suppressed nasm dependancy - rewrote forward DCT and motion estimation code

Originally committed as revision 104 to svn://svn.ffmpeg.org/ffmpeg/trunk

4 files changed, 535 insertions, 1305 deletions

diff --git a/libavcodec/i386/fdct_mmx.c b/libavcodec/i386/fdct_mmx.c
new file mode 100644
index 0000000000..614759f7d2
--- /dev/null
+++ b/libavcodec/i386/fdct_mmx.c
@@ -0,0 +1,291 @@
+/*
+ * MMX optimized forward DCT
+ * The gcc porting is Copyright (c) 2001 Gerard Lantau.
+ *
+ * from  fdctam32.c - AP922 MMX(3D-Now) forward-DCT
+ * 
+ *  Intel Application Note AP-922 - fast, precise implementation of DCT
+ *        http://developer.intel.com/vtune/cbts/appnotes.htm
+ */
+#include "../common.h"
+#include "mmx.h"
+
+#define ATTR_ALIGN(align) __attribute__ ((__aligned__ (align)))
+
+//////////////////////////////////////////////////////////////////////
+//
+// constants for the forward DCT
+// -----------------------------
+//
+// Be sure to check that your compiler is aligning all constants to QWORD
+// (8-byte) memory boundaries!  Otherwise the unaligned memory access will
+// severely stall MMX execution.
+//
+//////////////////////////////////////////////////////////////////////
+
+#define BITS_FRW_ACC	3 //; 2 or 3 for accuracy
+#define SHIFT_FRW_COL	BITS_FRW_ACC
+#define SHIFT_FRW_ROW	(BITS_FRW_ACC + 17)
+//#define RND_FRW_ROW		(262144 * (BITS_FRW_ACC - 1)) //; 1 << (SHIFT_FRW_ROW-1)
+#define RND_FRW_ROW		(1 << (SHIFT_FRW_ROW-1))
+//#define RND_FRW_COL		(2 * (BITS_FRW_ACC - 1)) //; 1 << (SHIFT_FRW_COL-1)
+#define RND_FRW_COL		(1 << (SHIFT_FRW_COL-1))
+
+//concatenated table, for forward DCT transformation
+const int16_t fdct_tg_all_16[] ATTR_ALIGN(8) = {
+    13036, 13036, 13036, 13036,		// tg * (2<<16) + 0.5
+    27146, 27146, 27146, 27146,		// tg * (2<<16) + 0.5
+    -21746, -21746, -21746, -21746,	// tg * (2<<16) + 0.5
+};
+const int16_t cos_4_16[4] = {
+    -19195, -19195, -19195, -19195,	//cos * (2<<16) + 0.5
+};
+
+const int16_t ocos_4_16[4] = {
+    23170, 23170, 23170, 23170,	//cos * (2<<15) + 0.5
+};
+
+const long long  fdct_one_corr ATTR_ALIGN(8) = 0x0001000100010001LL;
+const long fdct_r_row[2] ATTR_ALIGN(8) = {RND_FRW_ROW, RND_FRW_ROW };
+
+const int16_t tab_frw_01234567[] ATTR_ALIGN(8) = {  // forward_dct coeff table
+    //row0
+    16384, 16384, 21407, -8867,     //    w09 w01 w08 w00
+    16384, 16384, 8867, -21407,     //    w13 w05 w12 w04
+    16384, -16384, 8867, 21407,     //    w11 w03 w10 w02
+    -16384, 16384, -21407, -8867,   //    w15 w07 w14 w06
+    22725, 12873, 19266, -22725,    //    w22 w20 w18 w16
+    19266, 4520, -4520, -12873,     //    w23 w21 w19 w17
+    12873, 4520, 4520, 19266,       //    w30 w28 w26 w24
+    -22725, 19266, -12873, -22725,  //    w31 w29 w27 w25
+
+    //row1
+    22725, 22725, 29692, -12299,    //    w09 w01 w08 w00
+    22725, 22725, 12299, -29692,    //    w13 w05 w12 w04
+    22725, -22725, 12299, 29692,    //    w11 w03 w10 w02
+    -22725, 22725, -29692, -12299,  //    w15 w07 w14 w06
