bfin dsputils, basic pixel operations sads, diffs, motion compensation

and standard IEEE 8x8 block transforms
patch by Marc Hoffman, mmh pleasantst com

Originally committed as revision 8594 to svn://svn.ffmpeg.org/ffmpeg/trunk

6 files changed, 1678 insertions, 28 deletions

diff --git a/libavcodec/Makefile b/libavcodec/Makefile
index 5ef10bdf84..69c298df89 100644
--- a/libavcodec/Makefile
+++ b/libavcodec/Makefile
@@ -394,6 +394,11 @@ endif
 
 OBJS-$(TARGET_ARCH_BFIN)               += bfin/dsputil_bfin.o \
 
+ASM_OBJS-$(TARGET_ARCH_BFIN)           += bfin/pixels_bfin.o \
+                                          bfin/idct_bfin.o   \
+                                          bfin/fdct_bfin.o   \
+                                          bfin/xidct.o       \
+
 EXTRALIBS := -L$(BUILD_ROOT)/libavutil -lavutil$(BUILDSUF) $(EXTRALIBS)
 
 NAME=avcodec
diff --git a/libavcodec/bfin/config_bfin.h b/libavcodec/bfin/config_bfin.h
new file mode 100644
index 0000000000..1f5080900d
--- /dev/null
+++ b/libavcodec/bfin/config_bfin.h
@@ -0,0 +1,46 @@
+/*
+ * config_bfin.h
+ *
+ * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg 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 FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+/*
+   low level assembler interface wrapper
+
+DEFUN(put_pixels_clamped,mL1,
+        (DCTELEM *block, uint8_t *dest, int line_size)):
+
+      body
+
+      rts;
+*/
+#ifndef DEFUN
+
+#ifndef mL1
+#define mL1 .l1.text
+#endif
+#define mL3 .text
+
+#define DEFUN(fname,where,interface) \
+        .section where;              \
+        .global _ff_bfin_ ## fname ; \
+        .align 8;                    \
+        _ff_bfin_ ## fname
+
+#endif
+
diff --git a/libavcodec/bfin/dsputil_bfin.c b/libavcodec/bfin/dsputil_bfin.c
index 196ef6feab..b4d549fb72 100644
--- a/libavcodec/bfin/dsputil_bfin.c
+++ b/libavcodec/bfin/dsputil_bfin.c
@@ -1,5 +1,8 @@
 /*
- * Copyright (c) 2006 Michael Benjamin
+ * BlackFin DSPUTILS
+ *
+ * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
+ * Copyright (c) 2006 Michael Benjamin <michael.benjamin@analog.com>
  *
  * This file is part of FFmpeg.
  *
@@ -18,38 +21,290 @@
  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  */
 
+#include <unistd.h>
+#include <bits/bfin_sram.h>
 #include "../avcodec.h"
 #include "../dsputil.h"
 
-static int sad8x8_bfin( void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h )
-{
-    int sum;
-    __asm__ __volatile__ (
-    "P0 = %1;" // blk1
-    "P1 = %2;" // blk2
-    "P2 = %3;\n" // h
-    "I0 = P0;"
-    "I1 = P1;\n"
-    "A0 = 0;"
-    "A1 = 0;\n"
-    "M0 = P2;\n"
-    "P3 = 32;\n"
-    "LSETUP (sad8x8LoopBegin, sad8x8LoopEnd) LC0=P3;\n"
-    "sad8x8LoopBegin:\n"
-    "  DISALGNEXCPT || R0 = [I0] || R2 = [I1];\n"
-    "  DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];\n"
-    "sad8x8LoopEnd:\n"
-    "  SAA ( R1:0 , R3:2 );\n"
-    "R3 = A1.L + A1.H, R2 = A0.L + A0.H;\n"
-    "%0 = R2 + R3 (S);\n"
-    : "=&d" (sum)
-    : "m"(blk1), "m"(blk2), "m"(h)
-    : "P0","P1","P2","I0","I1","A0","A1","R0","R1","R2","R3");
-    return sum;
+#define USE_L1CODE
+
+#ifdef USE_L1CODE
+#define L1CODE __attribute__ ((l1_text))
+#else
+#define L1CODE
+#endif
+int off;
+
+
+extern void ff_bfin_idct (DCTELEM *block) L1CODE;
+extern void ff_bfin_fdct (DCTELEM *block) L1CODE;
+extern void ff_bfin_add_pixels_clamped (DCTELEM *block, uint8_t *dest, int line_size) L1CODE;
+extern void ff_bfin_put_pixels_clamped (DCTELEM *block, uint8_t *dest, int line_size) L1CODE;
+extern void ff_bfin_diff_pixels (DCTELEM *block, uint8_t *s1, uint8_t *s2, int stride)  L1CODE;
+extern void ff_bfin_get_pixels  (DCTELEM *restrict block, const uint8_t *pixels, int line_size) L1CODE;
+extern int  ff_bfin_pix_norm1  (uint8_t * pix, int line_size) L1CODE;
+extern int  ff_bfin_z_sad8x8   (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h) L1CODE;
+extern int  ff_bfin_z_sad16x16 (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h) L1CODE;
+
+extern void ff_bfin_z_put_pixels16_xy2     (uint8_t *block, const uint8_t *s0, int dest_size, int line_size, int h) L1CODE;
+extern void ff_bfin_z_put_pixels8_xy2      (uint8_t *block, const uint8_t *s0, int dest_size, int line_size, int h) L1CODE;
+extern void ff_bfin_put_pixels16_xy2_nornd (uint8_t *block, const uint8_t *s0, int line_size, int h) L1CODE;
+extern void ff_bfin_put_pixels8_xy2_nornd  (uint8_t *block, const uint8_t *s0, int line_size, int h) L1CODE;
+
+
+extern int  ff_bfin_pix_sum (uint8_t *p, int stride) L1CODE;
+
+extern void ff_bfin_put_pixels8uc        (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int dest_size, int line_size, int h) L1CODE;
+extern void ff_bfin_put_pixels16uc       (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int dest_size, int line_size, int h) L1CODE;
+extern void ff_bfin_put_pixels8uc_nornd  (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int line_size, int h) L1CODE;
+extern void ff_bfin_put_pixels16uc_nornd (uint8_t *block, const uint8_t *s0, const uint8_t *s1, int line_size, int h) L1CODE;
+
+extern int ff_bfin_sse4  (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) L1CODE;
+extern int ff_bfin_sse8  (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) L1CODE;
+extern int ff_bfin_sse16 (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h) L1CODE;
+
+
+#if 0
+void pblk (uint8_t *p, int w, int h, int s)
+{
+    int i,j;
+    av_log (0,0,"0x%08x:\n", p);
+    for (i = 0;i<h;i++) {
+        for (j=0;j<w;j++)
+            av_log (0,0,"%3d ", p[j]);
+        p+=s;
+        av_log (0,0,"\n");
+    }
+    av_log (0,0,"\n");
+}
+#endif
+
+static void bfin_idct_add (uint8_t *dest, int line_size, DCTELEM *block)
+{
+    ff_bfin_idct (block);
+    ff_bfin_add_pixels_clamped (block, dest, line_size);
+}
+
+static void bfin_idct_put (uint8_t *dest, int line_size, DCTELEM *block)
+{
+    ff_bfin_idct (block);
+    ff_bfin_put_pixels_clamped (block, dest, line_size);
+}
+
+
+static void bfin_clear_blocks (DCTELEM *blocks)
+{
+    // This is just a simple memset.
