        data[0] = (oldBaseValue + 4) >> 3;
+    }
+
+    // We handled the tranpose permutation above and we don't
+    // need to permute the "no" permutation case.
+    if ((lastNonZero > 0) &&
+        (s->idct_permutation_type != FF_TRANSPOSE_IDCT_PERM) &&
+        (s->idct_permutation_type != FF_NO_IDCT_PERM))
+    {
+        ff_block_permute(data, s->idct_permutation,
+                s->intra_scantable.scantable, lastNonZero);
+    }
+
+    return lastNonZero;
+}
+
diff --git a/libavcodec/ppc/mpegvideo_ppc.c b/libavcodec/ppc/mpegvideo_ppc.c
new file mode 100644
index 0000000000..1f5e5a4ef1
--- /dev/null
+++ b/libavcodec/ppc/mpegvideo_ppc.c
@@ -0,0 +1,77 @@
+/*
+ * Copyright (c) 2002 Dieter Shirley
+ *
+ * This library is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2 of the License, or (at your option) any later version.
+ *
+ * This library is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with this library; if not, write to the Free Software
+ * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ */
+ 
+#include <time.h>
+#include "../../config.h"
+#include "../dsputil.h"
+#include "../mpegvideo.h"
+
+#ifdef HAVE_ALTIVEC
+#include "dsputil_altivec.h"
+#endif
+
+extern int dct_quantize_altivec(MpegEncContext *s,  
+        DCTELEM *block, int n,
+        int qscale, int *overflow);
+
+extern void idct_put_altivec(UINT8 *dest, int line_size, INT16 *block);
+extern void idct_add_altivec(UINT8 *dest, int line_size, INT16 *block);
+
+
+void MPV_common_init_ppc(MpegEncContext *s)
+{
+#if HAVE_ALTIVEC
+    if (has_altivec())
+    {
+        if ((s->avctx->idct_algo == FF_IDCT_AUTO) ||
+                (s->avctx->idct_algo == FF_IDCT_ALTIVEC))
+        {
+            s->idct_put = idct_put_altivec;
+            s->idct_add = idct_add_altivec;
+            s->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM;
+        }
+
+        // Test to make sure that the dct required alignments are met.
+        if ((((long)(s->q_intra_matrix) & 0x0f) != 0) ||
+                (((long)(s->q_inter_matrix) & 0x0f) != 0))
+        {
+            fprintf(stderr, "Internal Error: q-matrix blocks must be 16-byte aligned "
+                    "to use Altivec DCT. Reverting to non-altivec version.\n");
+            return;
+        }
+
+        if (((long)(s->intra_scantable.inverse) & 0x0f) != 0)
+        {
+            fprintf(stderr, "Internal Error: scan table blocks must be 16-byte aligned "
+                    "to use Altivec DCT. Reverting to non-altivec version.\n");
+            return;
+        }
+
+
+        if ((s->avctx->dct_algo == FF_DCT_AUTO) ||
+                (s->avctx->dct_algo == FF_DCT_ALTIVEC))
+        {
+            s->dct_quantize = dct_quantize_altivec;
+        }
+    } else
+#endif
+    {
+        /* Non-AltiVec PPC optimisations here */
+    }
+}
+