+    31521, 17855, 26722, -31521,    //    w22 w20 w18 w16
+    26722, 6270, -6270, -17855,     //    w23 w21 w19 w17
+    17855, 6270, 6270, 26722,       //    w30 w28 w26 w24
+    -31521, 26722, -17855, -31521,  //    w31 w29 w27 w25
+
+    //row2
+    21407, 21407, 27969, -11585,    //    w09 w01 w08 w00
+    21407, 21407, 11585, -27969,    //    w13 w05 w12 w04
+    21407, -21407, 11585, 27969,    //    w11 w03 w10 w02
+    -21407, 21407, -27969, -11585,  //    w15 w07 w14 w06
+    29692, 16819, 25172, -29692,    //    w22 w20 w18 w16
+    25172, 5906, -5906, -16819,     //    w23 w21 w19 w17
+    16819, 5906, 5906, 25172,       //    w30 w28 w26 w24
+    -29692, 25172, -16819, -29692,  //    w31 w29 w27 w25
+
+    //row3
+    19266, 19266, 25172, -10426,    //    w09 w01 w08 w00
+    19266, 19266, 10426, -25172,    //    w13 w05 w12 w04
+    19266, -19266, 10426, 25172,    //    w11 w03 w10 w02
+    -19266, 19266, -25172, -10426,  //    w15 w07 w14 w06, 
+    26722, 15137, 22654, -26722,    //    w22 w20 w18 w16
+    22654, 5315, -5315, -15137,     //    w23 w21 w19 w17
+    15137, 5315, 5315, 22654,       //    w30 w28 w26 w24
+    -26722, 22654, -15137, -26722,  //    w31 w29 w27 w25, 
+
+    //row4
+    16384, 16384, 21407, -8867,     //    w09 w01 w08 w00
+    16384, 16384, 8867, -21407,     //    w13 w05 w12 w04
+    16384, -16384, 8867, 21407,     //    w11 w03 w10 w02
+    -16384, 16384, -21407, -8867,   //    w15 w07 w14 w06
+    22725, 12873, 19266, -22725,    //    w22 w20 w18 w16
+    19266, 4520, -4520, -12873,     //    w23 w21 w19 w17
+    12873, 4520, 4520, 19266,       //    w30 w28 w26 w24
+    -22725, 19266, -12873, -22725,  //    w31 w29 w27 w25 
+
+    //row5
+    19266, 19266, 25172, -10426,    //    w09 w01 w08 w00
+    19266, 19266, 10426, -25172,    //    w13 w05 w12 w04
+    19266, -19266, 10426, 25172,    //    w11 w03 w10 w02
+    -19266, 19266, -25172, -10426,  //    w15 w07 w14 w06
+    26722, 15137, 22654, -26722,    //    w22 w20 w18 w16
+    22654, 5315, -5315, -15137,     //    w23 w21 w19 w17
+    15137, 5315, 5315, 22654,       //    w30 w28 w26 w24
+    -26722, 22654, -15137, -26722,  //    w31 w29 w27 w25
+
+    //row6
+    21407, 21407, 27969, -11585,    //    w09 w01 w08 w00
+    21407, 21407, 11585, -27969,    //    w13 w05 w12 w04
+    21407, -21407, 11585, 27969,    //    w11 w03 w10 w02
+    -21407, 21407, -27969, -11585,  //    w15 w07 w14 w06, 
+    29692, 16819, 25172, -29692,    //    w22 w20 w18 w16
+    25172, 5906, -5906, -16819,     //    w23 w21 w19 w17
+    16819, 5906, 5906, 25172,       //    w30 w28 w26 w24
+    -29692, 25172, -16819, -29692,  //    w31 w29 w27 w25, 
+
+    //row7
+    22725, 22725, 29692, -12299,    //    w09 w01 w08 w00
+    22725, 22725, 12299, -29692,    //    w13 w05 w12 w04
+    22725, -22725, 12299, 29692,    //    w11 w03 w10 w02
+    -22725, 22725, -29692, -12299,  //    w15 w07 w14 w06, 
+    31521, 17855, 26722, -31521,    //    w22 w20 w18 w16