+    //
+    asm("P0=192; "
+        "I0=%0;  "
+        "R0=0;   "
+        "LSETUP(clear_blocks_blkfn_lab,clear_blocks_blkfn_lab)LC0=P0;"
+        "clear_blocks_blkfn_lab:"
+        "[I0++]=R0;"
+        ::"a" (blocks):"P0","I0","R0");
+}
+
+
+
+static void bfin_put_pixels8 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc (block, pixels, pixels, line_size, line_size, h);
+}
+
+static void bfin_put_pixels8_x2(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc (block, pixels, pixels+1, line_size, line_size, h);
+}
+
+static void bfin_put_pixels8_y2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc (block, pixels, pixels+line_size, line_size, line_size, h);
+}
+
+static void bfin_put_pixels8_xy2 (uint8_t *block, const uint8_t *s0, int line_size, int h)
+{
+    ff_bfin_z_put_pixels8_xy2 (block,s0,line_size, line_size, h);
+}
+
+static void bfin_put_pixels16 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc (block, pixels, pixels, line_size, line_size, h);
+}
+
+static void bfin_put_pixels16_x2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc (block, pixels, pixels+1, line_size, line_size, h);
+}
+
+static void bfin_put_pixels16_y2 (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc (block, pixels, pixels+line_size, line_size, line_size, h);
+}
+
+static void bfin_put_pixels16_xy2 (uint8_t *block, const uint8_t *s0, int line_size, int h)
+{
+    ff_bfin_z_put_pixels16_xy2 (block,s0,line_size, line_size, h);
+}
+
+void bfin_put_pixels8_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels, line_size, h);
+}
+
+static void bfin_put_pixels8_x2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels+1, line_size, h);
+}
+
+static void bfin_put_pixels8_y2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc_nornd (block, pixels, pixels+line_size, line_size, h);
 }
 
+
+void bfin_put_pixels16_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels, line_size, h);
+}
+
+static void bfin_put_pixels16_x2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels+1, line_size, h);
+}
+
+static void bfin_put_pixels16_y2_nornd (uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc_nornd (block, pixels, pixels+line_size, line_size, h);
+}
+
+static int bfin_pix_abs16 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    return ff_bfin_z_sad16x16 (blk1,blk2,line_size,line_size,h);
+}
+
+static uint8_t vtmp_blk[256] __attribute__((l1_data_B));
+
+static int bfin_pix_abs16_x2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc (vtmp_blk, blk2, blk2+1, 16, line_size, h);
+    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
+}
+
+static int bfin_pix_abs16_y2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_put_pixels16uc (vtmp_blk, blk2, blk2+line_size, 16, line_size, h);
+    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
+}
+
+static int bfin_pix_abs16_xy2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_z_put_pixels16_xy2 (vtmp_blk, blk2, 16, line_size, h);
+    return ff_bfin_z_sad16x16 (blk1, vtmp_blk, line_size, 16, h);
+}
+
+static int bfin_pix_abs8 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    return ff_bfin_z_sad8x8 (blk1,blk2,line_size,line_size, h);
+}
+
+static int bfin_pix_abs8_x2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc (vtmp_blk, blk2, blk2+1, 8, line_size, h);
+    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
+}
+
+static int bfin_pix_abs8_y2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_put_pixels8uc (vtmp_blk, blk2, blk2+line_size, 8, line_size, h);
+    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
+}
+
+static int bfin_pix_abs8_xy2 (void *c, uint8_t *blk1, uint8_t *blk2, int line_size, int h)
+{
+    ff_bfin_z_put_pixels8_xy2 (vtmp_blk, blk2, 8, line_size, h);
+    return ff_bfin_z_sad8x8 (blk1, vtmp_blk, line_size, 8, h);
+}
+
+
+/*
+  decoder optimization
+  start on 2/11 100 frames of 352x240@25 compiled with no optimization -g debugging
+  9.824s ~ 2.44x off
+  6.360s ~ 1.58x off with -O2
+  5.740s ~ 1.43x off with idcts
+
+  2.64s    2/20 same sman.mp4 decode only
+
+*/
+
 void dsputil_init_bfin( DSPContext* c, AVCodecContext *avctx )
 {
-    c->pix_abs[1][0] = sad8x8_bfin;
-    c->sad[1] = sad8x8_bfin;
+    c->get_pixels         = ff_bfin_get_pixels;
+    c->diff_pixels        = ff_bfin_diff_pixels;
+    c->put_pixels_clamped = ff_bfin_put_pixels_clamped;
+    c->add_pixels_clamped = ff_bfin_add_pixels_clamped;
+
+    c->clear_blocks       = bfin_clear_blocks;
+    c->pix_sum            = ff_bfin_pix_sum;
+    c->pix_norm1          = ff_bfin_pix_norm1;
+
+    c->sad[0]             = bfin_pix_abs16;
+    c->sad[1]             = bfin_pix_abs8;
+
+    /* TODO [0] 16  [1] 8 */
+    c->pix_abs[0][0] = bfin_pix_abs16;
+    c->pix_abs[0][1] = bfin_pix_abs16_x2;
+    c->pix_abs[0][2] = bfin_pix_abs16_y2;
+    c->pix_abs[0][3] = bfin_pix_abs16_xy2;
+
+    c->pix_abs[1][0] = bfin_pix_abs8;
+    c->pix_abs[1][1] = bfin_pix_abs8_x2;
+    c->pix_abs[1][2] = bfin_pix_abs8_y2;
+    c->pix_abs[1][3] = bfin_pix_abs8_xy2;
+
+
+    c->sse[0] = ff_bfin_sse16;
+    c->sse[1] = ff_bfin_sse8;
+    c->sse[2] = ff_bfin_sse4;
+
+
+    /**
+     * Halfpel motion compensation with rounding (a+b+1)>>1.
+     * This is an array[4][4] of motion compensation functions for 4
+     * horizontal blocksizes (8,16) and the 4 halfpel positions
+     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+     * @param block destination where the result is stored
+     * @param pixels source
+     * @param line_size number of bytes in a horizontal line of block
+     * @param h height
+     */
+
+    c->put_pixels_tab[0][0] = bfin_put_pixels16;
+    c->put_pixels_tab[0][1] = bfin_put_pixels16_x2;
+    c->put_pixels_tab[0][2] = bfin_put_pixels16_y2;
+    c->put_pixels_tab[0][3] = bfin_put_pixels16_xy2;
+
+    c->put_pixels_tab[1][0] = bfin_put_pixels8;
+    c->put_pixels_tab[1][1] = bfin_put_pixels8_x2;
+    c->put_pixels_tab[1][2] = bfin_put_pixels8_y2;
+    c->put_pixels_tab[1][3] = bfin_put_pixels8_xy2;
+
+    c->put_no_rnd_pixels_tab[1][0] = bfin_put_pixels8_nornd;
+    c->put_no_rnd_pixels_tab[1][1] = bfin_put_pixels8_x2_nornd;
+    c->put_no_rnd_pixels_tab[1][2] = bfin_put_pixels8_y2_nornd;
+    c->put_no_rnd_pixels_tab[1][3] = ff_bfin_put_pixels8_xy2_nornd;
+
+    c->put_no_rnd_pixels_tab[0][0] = bfin_put_pixels16_nornd;
+    c->put_no_rnd_pixels_tab[0][1] = bfin_put_pixels16_x2_nornd;
+    c->put_no_rnd_pixels_tab[0][2] = bfin_put_pixels16_y2_nornd;
+    c->put_no_rnd_pixels_tab[0][3] = ff_bfin_put_pixels16_xy2_nornd;
+
+    c->fdct               = ff_bfin_fdct;
+    c->idct               = ff_bfin_idct;
+    c->idct_add           = bfin_idct_add;
+    c->idct_put           = bfin_idct_put;
 }
+
+
+
diff --git a/libavcodec/bfin/fdct_bfin.S b/libavcodec/bfin/fdct_bfin.S
new file mode 100644
index 0000000000..8230673e83
--- /dev/null
+++ b/libavcodec/bfin/fdct_bfin.S
@@ -0,0 +1,324 @@
+/*
+ * fdct BlackFin
+ *
+ * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg 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 FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+/*
+  void ff_bfin_fdct (DCTELEM *buf);
+
+  This implementation works only for 8x8 input. The range of input
+  must be -256 to 255 i.e. 8bit input represented in a 16bit data
+  word. The original data must be sign extended into the 16bit data
+  words.
+
+
+   Chen factorization of
+
+           8
+   X(m) = sum (x(n) * cos ((2n+1)*m*pi/16))
+          n=0
+
+                                             C4
+ 0  --*-------------*0+7---*-----*0+3-------*-*-------------------> 0
+       \           /        \   /            X S4,S4
+ 1  --*-\---------/-*1+6---*-\-/-*1+2-------*-*-------------------> 4
+         \       /            \              -C4     C3
+ 2  --*---\-----/---*2+5---*-/-\-*1-2---------------*-*-----------> 2
+           \   /            /   \                    X S3,-S3
+ 3  --*-----\-/-----*3+4---*-----*0-3---------------*-*-----------> 6
+             /                                  C7   C3
+ 4  --*-----/-\-----*3-4------------*-*4+5--*-----*---------------> 1
+           /   \            -C4      X       \   /S7    C3
+ 5  --*---/-----\---*2-5---*-*------*=*4-5----\-/------*-*--------> 5
+         /       \          X S4,S4            /        X S3,-S3
+ 6  --*-/---------\-*1-6---*-*------*=*7-6----/-\------*-*--------> 3
+       /           \        C4       X       /   \-S7   C3
+    --*-------------*0-7------------*-*7+6--*-----*---------------> 7
+                                                C7
+
+Notation
+        Cn = cos(n*pi/8) used throughout the code.
+
+
+  Registers used:
+        R0, R1, R2, R3, R4, R5, R6,R7,  P0, P1, P2, P3, P4, P5, A0, A1.
+  Other registers used:
+        I0, I1, I2, I3, B0, B2, B3, M0, M1, L3 registers and LC0.
+
+  Input - r0 - pointer to start of DCTELEM *block
+
+  Output - The DCT output coefficients in the DCTELEM *block
+
+  Register constraint:
+               This code is called from jpeg_encode.
+               R6, R5, R4 if modified should be stored and restored.
+
+
+  Performance: (Timer version 0.6.33)
+               Code Size : 240 Bytes.
+               Memory Required :
+               Input Matrix : 8 * 8 * 2 Bytes.
+               Coefficients : 16 Bytes
+               Temporary matrix: 8 * 8 * 2 Bytes.
+               Cycle Count :26+{18+8*(14+2S)}*2  where S -> Stalls
+                            (7.45 c/pel)
+        -----------------------------------------
+        |  Size  |  Forward DCT  |  Inverse DCT |
+        -----------------------------------------
+        |  8x8   |   284 Cycles  |  311 Cycles  |
+        -----------------------------------------
+
+Ck = int16(cos(k/16*pi)*32767+.5)/2
+#define C4 23170
+#define C3 13623
+#define C6 6270
+#define C7 3196
+
+Sk = int16(sin(k/16*pi)*32767+.5)/2
+#define S4 11585
+#define S3 9102
+#define S6 15137
+#define S7 16069
+
+the coefficients are ordered as follows:
+short dct_coef[]
+  C4,S4,
+  C6,S6,
+  C7,S7,
+  S3,C3,
+
+-----------------------------------------------------------
+FFMPEG conformance testing results
+-----------------------------------------------------------
+dct-test: modified with the following
+            dct_error("BFINfdct", 0, ff_bfin_fdct, fdct, test);
+produces the following output:
+
+root:/u/ffmpeg/bhead/libavcodec> ./dct-test
+ffmpeg DCT/IDCT test
+
+    2  -131    -6   -48   -36    33   -83    24
+   34    52   -24   -15     5    92    57   143
+  -67   -43    -1    74   -16     5   -71    32
+  -78   106    92   -34   -38    81    20   -18
+    7   -62    40     2   -15    90   -62   -83
+  -83     1  -104   -13    43   -19     7    11
+  -63    31    12   -29    83    72    21    10
+  -17   -63   -15    73    50   -91   159   -14
+DCT BFINfdct: err_inf=2 err2=0.16425938 syserr=0.00795000 maxout=2098 blockSumErr=27
+DCT BFINfdct: 92.1 kdct/s
+root:/u/ffmpeg/bhead/libavcodec>
+
+*/
+
+#include "config_bfin.h"
+
+.section .l1.data.B,"aw",@progbits
+.align 4;
+dct_coeff:
+.short 0x5a82, 0x2d41, 0x187e, 0x3b21, 0x0c7c, 0x3ec5, 0x238e, 0x3537;
+
+.section .l1.data.A,"aw",@progbits
+.align 4
+vtmp:   .space 128
+
+DEFUN(fdct,mL1,
+        (DCTELEM *block)):
+    [--SP] = (R7:4, P5:3);          // Push the registers onto the stack.
+
+    b0 = r0;
+    r0 = [P3+dct_coeff@GOT17M4];
+    b3 = r0;
+    r0 = [P3+vtmp@GOT17M4];
+    b2 = r0;
+
+    L3 = 16;                        // L3 is set to 16 to make the coefficient
+                                    // array Circular.
+
+
+//----------------------------------------------------------------------------
+
+/*
+ * I0, I1, and I2 registers are used to read the input data. I3 register is used
+ * to read the coefficients. P0 and P1 registers are used for writing the output
+ * data.
+ */
+    M0 = 12 (X);                    // All these initializations are used in the
+    M1 = 16 (X);                    // modification of address offsets.
+
+    M2 = 128 (X);
+
+    P2 = 16;
+    P3 = 32 (X);
+    P4 = -110 (X);
+    P5 = -62 (X);
+    P0 = 2(X);
+
+
+    // Prescale the input to get the correct precision.
+    i0=b0;
+    i1=b0;
+
+    lsetup (.0, .1) LC0 = P3;
+    r0=[i0++];
+.0:     r1=r0<<3 (v) || r0=[i0++] ;
+.1:     [i1++]=r1;
+
+        /*
+         * B0 points to the "in" buffer.
+         * B2 points to "temp" buffer in the first iteration.
+         */
+
+    lsetup (.2, .3) LC0 = P0;
+.2:
+        I0 = B0;                        // I0 points to Input Element (0, 0).
+        I1 = B0;                        // Element 1 and 0 is read in R0.
+        I1 += M0  || R0 = [I0++];       // I1 points to Input Element (0, 6).
+        I2 = I1;                        // Element 6 is read into R3.H.
+        I2 -= 4   || R3.H = W[I1++];    // I2 points to Input Element (0, 4).
+
+        I3 = B3;                        // I3 points to Coefficients.
+        P0 = B2;                        // P0 points to temporary array Element
+                                        //   (0, 0).
+        P1 = B2;                        // P1 points to temporary array.
+        R7 = [P1++P2] || R2 = [I2++];   // P1 points to temporary array
+                                        //   Element (1, 0).
+                                        // R7 is a dummy read. X4,X5
+                                        //   are read into R2.
+        R3.L = W[I1--];                 // X7 is read into R3.L.
+        R1.H = W[I0++];                 // X2 is read into R1.H.
+
+
+        /*
+         *  X0 = (X0 + X7) / 2.
+         *  X1 = (X1 + X6) / 2.
+         *  X6 = (X1 - X6) / 2.
+         *  X7 = (X0 - X7) / 2.
+         *  It reads the data 3 in R1.L.
+         */
+
+        R0 = R0 +|+ R3, R3 = R0 -|- R3 || R1.L = W[I0++] || NOP;
+
+        /*
+         *       X2 = (X2 + X5) / 2.
+         *       X3 = (X3 + X4) / 2.
+         *       X4 = (X3 - X4) / 2.
+         *       X5 = (X2 - X5) / 2.
+         *          R7 = C4 = cos(4*pi/16)
+         */
+
+        R1 = R1 +|+ R2, R2 = R1 -|- R2 (CO) || NOP      ||  R7 = [I3++];
+
+        /*
+         * At the end of stage 1 R0 has (1,0), R1 has (2,3), R2 has (4, 5) and
+         * R3 has (6,7).
+         * Where the notation (x, y) represents uper/lower half pairs.
+         */
+
+        /*
+         *       X0 = X0 + X3.
+         *       X1 = X1 + X2.
+         *       X2 = X1 - X2.
+         *       X3 = X0 - X3.
+         */
+        R0 = R0 +|+ R1, R1 = R0 -|- R1;
+
+        lsetup (.row0, .row1) LC1 = P2 >> 1;  // 1d dct, loops 8x
+.row0:
+
+        /*
+         *       This is part 2 computation continued.....
+         *       A1 =      X6 * cos(pi/4)
+         *       A0 =      X6 * cos(pi/4)
+         *       A1 = A1 - X5 * cos(pi/4)
+         *       A0 = A0 + X5 * cos(pi/4).
+         *       The instruction W[I0] = R3.L is used for packing it to R2.L.
+         */
+
+        A1=R3.H*R7.l,         A0=R3.H*R7.l            ||  I1+=M1 || W[I0] = R3.L;
+        R4.H=(A1-=R2.L*R7.l), R4.L=(A0+=R2.L*R7.l)    ||  I2+=M0 || NOP;
+
+        /*       R0 = (X1,X0)      R1 = (X2,X3)       R4 = (X5, X6). */
+
+        /*
+         *       A1 =      X0 * cos(pi/4)
+         *       A0 =      X0 * cos(pi/4)
+         *       A1 = A1 - X1 * cos(pi/4)
+         *       A0 = A0 + X1 * cos(pi/4)
+         *       R7 = (C2,C6)
+         */
+        A1=R0.L*R7.h,        A0=R0.L*R7.h             || NOP       || R3.H=W[I1++];
+        R5.H=(A1-=R0.H*R7.h),R5.L=(A0+=R0.H*R7.h)     || R7=[I3++] || NOP;
+
+        /*
+         *       A1 =      X2 * cos(3pi/8)
+         *       A0 =      X3 * cos(3pi/8)
+         *       A1 = A1 + X3 * cos(pi/8)
+         *       A0 = A0 - X2 * cos(pi/8)
+         *         R3 = cos(pi/4)
+         *         R7 = (cos(7pi/8),cos(pi/8))
+         *       X4 = X4 + X5.
+         *       X5 = X4 - X5.
+         *       X6 = X7 - X6.
+         *       X7 = X7 + X6.