+    26722, 6270, -6270, -17855,     //    w23 w21 w19 w17
+    17855, 6270, 6270, 26722,       //    w30 w28 w26 w24
+    -31521, 26722, -17855, -31521   //    w31 w29 w27 w25
+};
+
+
+static inline void fdct_col(const int16_t *in, int16_t *out, int offset)
+{
+    movq_m2r(*(in + offset + 1 * 8), mm0);
+    movq_m2r(*(in + offset + 6 * 8), mm1);
+    movq_r2r(mm0, mm2);
+    movq_m2r(*(in + offset + 2 * 8), mm3);
+    paddsw_r2r(mm1, mm0);
+    movq_m2r(*(in + offset + 5 * 8), mm4);
+    psllw_i2r(SHIFT_FRW_COL, mm0);
+    movq_m2r(*(in + offset + 0 * 8), mm5);
+    paddsw_r2r(mm3, mm4);
+    paddsw_m2r(*(in + offset + 7 * 8), mm5);
+    psllw_i2r(SHIFT_FRW_COL, mm4);
+    movq_r2r(mm0, mm6);
+    psubsw_r2r(mm1, mm2);
+    movq_m2r(*(fdct_tg_all_16 + 4), mm1);
+    psubsw_r2r(mm4, mm0);
+    movq_m2r(*(in + offset + 3 * 8), mm7);
+    pmulhw_r2r(mm0, mm1);
+    paddsw_m2r(*(in + offset + 4 * 8), mm7);
+    psllw_i2r(SHIFT_FRW_COL, mm5);
+    paddsw_r2r(mm4, mm6);
+    psllw_i2r(SHIFT_FRW_COL, mm7);
+    movq_r2r(mm5, mm4);
+    psubsw_r2r(mm7, mm5);
+    paddsw_r2r(mm5, mm1);
+    paddsw_r2r(mm7, mm4);
+    por_m2r(fdct_one_corr, mm1);
+    psllw_i2r(SHIFT_FRW_COL + 1, mm2);
+    pmulhw_m2r(*(fdct_tg_all_16 + 4), mm5);
+    movq_r2r(mm4, mm7);
+    psubsw_m2r(*(in + offset + 5 * 8), mm3);
+    psubsw_r2r(mm6, mm4);
+    movq_r2m(mm1, *(out + offset + 2 * 8));
+    paddsw_r2r(mm6, mm7);
+    movq_m2r(*(in + offset + 3 * 8), mm1);
+    psllw_i2r(SHIFT_FRW_COL + 1, mm3);
+    psubsw_m2r(*(in + offset + 4 * 8), mm1);
+    movq_r2r(mm2, mm6);
+    movq_r2m(mm4, *(out + offset + 4 * 8));
+    paddsw_r2r(mm3, mm2);
+    pmulhw_m2r(*ocos_4_16, mm2);
+    psubsw_r2r(mm3, mm6);
+    pmulhw_m2r(*ocos_4_16, mm6);
+    psubsw_r2r(mm0, mm5);
+    por_m2r(fdct_one_corr, mm5);
+    psllw_i2r(SHIFT_FRW_COL, mm1);
+    por_m2r(fdct_one_corr, mm2);
+    movq_r2r(mm1, mm4);
+    movq_m2r(*(in + offset + 0 * 8), mm3);
+    paddsw_r2r(mm6, mm1);
+    psubsw_m2r(*(in + offset + 7 * 8), mm3);
+    psubsw_r2r(mm6, mm4);
+    movq_m2r(*(fdct_tg_all_16 + 0), mm0);
+    psllw_i2r(SHIFT_FRW_COL, mm3);
+    movq_m2r(*(fdct_tg_all_16 + 8), mm6);
+    pmulhw_r2r(mm1, mm0);
+    movq_r2m(mm7, *(out + offset + 0 * 8));
+    pmulhw_r2r(mm4, mm6);
+    movq_r2m(mm5, *(out + offset + 6 * 8));
+    movq_r2r(mm3, mm7);
+    movq_m2r(*(fdct_tg_all_16 + 8), mm5);
+    psubsw_r2r(mm2, mm7);
+    paddsw_r2r(mm2, mm3);
+    pmulhw_r2r(mm7, mm5);
+    paddsw_r2r(mm3, mm0);
+    paddsw_r2r(mm4, mm6);
+    pmulhw_m2r(*(fdct_tg_all_16 + 0), mm3);
+    por_m2r(fdct_one_corr, mm0);
+    paddsw_r2r(mm7, mm5);
+    psubsw_r2r(mm6, mm7);
+    movq_r2m(mm0, *(out + offset + 1 * 8));
+    paddsw_r2r(mm4, mm5);
+    movq_r2m(mm7, *(out + offset + 3 * 8));
+    psubsw_r2r(mm1, mm3);
+    movq_r2m(mm5, *(out + offset + 5 * 8));
+    movq_r2m(mm3, *(out + offset + 7 * 8));
+}
+
+static inline void fdct_row(const int16_t *in, int16_t *out, const int16_t *table)
+{