+         */
+        A1=R1.H*R7.L,        A0=R1.L*R7.L             || W[P0++P3]=R5.L || R2.L=W[I0];
+        R2=R2+|+R4,          R4=R2-|-R4               || I0+=4          || R3.L=W[I1--];
+        R6.H=(A1+=R1.L*R7.H),R6.L=(A0 -= R1.H * R7.H) || I0+=4          || R7=[I3++];
+
+        /*         R2 = (X4, X7)        R4 = (X5,X6)      R5 = (X1, X0)       R6 = (X2,X3). */
+
+        /*
+         *       A1 =      X4 * cos(7pi/16)
+         *       A0 =      X7 * cos(7pi/16)
+         *       A1 = A1 + X7 * cos(pi/16)
+         *       A0 = A0 - X4 * cos(pi/16)
+         */
+
+        A1=R2.H*R7.L,        A0=R2.L*R7.L             || W[P0++P3]=R6.H || R0=[I0++];
+        R2.H=(A1+=R2.L*R7.H),R2.L=(A0-=R2.H*R7.H)     || W[P0++P3]=R5.H || R7=[I3++];
+
+        /*
+         *       A1 =      X5 * cos(3pi/16)
+         *       A0 =      X6 * cos(3pi/16)
+         *       A1 = A1 + X6 * cos(5pi/16)
+         *       A0 = A0 - X5 * cos(5pi/16)
+         *       The output values are written.
+         */
+
+        A1=R4.H*R7.H,        A0=R4.L*R7.H             || W[P0++P2]=R6.L || R1.H=W[I0++];
+        R4.H=(A1+=R4.L*R7.L),R4.L=(A0-=R4.H*R7.L)     || W[P0++P4]=R2.L || R1.L=W[I0++];
+
+
+        /* Beginning of next stage, **pipelined** + drain and store the
+           rest of the column store. */
+
+        R0=R0+|+R3,R3=R0-|-R3                         || W[P1++P3]=R2.H || R2=[I2++];
+        R1=R1+|+R2,R2=R1-|-R2 (CO)                    || W[P1++P3]=R4.L || R7=[I3++];
+.row1:  R0=R0+|+R1,R1=R0-|-R1                         || W[P1++P5]=R4.H || NOP;
+
+        // Exchange input with output.
+        B1 = B0;
+        B0 = B2;
+.3:     B2 = B1;
+
+        L3=0;
+        (r7:4,p5:3) = [sp++];
+        RTS;
+
diff --git a/libavcodec/bfin/idct_bfin.S b/libavcodec/bfin/idct_bfin.S
new file mode 100644
index 0000000000..f6904c189b
--- /dev/null
+++ b/libavcodec/bfin/idct_bfin.S
@@ -0,0 +1,297 @@
+/*
+ * idct BlackFin
+ *
+ * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg 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 FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+/*
+   This blackfin DSP code implements an 8x8 inverse type II DCT.
+
+Prototype       : void ff_bfin_idct(DCTELEM *in)
+
+Registers Used  : A0, A1, R0-R7, I0-I3, B0, B2, B3, M0-M2, L0-L3, P0-P5, LC0.
+
+Performance     :
+                    Code Size   : 498 Bytes.
+                    Cycle Count : 417 Cycles
+
+
+-----------------------------------------------------------
+FFMPEG conformance testing results
+-----------------------------------------------------------
+
+dct-test: modified with the following
+            dct_error("BFINidct", 1, ff_bfin_idct, idct, test);
+produces the following output
+
+root:/u/ffmpeg/bhead/libavcodec> ./dct-test -i
+ffmpeg DCT/IDCT test
+
+    8    15    -2    21    24    17     0    10
+    2   -10    -5    -5    -3     7   -14    -3
+    2   -13   -10   -19    18    -6     6    -2
+    9     4    16    -3     9    12    10    15
+   15    -9    -2    10     1    16     0   -15
+  -15     5     7     3    13     0    13    20
+   -6   -15    24     9   -18     1     9   -22
+   -8    25    23     2    -7     0    30    13
+IDCT BFINidct: err_inf=1 err2=0.01002344 syserr=0.00150000 maxout=266 blockSumErr=64
+IDCT BFINidct: 88.3 kdct/s
+
+*/
+
+#include "config_bfin.h"
+
+.section .l1.data.B,"aw",@progbits
+
+.align 4;
+coefs:
+.short 0x5a82;           //  C4
+.short 0x5a82;           //  C4
+.short 0x30FC;           //cos(3pi/8)  C6
+.short 0x7642;           //cos(pi/8)   C2
+.short 0x18F9;           //cos(7pi/16)
+.short 0x7D8A;           //cos(pi/16)
+.short 0x471D;           //cos(5pi/16)
+.short 0x6A6E;           //cos(3pi/16)
+.short 0x18F9;           //cos(7pi/16)
+.short 0x7D8A;           //cos(pi/16)
+
+.section .l1.data.A
+
+vtmp: .space 256
+
+#define TMP0 FP-8
+#define TMP1 FP-12
+#define TMP2 FP-16
+
+
+DEFUN(idct,mL1,
+        (DCTELEM *block)):
+
+/********************** Function Prologue *********************************/
+    link 16;
+    [--SP] = (R7:4, P5:3);   // Push the registers onto the stack.
+    B0 = R0;                 // Pointer to Input matrix
+    R1 = [P3+coefs@GOT17M4]; // Pointer to Coefficients
+    R2 = [P3+vtmp@GOT17M4];  // Pointer to Temporary matrix
+    B3 = R1;
+    B2 = R2;
+    L3 = 20;                // L3 is used for making the coefficient array
+                            // circular.
+                            // MUST BE RESTORED TO ZERO at function exit.
+    M1 = 16 (X);            // All these registers are initialized for
+    M3 = 8(X);              // modifying address offsets.
+
+    I0 = B0;                // I0 points to Input Element (0, 0).
+    I2 = B0;                // I2 points to Input Element (0, 0).
+    I2 += M3 || R0.H = W[I0];
+                            // Element 0 is read into R0.H
+    I1 = I2;                // I1 points to input Element (0, 6).
+    I1 += 4  || R0.L = W[I2++];
+                            // I2 points to input Element (0, 4).
+                            // Element 4 is read into R0.L.
+    P2 = 8 (X);
+    P3 = 32 (X);
+    P4 = -32 (X);
+    P5 = 98 (X);
+    R7 = 0x8000(Z);
+    I3 = B3;                // I3 points to Coefficients
+    P0 = B2;                // P0 points to array Element (0, 0) of temp
+    P1 = B2;
+    R7 = [I3++] || [TMP2]=R7;            // Coefficient C4 is read into R7.H and R7.L.
+    MNOP;
+    NOP;
+
+    /*
+     *   A1 =      Y0 * cos(pi/4)
+     *   A0 =      Y0 * cos(pi/4)
+     *   A1 = A1 + Y4 * cos(pi/4)
+     *   A0 = A0 - Y4 * cos(pi/4)
+     *   load:
+     *     R1=(Y2,Y6)
+     *     R7=(C2,C6)
+     *   res:
+     *     R3=Y0, R2=Y4
+     */
+    A1=R7.H*R0.H,       A0=R7.H*R0.H (IS)       || I0+= 4       || R1.L=W[I1++];
+    R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || R1.H=W[I0--] || R7=[I3++];
+
+    LSETUP (.0, .1) LC0 = P2; // perform 8 1d idcts
+
+    P2 = 112 (X);
+    P1 = P1 + P2;           // P1 points to element (7, 0) of temp buffer.
+    P2 = -94(X);
+
+.0:
+       /*
+        *   A1 =      Y2 * cos(3pi/8)
+        *   A0 =      Y2 * cos(pi/8)
+        *   A1 = A1 - Y6 * cos(pi/8)
+        *   A0 = A0 + Y6 * cos(3pi/8)
+        *      R5 = (Y1,Y7)
+        *      R7 = (C1,C7)
+        *   res:
+        *      R1=Y2, R0=Y6
+        */
+        A1=R7.L*R1.H,       A0=R7.H*R1.H (IS)        || I0+=4        || R5.H=W[I0];
+        R1=(A1-=R7.H*R1.L), R0=(A0+=R7.L*R1.L) (IS)  || R5.L=W[I1--] || R7=[I3++];
+        /*
+        *   Y0 = Y0 + Y6.
+        *   Y4 = Y4 + Y2.
+        *   Y2 = Y4 - Y2.
+        *   Y6 = Y0 - Y6.
+        *     R3 is saved
+        *     R6.l=Y3
+        * note: R3: Y0, R2: Y4, R1: Y2, R0: Y6
+        */
+        R3=R3+R0, R0=R3-R0;
+        R2=R2+R1, R1=R2-R1 || [TMP0]=R3 || R6.L=W[I0--];
+        /*
+         *  Compute the odd portion (1,3,5,7) even is done.
+         *
+         *  Y1 = C7 * Y1 - C1 * Y7 + C3 * Y5 - C5 * Y3.
+         *  Y7 = C1 * Y1 + C7 * Y7 + C5 * Y5 + C3 * Y3.
+         *  Y5 = C5 * Y1 + C3 * Y7 + C7 * Y5 - C1 * Y3.
+         *  Y3 = C3 * Y1 - C5 * Y7 - C1 * Y5 - C7 * Y3.
+         */
+        //  R5=(Y1,Y7)  R6=(Y5,Y3)                                                   // R7=(C1,C7)
+        A1 =R7.L*R5.H,       A0 =R7.H*R5.H (IS)       || [TMP1]=R2 || R6.H=W[I2--];
+        A1-=R7.H*R5.L,       A0+=R7.L*R5.L (IS)       || I0-=4     || R7=[I3++];
+        A1+=R7.H*R6.H,       A0+=R7.L*R6.H (IS)       || I0+=M1;                     // R7=(C3,C5)
+        R3 =(A1-=R7.L*R6.L), R2 =(A0+=R7.H*R6.L) (IS);
+        A1 =R7.L*R5.H,       A0 =R7.H*R5.H (IS)       || R4=[TMP0];
+        A1+=R7.H*R5.L,       A0-=R7.L*R5.L (IS)       || I1+=M1    || R7=[I3++];     // R7=(C1,C7)
+        A1+=R7.L*R6.H,       A0-=R7.H*R6.H (IS);
+        R7 =(A1-=R7.H*R6.L), R6 =(A0-=R7.L*R6.L) (IS) || I2+=M1;
+        // R3=Y1, R2=Y7, R7=Y5, R6=Y3
+
+        /* Transpose write column. */
+        R5.H=R4+R2 (RND12);                                   // Y0=Y0+Y7
+        R5.L=R4-R2 (RND12) || R4 = [TMP1];                    // Y7=Y7-Y0
+        R2.H=R1+R7 (RND12) || W[P0++P3]=R5.H;                 // Y2=Y2+Y5 st Y0
+        R2.L=R1-R7 (RND12) || W[P1++P4]=R5.L || R7=[I3++];    // Y5=Y2-Y5 st Y7
+        R5.H=R0-R3 (RND12) || W[P0++P3]=R2.H || R1.L=W[I1++]; // Y1=Y6-Y1 st Y2
+        R5.L=R0+R3 (RND12) || W[P1++P4]=R2.L || R0.H=W[I0++]; // Y6=Y6+Y1 st Y5
+        R3.H=R4-R6 (RND12) || W[P0++P3]=R5.H || R0.L=W[I2++]; // Y3=Y3-Y4 st Y1
+        R3.L=R4+R6 (RND12) || W[P1++P4]=R5.L || R1.H=W[I0++]; // Y4=Y3+Y4 st Y6
+
+        /* pipeline loop start, + drain Y3, Y4 */
+        A1=R7.H*R0.H,       A0=R7.H*R0.H (IS)       || W[P0++P2]= R3.H || R1.H = W[I0--];
+.1:     R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || W[P1++P5]= R3.L || R7 = [I3++];
+
+
+
+    I0 = B2;                // I0 points to Input Element (0, 0)
+    I2 = B2;                // I2 points to Input Element (0, 0)
+    I2 += M3 || R0.H = W[I0];
+                            // Y0 is read in R0.H
+    I1 = I2;                // I1 points to input Element (0, 6)
+    I1 += 4  || R0.L = W[I2++];
+                            // I2 points to input Element (0, 4)
+                            // Y4 is read in R0.L
+    P2 = 8 (X);
+    I3 = B3;                // I3 points to Coefficients
+    P0 = B0;                // P0 points to array Element (0, 0) for writing
+                            // output
+    P1 = B0;
+    R7 = [I3++];            // R7.H = C4 and R7.L = C4
+    NOP;
+
+    /*
+     *   A1 =      Y0 * cos(pi/4)
+     *   A0 =      Y0 * cos(pi/4)
+     *   A1 = A1 + Y4 * cos(pi/4)
+     *   A0 = A0 - Y4 * cos(pi/4)
+     *   load:
+     *     R1=(Y2,Y6)
+     *     R7=(C2,C6)
+     *   res:
+     *     R3=Y0, R2=Y4
+     */
+    A1=R7.H*R0.H,       A0=R7.H*R0.H (IS)       || I0+=4        || R1.L=W[I1++];
+    R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || R1.H=W[I0--] || R7=[I3++];
+
+    LSETUP (.2, .3) LC0 = P2; // peform 8 1d idcts
+    P2 = 112 (X);
+    P1 = P1 + P2;
+    P2 = -94(X);
+
+.2:
+        /*
+         *   A1 =      Y2 * cos(3pi/8)
+         *   A0 =      Y2 * cos(pi/8)
+         *   A1 = A1 - Y6 * cos(pi/8)
+         *   A0 = A0 + Y6 * cos(3pi/8)
+         *      R5 = (Y1,Y7)
+         *      R7 = (C1,C7)
+         *   res:
+         *      R1=Y2, R0=Y6
+         */
+        A1=R7.L*R1.H,       A0=R7.H*R1.H (IS)        || I0+=4        || R5.H=W[I0];
+        R1=(A1-=R7.H*R1.L), R0=(A0+=R7.L*R1.L) (IS)  || R5.L=W[I1--] || R7=[I3++];
+        /*
+        *   Y0 = Y0 + Y6.
+        *   Y4 = Y4 + Y2.
+        *   Y2 = Y4 - Y2.
+        *   Y6 = Y0 - Y6.
+        *     R3 is saved
+        *     R6.l=Y3
+        * note: R3: Y0, R2: Y4, R1: Y2, R0: Y6
+        */
+        R3=R3+R0, R0=R3-R0;
+        R2=R2+R1, R1=R2-R1 || [TMP0]=R3 || R6.L=W[I0--];
+        /*
+         *  Compute the odd portion (1,3,5,7) even is done.
+         *
+         *  Y1 = C7 * Y1 - C1 * Y7 + C3 * Y5 - C5 * Y3.
+         *  Y7 = C1 * Y1 + C7 * Y7 + C5 * Y5 + C3 * Y3.
+         *  Y5 = C5 * Y1 + C3 * Y7 + C7 * Y5 - C1 * Y3.
+         *  Y3 = C3 * Y1 - C5 * Y7 - C1 * Y5 - C7 * Y3.
+         */
+        //  R5=(Y1,Y7)  R6=(Y5,Y3)                                                   // R7=(C1,C7)
+        A1 =R7.L*R5.H,       A0 =R7.H*R5.H (IS)       || [TMP1]=R2 || R6.H=W[I2--];
+        A1-=R7.H*R5.L,       A0+=R7.L*R5.L (IS)       || I0-=4     || R7=[I3++];
+        A1+=R7.H*R6.H,       A0+=R7.L*R6.H (IS)       || I0+=M1;                     // R7=(C3,C5)
+        R3 =(A1-=R7.L*R6.L), R2 =(A0+=R7.H*R6.L) (IS);
+        A1 =R7.L*R5.H,       A0 =R7.H*R5.H (IS)       || R4=[TMP0];
+        A1+=R7.H*R5.L,       A0-=R7.L*R5.L (IS)       || I1+=M1    || R7=[I3++];     // R7=(C1,C7)
+        A1+=R7.L*R6.H,       A0-=R7.H*R6.H (IS);
+        R7 =(A1-=R7.H*R6.L), R6 =(A0-=R7.L*R6.L) (IS) || I2+=M1;
+        // R3=Y1, R2=Y7, R7=Y5, R6=Y3
+
+        /* Transpose write column. */
+        R5.H=R4+R2 (RND20);                                   // Y0=Y0+Y7
+        R5.L=R4-R2 (RND20) || R4 = [TMP1];                    // Y7=Y7-Y0
+        R2.H=R1+R7 (RND20) || W[P0++P3]=R5.H;                 // Y2=Y2+Y5 st Y0
+        R2.L=R1-R7 (RND20) || W[P1++P4]=R5.L || R7=[I3++];    // Y5=Y2-Y5 st Y7
+        R5.H=R0-R3 (RND20) || W[P0++P3]=R2.H || R1.L=W[I1++]; // Y1=Y6-Y1 st Y2
+        R5.L=R0+R3 (RND20) || W[P1++P4]=R2.L || R0.H=W[I0++]; // Y6=Y6+Y1 st Y5
+        R3.H=R4-R6 (RND20) || W[P0++P3]=R5.H || R0.L=W[I2++]; // Y3=Y3-Y4 st Y1
+        R3.L=R4+R6 (RND20) || W[P1++P4]=R5.L || R1.H=W[I0++]; // Y4=Y3+Y4 st Y6
+
+        /* pipeline loop start, + drain Y3, Y4 */
+        A1=R7.H*R0.H,       A0=R7.H*R0.H (IS)       || W[P0++P2]= R3.H || R1.H = W[I0--];
+.3:     R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || W[P1++P5]= R3.L || R7 = [I3++];
+
+    L3 = 0;
+    (R7:4,P5:3)=[SP++];
+    unlink;
+    RTS;
+
+
diff --git a/libavcodec/bfin/pixels_bfin.S b/libavcodec/bfin/pixels_bfin.S
new file mode 100644
index 0000000000..2968fcff6f
--- /dev/null
+++ b/libavcodec/bfin/pixels_bfin.S
@@ -0,0 +1,723 @@
+/*
+ * Blackfin Pixel Operations
+ * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg 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 FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+#include "config_bfin.h"
+
+DEFUN(put_pixels_clamped,mL1,
+        (DCTELEM *block, uint8_t *dest, int line_size)):
+    [--SP] = (R7:4);
+    R4 = 0;
+    R5.l = 0x00ff;
+    R5.h = 0x00ff;
+    I0 = R0;         // block
+    I1 = R1;         // dest
+    R2 += -4;        // line_size
+    M1 = R2;
+    P0 = 8;
+    R0 = [I0++];
+    R1 = [I0++];
+    R2 = MAX(R0, R4) (V);
+    LSETUP (ppc$0,ppc$1) LC0=P0;
+ppc$0: R2 = MIN(R2, R5) (V);
+       R3 = MAX(R1, R4) (V);
+       R3 = MIN(R3, R5) (V)      || R0 = [I0++];
+       R6 = BYTEPACK (R2,R3)     || R1 = [I0++];
+       R2 = MAX(R0, R4) (V)      || [I1++] = R6;
+       R2 = MIN(R2, R5) (V);
+       R3 = MAX(R1, R4) (V);
+       R3 = MIN(R3, R5) (V)      || R0 = [I0++];
+       R6 = BYTEPACK (R2,R3)     || R1 = [I0++];
+ppc$1: R2 = Max(R0, R4) (V)      || [I1++M1] = R6;
+
+    (R7:4) = [SP++];
+    RTS;
+
+DEFUN(add_pixels_clamped,mL1,
+        (DCTELEM *block, uint8_t *dest, int line_size)):
+    [-- SP] = (R7:4);
+    R4 = 0;
+    I0 = 0;
+    R2 += -4;        // line_size
+    M0 = R2;