+    movd_m2r(*(in + 6), mm5);
+    punpcklwd_m2r(*(in + 4), mm5);
+    movq_r2r(mm5, mm2);
+    psrlq_i2r(0x20, mm5);
+    movq_m2r(*(in + 0), mm0);
+    punpcklwd_r2r(mm2, mm5);
+    movq_r2r(mm0, mm1);
+    paddsw_r2r(mm5, mm0);
+    psubsw_r2r(mm5, mm1);
+    movq_r2r(mm0, mm2);
+    punpcklwd_r2r(mm1, mm0);
+    punpckhwd_r2r(mm1, mm2);
+    movq_r2r(mm2, mm1);
+    movq_r2r(mm0, mm2);
+    movq_m2r(*(table + 0), mm3);
+    punpcklwd_r2r(mm1, mm0);
+    movq_r2r(mm0, mm5);
+    punpckldq_r2r(mm0, mm0);
+    movq_m2r(*(table + 4), mm4);
+    punpckhwd_r2r(mm1, mm2);
+    pmaddwd_r2r(mm0, mm3);
+    movq_r2r(mm2, mm6);
+    movq_m2r(*(table + 16), mm1);
+    punpckldq_r2r(mm2, mm2);
+    pmaddwd_r2r(mm2, mm4);
+    punpckhdq_r2r(mm5, mm5);
+    pmaddwd_m2r(*(table + 8), mm0);
+    punpckhdq_r2r(mm6, mm6);
+    movq_m2r(*(table + 20), mm7);
+    pmaddwd_r2r(mm5, mm1);
+    paddd_m2r(*fdct_r_row, mm3);
+    pmaddwd_r2r(mm6, mm7);
+    pmaddwd_m2r(*(table + 12), mm2);
+    paddd_r2r(mm4, mm3);
+    pmaddwd_m2r(*(table + 24), mm5);
+    pmaddwd_m2r(*(table + 28), mm6);
+    paddd_r2r(mm7, mm1);
+    paddd_m2r(*fdct_r_row, mm0);
+    psrad_i2r(SHIFT_FRW_ROW, mm3);
+    paddd_m2r(*fdct_r_row, mm1);
+    paddd_r2r(mm2, mm0);
+    paddd_m2r(*fdct_r_row, mm5);
+    psrad_i2r(SHIFT_FRW_ROW, mm1);
+    paddd_r2r(mm6, mm5);
+    psrad_i2r(SHIFT_FRW_ROW, mm0);
+    psrad_i2r(SHIFT_FRW_ROW, mm5);
+    packssdw_r2r(mm0, mm3);
+    packssdw_r2r(mm5, mm1);
+    movq_r2r(mm3, mm6);
+    punpcklwd_r2r(mm1, mm3);
+    punpckhwd_r2r(mm1, mm6);
+    movq_r2m(mm3, *(out + 0));
+    movq_r2m(mm6, *(out + 4));
+}
+
+void fdct_mmx(int16_t *block)
+{
+    /* XXX: not thread safe */
+    static int16_t block_tmp[64] ATTR_ALIGN(8);
+    int16_t *block1, *out;
+    const int16_t *table;
+    int i;
+
+    block1 = block_tmp;
+    fdct_col(block, block1, 0);
+    fdct_col(block, block1, 4);
+
+    block1 = block_tmp;
+    table = tab_frw_01234567;
+    out = block;
+    for(i=8;i>0;i--) {
+        fdct_row(block1, out, table);
+        block1 += 8;
+        table += 32;
+        out += 8;
+    }
+}
diff --git a/libavcodec/i386/fdct_mmx.s b/libavcodec/i386/fdct_mmx.s
deleted file mode 100644
index 75c67bfe5e..0000000000
--- a/libavcodec/i386/fdct_mmx.s
+++ /dev/null
@@ -1,507 +0,0 @@
-; //////////////////////////////////////////////////////////////////////////////
-; //
-; //  fdctam32.c - AP922 MMX(3D-Now) forward-DCT
-; //  ----------
-; //  Intel Application Note AP-922 - fast, precise implementation of DCT
-; //        http://developer.intel.com/vtune/cbts/appnotes.htm
-; //  ----------
-; //  
-; //       This routine can use a 3D-Now/MMX enhancement to increase the
-; //  accuracy of the fdct_col_4 macro.  The dct_col function uses 3D-Now's
-; //  PMHULHRW instead of MMX's PMHULHW(and POR).  The substitution improves
-; //  accuracy very slightly with performance penalty.  If the target CPU
-; //  does not support 3D-Now, then this function cannot be executed.