+    I1 = R1;         // dest
+    I3 = R0;         // block
+    I2 = R1;         // dest
+    P0 = 8;
+    M3 = 2;
+    R0 = [I3++]  || R2 = [I1];
+    R2 = R2 << 8                      || R0.H = W[I3--]  || R3 = [I1++];
+    R3 = R3 >> 8                      || R1.L = W[I3]    || I3 += 4;
+    R6 = BYTEOP3P(R1:0, R3:2) (LO)    || R1.H = W[I3++]  || R2 = [I1];
+
+    LSETUP(apc$2,apc$3) LC1 = P0;
+apc$2: R7 = BYTEOP3P(R1:0, R3:2) (HI, R) || R0 = [I3++]     || R3 = [I1++M0];
+       R2 = R2 << 8                      || R0.H = W[I3--];
+       R3 = R3 >> 8                      || R1.L = W[I3]    || I3 += 4;
+       R6 = R6 + R7 (S)                  || R1.H = W[I3];
+       R6 = BYTEOP3P(R1:0, R3:2) (LO)    || I3+=M3          || [I2++]=R6;
+       R7 = BYTEOP3P(R1:0, R3:2) (HI, R) || R0 = [I3++]     || R2 = [I1];
+       R2 = R2 << 8                      || R0.H = W[I3--]  || R3 = [I1++];
+       R3 = R3 >> 8                      || R1.L = W[I3]    || I3 += 4;
+       R6 = R6 + R7 (S)                  || R1.H = W[I3++];
+apc$3: R6 = BYTEOP3P(R1:0, R3:2) (LO)    || [I2++M0] = R6   || R2 = [I1];
+
+    (R7:4) = [SP++];
+    RTS;
+
+
+/*
+  motion compensation
+  primitives
+
+     * Halfpel motion compensation with rounding (a+b+1)>>1.
+     * This is an array[4][4] of motion compensation funcions for 4
+     * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
+     * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
+     * @param block destination where the result is stored
+     * @param pixels source
+     * @param line_size number of bytes in a horizontal line of block
+     * @param h height
+
+*/
+
+DEFUN(put_pixels8uc,mL1,
+        (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
+                 int dest_size, int line_size, int h)):
+        i3=r0;        // dest
+        i0=r1;        // src0
+        i1=r2;        // src1
+        r0=[sp+12];   // dest_size
+        r2=[sp+16];   // line_size
+        p0=[sp+20];   // h
+        [--sp] = (r7:6);
+        r0+=-4;
+        m3=r0;
+        r2+=-8;
+        m0=r2;
+        LSETUP(pp8$0,pp8$1) LC0=P0;
+        DISALGNEXCPT                || R0 = [I0++]  || R2  =[I1++];
+
+pp8$0:  DISALGNEXCPT                || R1 = [I0++]  || R3  =[I1++];
+        R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++M0]|| R2  =[I1++M0];
+        R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++]  || [I3++] = R6 ;
+pp8$1:  DISALGNEXCPT                || R2 = [I1++]  || [I3++M3] = R7;
+
+        (r7:6) = [sp++];
+        RTS;
+
+DEFUN(put_pixels16uc,mL1,
+        (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
+                 int dest_size, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:6);
+        i3=r0;        // dest
+        i0=r1;        // src0
+        i1=r2;        // src1
+        r0=[fp+20];   // dest_size
+        r2=[fp+24];   // line_size
+        p0=[fp+28];   // h
+
+
+        r0+=-12;
+        m3=r0;        // line_size
+        r2+=-16;
+        m0=r2;
+
+        LSETUP(pp16$0,pp16$1) LC0=P0;
+         DISALGNEXCPT                || R0 = [I0++]   || R2  =[I1++];
+
+pp16$0:  DISALGNEXCPT                || R1 = [I0++]   || R3  =[I1++];
+         R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++]   || R2  =[I1++];
+         R7 = BYTEOP1P(R1:0,R3:2)(R) || R1 = [I0++]   || R3  =[I1++];
+         [I3++] = R6;
+         R6 = BYTEOP1P(R1:0,R3:2)    || R0 = [I0++M0] || R2  =[I1++M0];
+         R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++]   || [I3++] = R7 ;
+         [I3++] = R6;
+pp16$1:  DISALGNEXCPT                || R2 = [I1++]   || [I3++M3] = R7;
+
+        (r7:6) = [sp++];
+        unlink;
+        RTS;
+
+
+
+
+
+
+DEFUN(put_pixels8uc_nornd,mL1,
+        (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
+                 int line_size, int h)):
+        i3=r0;        // dest
+        i0=r1;        // src0
+        i1=r2;        // src1
+        r2=[sp+12];   // line_size
+        p0=[sp+16];   // h
+        [--sp] = (r7:6);
+        r2+=-4;
+        m3=r2;
+        r2+=-4;
+        m0=r2;
+        LSETUP(pp8$2,pp8$3) LC0=P0;
+        DISALGNEXCPT                || R0 = [I0++]  || R2  =[I1++];
+
+pp8$2:  DISALGNEXCPT                || R1 = [I0++]  || R3  =[I1++];
+        R6 = BYTEOP1P(R1:0,R3:2)(T)  || R0 = [I0++M0]|| R2  =[I1++M0];
+        R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++]  || [I3++] = R6 ;
+pp8$3:  DISALGNEXCPT                || R2 = [I1++]  || [I3++M3] = R7;
+
+        (r7:6) = [sp++];
+        RTS;
+
+DEFUN(put_pixels16uc_nornd,mL1,
+        (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
+                 int line_size, int h)):
+        i3=r0;        // dest
+        i0=r1;        // src0
+        i1=r2;        // src1
+        r2=[sp+12];   // line_size
+        p0=[sp+16];   // h
+
+        [--sp] = (r7:6);
+        r2+=-12;
+        m3=r2;        // line_size
+        r2+=-4;
+        m0=r2;
+
+        LSETUP(pp16$2,pp16$3) LC0=P0;
+        DISALGNEXCPT                || R0 = [I0++]   || R2  =[I1++];
+
+pp16$2:
+        DISALGNEXCPT                || R1 = [I0++]   || R3  =[I1++];
+        R6 = BYTEOP1P(R1:0,R3:2)(T)    || R0 = [I0++]   || R2  =[I1++];
+        R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R1 = [I0++]   || R3  =[I1++];
+        [I3++] = R6;
+
+        R6 = BYTEOP1P(R1:0,R3:2)(T)    || R0 = [I0++M0] || R2  =[I1++M0];
+        R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++]   || [I3++] = R7 ;
+        [I3++] = R6;
+pp16$3: DISALGNEXCPT                || R2 = [I1++]   || [I3++M3] = R7;
+
+        (r7:6) = [sp++];
+
+        RTS;
+
+DEFUN(z_put_pixels16_xy2,mL1,
+        (uint8_t *block, const uint8_t *s0,
+                 int dest_size, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:4);
+        i3=r0;        // dest
+        i0=r1;        // src0--> pixels
+        i1=r1;        // src1--> pixels + line_size
+        r2+=-12;
+        m2=r2;        // m2=dest_width-4
+        r2=[fp+20];
+        m3=r2;        // line_size
+        p0=[fp+24];   // h
+        r2+=-16;
+        i1+=m3;       /* src1 + line_size */
+        m0=r2;        /* line-size - 20 */
+
+        B0 = I0;
+        B1 = I1;
+        B3 = I3;
+
+        DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
+
+        LSETUP(LS$16E,LE$16E) LC0=P0;
+LS$16E: DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++] || R2  =[I1++];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R1 = [I0++] || [I3++] = R4 ;
+        DISALGNEXCPT                       || R3 = [I1++] || [I3++] = R5;
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++M0]|| R2  = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
+LE$16E: DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
+
+        M1 = 1;
+        I3 = B3;
+        I1 = B1;
+        I0 = B0;
+
+        I0 += M1;
+        I1 += M1;
+
+        DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
+        LSETUP(LS$16O,LE$16O) LC0=P0;
+LS$16O: DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++] || R2  =[I1++];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R1 = [I0++] || R6  =[I3++];
+        R4 = R4 +|+ R6                       || R7 = [I3--];
+        R5 = R5 +|+ R7                       || [I3++] = R4;