-; //  
-; //  For a fast, precise MMX implementation of inverse-DCT 
-; //              visit http://www.elecard.com/peter
-; //
-; //  v1.0 07/22/2000 (initial release)
-; //     
-; //  liaor@iname.com  http://members.tripod.com/~liaor  
-; //////////////////////////////////////////////////////////////////////////////
-
-;;;
-;;; A.Stevens Jul 2000:	 ported to nasm syntax and disentangled from
-;;; from Win**** compiler specific stuff.
-;;; All the real work was done above though.
-;;; See above for how to optimise quality on 3DNow! CPU's
-
-		;;
-		;;		Macros for code-readability...
-		;; 
-%define INP eax		;	 pointer to (short *blk) 
-%define OUT ecx		;	 pointer to output (temporary store space qwTemp[])
-%define TABLE ebx	; pointer to tab_frw_01234567[]
-%define TABLEF ebx  ; pointer to tg_all_16
-%define round_frw_row edx
-
-
-%define x0 INP + 0*16
-%define x1 INP + 1*16
-%define x2 INP + 2*16
-%define x3 INP + 3*16
-%define x4 INP + 4*16
-%define x5 INP + 5*16
-%define x6 INP + 6*16
-%define x7 INP + 7*16
-%define y0 OUT + 0*16
-%define y1 OUT + 1*16
-%define y2 OUT + 2*16
-%define y3 OUT + 3*16
-%define y4 OUT + 4*16
-%define y5 OUT + 5*16
-%define y6 OUT + 6*16
-%define y7 OUT + 7*16
-				
-		;;
-		;; Constants for DCT
-		;;
-%define BITS_FRW_ACC	3 ; 2 or 3 for accuracy
-%define SHIFT_FRW_COL	BITS_FRW_ACC
-%define SHIFT_FRW_ROW	(BITS_FRW_ACC + 17)
-%define RND_FRW_ROW		(1 << (SHIFT_FRW_ROW-1))
-%define RND_FRW_COL		(1 << (SHIFT_FRW_COL-1))
-
-extern fdct_one_corr		
-extern fdct_r_row				;  Defined in C for convenience
-		;;
-		;; Concatenated table of forward dct transformation coeffs.
-		;; 
-extern  fdct_tg_all_16			; Defined in C for convenience
-		;; Offsets into table..
-		
-%define tg_1_16 (TABLEF + 0)
-%define tg_2_16 (TABLEF + 8)
-%define tg_3_16 (TABLEF + 16)
-%define cos_4_16 (TABLEF + 24)
-%define ocos_4_16 (TABLEF + 32)		
-
-		;;
-		;; Concatenated table of forward dct coefficients
-		;; 
-extern tab_frw_01234567		; Defined in C for convenience
-
-		;; Offsets into table..
-SECTION .text
-		
-global fdct_mmx
-		
-;;; 
-;;; void fdct_mmx( short *blk )
-;;; 
-
-
-
-;     ////////////////////////////////////////////////////////////////////////
-;     //
-;     // The high-level pseudocode for the fdct_am32() routine :
-;     //
-;     // fdct_am32()
-;     // {
-;     //    forward_dct_col03(); // dct_column transform on cols 0-3
-;     //    forward_dct_col47(); // dct_column transform on cols 4-7
-;     //    for ( j = 0; j < 8; j=j+1 )
-;     //      forward_dct_row1(j); // dct_row transform on row #j
-;     // }
-;     //
-;    
-
-align 32
-fdct_mmx:
-	push ebp			; save stack pointer
-	mov ebp, esp		; link
-
-	push ebx		
-	push ecx		
-	push edx
-	push edi
-					
-	mov INP, [ebp+8];		; input data is row 0 of blk[]
-    ;// transform the left half of the matrix (4 columns)
-
-    lea TABLEF,  [fdct_tg_all_16];
-    mov OUT, INP;
-
-;	lea round_frw_col,  [r_frw_col]
-    ; for ( i = 0; i < 2; i = i + 1)
-    ; the for-loop is executed twice.  We are better off unrolling the 
-    ; loop to avoid branch misprediction.