+        DISALGNEXCPT                       || R3  =[I1++] || [I3++] = R5;
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++M0]|| R2  = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 = [I3++];
+        R4 = R4 +|+ R6                       || R7 = [I3--];
+        R5 = R5 +|+ R7                       || [I3++] = R4;
+LE$16O: DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(put_pixels16_xy2_nornd,mL1,
+        (uint8_t *block, const uint8_t *s0,
+                 int line_size, int h)):
+        link 0;
+        [--sp] = (r7:4);
+        i3=r0;        // dest
+        i0=r1;        // src0--> pixels
+        i1=r1;        // src1--> pixels + line_size
+        m3=r2;
+        r2+=-12;
+        m2=r2;
+        r2+=-4;
+        i1+=m3;       /* src1 + line_size */
+        m0=r2;        /* line-size - 20 */
+        p0=[fp+20];   // h
+
+        B0=I0;
+        B1=I1;
+        B3=I3;
+
+        DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
+
+        LSETUP(LS$16ET,LE$16ET) LC0=P0;
+LS$16ET:DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++] || R2  =[I1++];
+        R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R1 = [I0++] || [I3++] = R4 ;
+        DISALGNEXCPT                       || R3 = [I1++] || [I3++] = R5;
+        R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++M0]|| R2  = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R0 = [I0++] || [I3++] = R4 ;
+LE$16ET:DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
+
+        M1 = 1;
+        I3=B3;
+        I1=B1;
+        I0=B0;
+
+        I0 += M1;
+        I1 += M1;
+
+        DISALGNEXCPT                       || R0 = [I0++] || R2  =[I1++];
+        LSETUP(LS$16OT,LE$16OT) LC0=P0;
+LS$16OT:DISALGNEXCPT                       || R1 = [I0++] || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++] || R2  =[I1++];
+        R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R1 = [I0++] || R6  =[I3++];
+        R4 = R4 +|+ R6                                    || R7 = [I3--];
+        R5 = R5 +|+ R7                                    || [I3++] = R4;
+        DISALGNEXCPT                       || R3  =[I1++] || [I3++] = R5;
+        R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++M0]|| R2  = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R0 = [I0++] || R6 = [I3++];
+        R4 = R4 +|+ R6                                    || R7 = [I3--];
+        R5 = R5 +|+ R7                                    || [I3++] = R4;
+LE$16OT:DISALGNEXCPT                       || R2 = [I1++] || [I3++M2] = R5;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(z_put_pixels8_xy2,mL1,
+        (uint8_t *block, const uint8_t *s0,
+                 int dest_size, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:4);
+        i3=r0;        // dest
+        i0=r1;        // src0--> pixels
+        i1=r1;        // src1--> pixels + line_size
+        r2+=-4;
+        m2=r2;        // m2=dest_width-4
+        r2=[fp+20];
+        m3=r2;        // line_size
+        p0=[fp+24];   // h
+        r2+=-8;
+        i1+=m3;       /* src1 + line_size */
+        m0=r2;        /* line-size - 20 */
+
+        b0 = I0;
+        b1 = I1;
+        b3 = I3;
+
+        LSETUP(LS$8E,LE$8E) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+LS$8E:  DISALGNEXCPT                       || R1 = [I0++]   || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDL)   || R0 = [I0++M0] || R2  =[I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++]   || [I3++] = R4 ;
+LE$8E:  DISALGNEXCPT                       || R2 = [I1++]   || [I3++M2] = R5;
+
+        M1 = 1;
+        I3 = b3;
+        I1 = b1;
+        I0 = b0;
+
+        I0 += M1;
+        I1 += M1;
+
+        LSETUP(LS$8O,LE$8O) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+LS$8O:  DISALGNEXCPT                       || R1 = [I0++]   || R3  =[I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (RNDH)   || R0 = [I0++M0] || R2  =[I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++]   || R6  =[I3++];
+        R4 = R4 +|+ R6                                      || R7 = [I3--];
+        R5 = R5 +|+ R7                                      || [I3++] = R4;
+LE$8O:  DISALGNEXCPT                       || R2  =[I1++]   || [I3++M2] = R5;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(put_pixels8_xy2_nornd,mL1,
+        (uint8_t *block, const uint8_t *s0, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:4);
+        i3=r0;        // dest
+        i0=r1;        // src0--> pixels
+        i1=r1;        // src1--> pixels + line_size
+        m3=r2;
+        r2+=-4;
+        m2=r2;
+        r2+=-4;
+        i1+=m3;       /* src1 + line_size */
+        m0=r2;        /* line-size - 20 */
+        p0=[fp+20];   // h
+
+
+        b0 = I0;
+        b1 = I1;
+        b3 = I3;
+
+        LSETUP(LS$8ET,LE$8ET) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+
+LS$8ET: DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (TL)     || R0 = [I0++M0] || R2 = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (TL,R)   || R0 = [I0++]   || [I3++] = R4 ;
+LE$8ET: DISALGNEXCPT                       || R2 = [I1++]   || [I3++M2] = R5;
+
+        M1 = 1;
+        I3 = b3;
+        I1 = b1;
+        I0 = b0;
+
+        I0 += M1;
+        I1 += M1;
+
+        LSETUP(LS$8OT,LE$8OT) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2 = [I1++];
+
+LS$8OT: DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        R4 = BYTEOP2P (R3:2,R1:0) (TH)     || R0 = [I0++M0] || R2 = [I1++M0];
+        R5 = BYTEOP2P (R3:2,R1:0) (TH,R)   || R0 = [I0++]   || R6 = [I3++];
+        R4 = R4 +|+ R6                                      || R7 = [I3--];
+        R5 = R5 +|+ R7                                      || [I3++] = R4;
+LE$8OT: DISALGNEXCPT                       || R2  =[I1++]   || [I3++M2] = R5;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(diff_pixels,mL1,
+       (DCTELEM *block, uint8_t *s1, uint8_t *s2, int stride)):
+        link 0;
+        [--sp] = (r7:4);
+        p0=8;
+        i3=r0;        // block
+        i0=r1;        // s1
+        i1=r2;        // s2
+        r2=[fp+20];   // stride
+        r2+=-8;
+        m0=r2;
+
+
+        LSETUP(.LS0,.LE0) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+
+.LS0:   DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        (R5,R4) = BYTEOP16M (R1:0,R3:2)    || R0 = [I0++M0] || R2 = [I1++M0];
+        (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++]   || [I3++] = R4;
+        DISALGNEXCPT                       || R2 = [I1++]   || [I3++] = R5;
+        [i3++]=r6;
+.LE0:  [i3++]=r7;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+/*
+    for (i = 0; i < 16; i++) {
+        for (j = 0; j < 16; j++) {
+          sum += pix[j];
+        }
+        pix += line_size;
+    }
+*/
+DEFUN(pix_sum,mL1,
+        (uint8_t *p, int stride)):
+        link 0;
+        [--sp] = (r7:4);
+        p0=8;
+        i0=r0;        // s1
+        i1=r0;
+        m1=r1;
+        r1=r1+r1;
+        r1+=-16;       // stride
+        m0=r1;
+        i1+=m1;
+
+        r6=0;
+
+        LSETUP(LS$PS,LE$PS) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+
+LS$PS:  DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        (R5,R4) = BYTEOP16P (R3:2,R1:0)    || R0 = [I0++]   || R2 = [I1++];
+        r6=r6+|+r5;