-.mmx32_fdct_col03: 
-    movq mm0, [x1] ; 0 ; x1
-     ;;
-
-    movq mm1, [x6] ; 1 ; x6
-    movq mm2, mm0 ; 2 ; x1
-
-    movq mm3, [x2] ; 3 ; x2
-    paddsw mm0, mm1 ; t1 = x[1] + x[6]
-
-    movq mm4, [x5] ; 4 ; x5
-    psllw mm0, SHIFT_FRW_COL ; t1
-
-    movq mm5, [x0] ; 5 ; x0
-    paddsw mm4, mm3 ; t2 = x[2] + x[5]
-
-    paddsw mm5, [x7] ; t0 = x[0] + x[7]
-    psllw mm4, SHIFT_FRW_COL ; t2
-
-    movq mm6, mm0 ; 6 ; t1
-    psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6]
-
-    movq mm1,  [tg_2_16] ; 1 ; tg_2_16
-    psubsw mm0, mm4 ; tm12 = t1 - t2
-
-    movq mm7, [x3] ; 7 ; x3
-    pmulhw mm1, mm0 ; tm12*tg_2_16
-
-    paddsw mm7, [x4] ; t3 = x[3] + x[4]
-    psllw mm5, SHIFT_FRW_COL ; t0
-
-    paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2
-    psllw mm7, SHIFT_FRW_COL ; t3
-
-    movq mm4, mm5 ; 4 ; t0
-    psubsw mm5, mm7 ; tm03 = t0 - t3
-
-    paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16
-    paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3
-
-    por mm1,  [fdct_one_corr] ; correction y2 +0.5
-    psllw mm2, SHIFT_FRW_COL+1 ; t6
-
-    pmulhw mm5,  [tg_2_16] ; tm03*tg_2_16
-    movq mm7, mm4 ; 7 ; tp03
-
-    psubsw mm3, [x5] ; t5 = x[2] - x[5]
-    psubsw mm4, mm6 ; y4 = tp03 - tp12
-
-    movq [y2], mm1 ; 1 ; save y2
-    paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12
-    
-    movq mm1, [x3] ; 1 ; x3
-    psllw mm3, SHIFT_FRW_COL+1 ; t5
-
-    psubsw mm1, [x4] ; t4 = x[3] - x[4]
-    movq mm6, mm2 ; 6 ; t6
-    
-    movq [y4], mm4 ; 4 ; save y4
-    paddsw mm2, mm3 ; t6 + t5
-
-    pmulhw mm2,  [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16
-    psubsw mm6, mm3 ; 3 ; t6 - t5
-
-    pmulhw mm6,  [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16
-    psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12
-
-    por mm5,  [fdct_one_corr] ; correction y6 +0.5
-    psllw mm1, SHIFT_FRW_COL ; t4
-
-    por mm2,  [fdct_one_corr] ; correction tp65 +0.5
-    movq mm4, mm1 ; 4 ; t4
-
-    movq mm3, [x0] ; 3 ; x0
-    paddsw mm1, mm6 ; tp465 = t4 + tm65
-
-    psubsw mm3, [x7] ; t7 = x[0] - x[7]
-    psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65
-
-    movq mm0,  [tg_1_16] ; 0 ; tg_1_16
-    psllw mm3, SHIFT_FRW_COL ; t7
-
-    movq mm6,  [tg_3_16] ; 6 ; tg_3_16
-    pmulhw mm0, mm1 ; tp465*tg_1_16
-
-    movq [y0], mm7 ; 7 ; save y0
-    pmulhw mm6, mm4 ; tm465*tg_3_16
-
-    movq [y6], mm5 ; 5 ; save y6
-    movq mm7, mm3 ; 7 ; t7
-
-    movq mm5,  [tg_3_16] ; 5 ; tg_3_16
-    psubsw mm7, mm2 ; tm765 = t7 - tp65
-
-    paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65
-    pmulhw mm5, mm7 ; tm765*tg_3_16
-
-    paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16
-    paddsw mm6, mm4 ; tm465*tg_3_16
-
-    pmulhw mm3,  [tg_1_16] ; tp765*tg_1_16
-    ;;
-
-    por mm0,  [fdct_one_corr] ; correction y1 +0.5
-    paddsw mm5, mm7 ; tm765*tg_3_16
-