+        r6=r6+|+r4;
+        (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R1 = [I0++]   || R3 = [I1++];
+        r6=r6+|+r5;
+        r6=r6+|+r4;
+        (R5,R4) = BYTEOP16P (R3:2,R1:0)    || R0 = [I0++m0] || R2 = [I1++m0];
+        r6=r6+|+r5;
+        r6=r6+|+r4;
+        (R5,R4) = BYTEOP16P (R3:2,R1:0) (R)|| R0 = [I0++]   || R2 = [I1++];
+        r6=r6+|+r5;
+LE$PS:  r6=r6+|+r4;
+        r0.l=r6.l+r6.h;
+        r0.h=0;
+
+        (r7:4) = [sp++];
+        unlink;
+        rts;
+
+
+DEFUN(get_pixels,mL1,
+        (DCTELEM *restrict block, const uint8_t *pixels, int line_size)):
+        [--sp] = (r7:4);
+        i3=r0;        // dest
+        i0=r1;        // src0
+        p0=8;
+        r2+=-8;
+        m0=r2;
+        LSETUP(gp8$0,gp8$1) LC0=P0;
+
+        DISALGNEXCPT                   || R0 = [I0++];
+        DISALGNEXCPT                   || R1 = [I0++];
+
+gp8$0:  (R7,R6) = byteunpack R1:0      || R0 = [I0++M0];
+        (R5,R4) = byteunpack R1:0 (R)  || R0 = [I0++]    || [I3++]=R6;
+        DISALGNEXCPT                   || R1 = [I0++]    || [I3++]=R7;
+        [I3++]=R4;
+gp8$1:  [I3++]=R5
+
+
+        (r7:4) = [sp++];
+        RTS;
+
+
+/* sad = sad16x16 (ubyte *mb, ubyte *refwin, srcwidth, refwinwidth, h) */
+/* 91 cycles */
+DEFUN(z_sad16x16,mL1,
+        (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)):
+        link 0;
+        I0 = R0;
+        I1 = R1;
+
+        A1 = A0 = 0;
+        R0 = [sp+20]; // rwidth
+        P2 = [sp+24]; // height
+        R3 = 16;
+        R0 = R0 - R3;
+        R3 = R2 - R3;
+        M1 = R0;
+        M0 = R3;
+
+        DISALGNEXCPT         || R0 = [I0++]    || R2 = [I1++];
+        LSETUP (s$16, e$16) LC0=P2;
+s$16:   DISALGNEXCPT         || R1 = [I0++]    || R3 = [I1++];
+        SAA (R1:0,R3:2)      || R0 = [I0++]    || R2 = [I1++];
+        SAA (R1:0,R3:2) (R)  || R1 = [I0++]    || R3 = [I1++];
+        SAA (R1:0,R3:2)      || R0 = [I0++M0]  || R2 = [I1++M1];
+e$16:   SAA (R1:0,R3:2) (R)  || R0 = [I0++]    || R2 = [I1++];
+
+        R3=A1.L+A1.H,  R2=A0.L+A0.H ;
+        R0 = R2 + R3 ;
+        unlink;
+        RTS;
+
+/* sad = sad8x8 (ubyte *mb, ubyte *refwin, int srcwidth, int refwinwidth, int h) */
+/* 36 cycles */
+DEFUN(z_sad8x8,mL1,
+        (uint8_t *blk1, uint8_t *blk2, int dsz, int line_size, int h)):
+        I0 = R0;
+        I1 = R1;
+
+        A1 = A0 = 0;
+        r0 = [sp+12]; // rwidth
+        P2 = [sp+16]; //height
+        R3 = 8;
+        R0 = R0 - R3;
+        R3 = R2 - R3;
+        M0 = R3;
+        M1 = R0;
+
+        LSETUP (s$8, e$8) LC0=P2;
+        DISALGNEXCPT         || R0 = [I0++]   || R2 = [I1++];
+        DISALGNEXCPT         || R1 = [I0++]   || R3 = [I1++];
+s$8:    SAA (R1:0,R3:2)      || R0 = [I0++M0] || R2 = [I1++M1];
+        SAA (R1:0,R3:2) (R)  || R0 = [I0++]   || R2 = [I1++];
+e$8:    DISALGNEXCPT         || R1 = [I0++]   || R3 = [I1++];
+
+        R3=A1.L+A1.H,  R2=A0.L+A0.H ;
+        R0 = R2 + R3 ;
+        RTS;
+
+DEFUN(pix_norm1,mL1,
+        (uint8_t * pix, int line_size)):
+        [--SP]=(R7:4,P5:3);
+
+        // Fetch the input arguments.
+        P1 = R0;  // pix
+        P0 = R1;  // line_size
+        P5 = 16;  // loop ctr.
+        P0 -= P5;
+        M0 = P0;  // M0 = line_size-16;
+        // Now for the real work.
+        A1 = A0 = 0;
+        lsetup(_pix_norm1_blkfn_loopStart, _pix_norm1_blkfn_loopEnd) LC1 = P5;
+        I0 = P1;
+        DISALGNEXCPT || r0 = [i0++];
+
+_pix_norm1_blkfn_loopStart:
+        // following unpacks pix1[0..15] pix1+line_size[0..15]
+        DISALGNEXCPT || r1 = [i0++];
+
+        (r5, r4) = byteunpack r1:0 || r0 = [i0++];
+        a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
+        a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
+        (r5, r4) = byteunpack r1:0(r) || r1 = [i0++];
+        a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
+        a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
+        (r5, r4) = byteunpack r1:0 || r0 = [i0++M0];
+        a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
+        a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
+        (r5, r4) = byteunpack r1:0(r) || r0 = [i0++];
+        a1 += r5.h * r5.h, a0 += r5.l * r5.l (is);
+_pix_norm1_blkfn_loopEnd:
+        a1 += r4.h * r4.h, a0 += r4.l * r4.l (is);
+
+
+// Clean up at the end:
+        R2 = A0, R3 = A1;
+        R0 = R2 + R3 (S);
+
+        (R7:4,P5:3)=[SP++];
+
+        RTS;
+
+DEFUN(sse4,mL1,
+        (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:6);
+        p0=[fp+24];   // h
+        i0=r1;        // pix1
+        i1=r2;        // pix2
+        r2=[fp+20];   // line_size
+        r2+=-4;
+        m0=r2;
+
+        a0=a1=0;
+        LSETUP(.S40,.E40) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+
+.S40:   DISALGNEXCPT                       || R1 = [I0++M0] || R3 = [I1++M0];
+        (R7,R6) = BYTEOP16M (R1:0,R3:2);
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+.E40:   a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        a0 += a1;
+        r0 = a0;
+
+        (r7:6) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(sse8,mL1,
+        (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:6);
+        p0=[fp+24];   // h
+        i0=r1;        // pix1
+        i1=r2;        // pix2
+        r2=[fp+20];   // line_size
+        r2+=-8;
+        m0=r2;
+
+        a0=a1=0;
+        LSETUP(.S80,.E80) LC0=P0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+
+.S80:   DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        (R7,R6) = BYTEOP16M (R1:0,R3:2)    || R0 = [I0++M0] || R2 = [I1++M0];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+        a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++]   || R2 = [I1++];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+.E80:   a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        a0 += a1;
+        r0 = a0;
+
+        (r7:6) = [sp++];
+        unlink;
+        rts;
+
+DEFUN(sse16,mL1,
+        (void *v, uint8_t *pix1, uint8_t *pix2, int line_size, int h)):
+        link 0;
+        [--sp] = (r7:6);
+        p0=[fp+24];   // h
+        i0=r1;        // pix1
+        i1=r2;        // pix2
+        r2=[fp+20];   // line_size
+        r2+=-16;
+        m0=r2;
+
+        a0=a1=0;
+        DISALGNEXCPT                       || R0 = [I0++]   || R2  =[I1++];
+        LSETUP(.S160,.E160) LC0=P0;
+
+.S160:  DISALGNEXCPT                       || R1 = [I0++]   || R3 = [I1++];
+        (R7,R6) = BYTEOP16M (R1:0,R3:2)    || R0 = [I0++]   || R2 = [I1++];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+        a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R1 = [I0++]   || R3 = [I1++];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+        a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        (R7,R6) = BYTEOP16M (R1:0,R3:2)    || R0 = [I0++M0] || R2 = [I1++M0];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+        a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        (R7,R6) = BYTEOP16M (R1:0,R3:2) (R)|| R0 = [I0++]   || R2 = [I1++];
+        a0 += r7.l * r7.l, a1 += r7.h * r7.h (is);
+.E160:  a0 += r6.l * r6.l, a1 += r6.h * r6.h (is);
+        a0 += a1;
+        r0 = a0;
+
+        (r7:6) = [sp++];
+        unlink;
+        rts;
+
+