-    psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16
-    add INP, 0x08   ; ; increment pointer
-
-    movq [y1], mm0 ; 0 ; save y1
-    paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465
-
-    movq [y3], mm7 ; 7 ; save y3
-    psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465
-
-    movq [y5], mm5 ; 5 ; save y5
-
-
-.mmx32_fdct_col47: ; begin processing last four columns
-    movq mm0, [x1] ; 0 ; x1
-    ;;
-    movq [y7], mm3 ; 3 ; save y7 (columns 0-4)
-    ;;
-
-    movq mm1, [x6] ; 1 ; x6
-    movq mm2, mm0 ; 2 ; x1
-
-    movq mm3, [x2] ; 3 ; x2
-    paddsw mm0, mm1 ; t1 = x[1] + x[6]
-
-    movq mm4, [x5] ; 4 ; x5
-    psllw mm0, SHIFT_FRW_COL ; t1
-
-    movq mm5, [x0] ; 5 ; x0
-    paddsw mm4, mm3 ; t2 = x[2] + x[5]
-
-    paddsw mm5, [x7] ; t0 = x[0] + x[7]
-    psllw mm4, SHIFT_FRW_COL ; t2
-
-    movq mm6, mm0 ; 6 ; t1
-    psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6]
-
-    movq mm1,  [tg_2_16] ; 1 ; tg_2_16
-    psubsw mm0, mm4 ; tm12 = t1 - t2
-
-    movq mm7, [x3] ; 7 ; x3
-    pmulhw mm1, mm0 ; tm12*tg_2_16
-
-    paddsw mm7, [x4] ; t3 = x[3] + x[4]
-    psllw mm5, SHIFT_FRW_COL ; t0
-
-    paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2
-    psllw mm7, SHIFT_FRW_COL ; t3
-
-    movq mm4, mm5 ; 4 ; t0
-    psubsw mm5, mm7 ; tm03 = t0 - t3
-
-    paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16
-    paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3
-
-    por mm1,  [fdct_one_corr] ; correction y2 +0.5
-    psllw mm2, SHIFT_FRW_COL+1 ; t6
-
-    pmulhw mm5,  [tg_2_16] ; tm03*tg_2_16
-    movq mm7, mm4 ; 7 ; tp03
-
-    psubsw mm3, [x5] ; t5 = x[2] - x[5]
-    psubsw mm4, mm6 ; y4 = tp03 - tp12
-
-    movq [y2+8], mm1 ; 1 ; save y2
-    paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12
-    
-    movq mm1, [x3] ; 1 ; x3
-    psllw mm3, SHIFT_FRW_COL+1 ; t5
-
-    psubsw mm1, [x4] ; t4 = x[3] - x[4]
-    movq mm6, mm2 ; 6 ; t6
-    
-    movq [y4+8], mm4 ; 4 ; save y4
-    paddsw mm2, mm3 ; t6 + t5
-
-    pmulhw mm2,  [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16
-    psubsw mm6, mm3 ; 3 ; t6 - t5
-
-    pmulhw mm6,  [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16
-    psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12
-
-    por mm5,  [fdct_one_corr] ; correction y6 +0.5
-    psllw mm1, SHIFT_FRW_COL ; t4
-
-    por mm2,  [fdct_one_corr] ; correction tp65 +0.5
-    movq mm4, mm1 ; 4 ; t4
-
-    movq mm3, [x0] ; 3 ; x0
-    paddsw mm1, mm6 ; tp465 = t4 + tm65
-
-    psubsw mm3, [x7] ; t7 = x[0] - x[7]
-    psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65
-
-    movq mm0,  [tg_1_16] ; 0 ; tg_1_16
-    psllw mm3, SHIFT_FRW_COL ; t7
-
-    movq mm6,  [tg_3_16] ; 6 ; tg_3_16
-    pmulhw mm0, mm1 ; tp465*tg_1_16
-
-    movq [y0+8], mm7 ; 7 ; save y0
-    pmulhw mm6, mm4 ; tm465*tg_3_16
-
-    movq [y6+8], mm5 ; 5 ; save y6
-    movq mm7, mm3 ; 7 ; t7
-
-    movq mm5,  [tg_3_16] ; 5 ; tg_3_16
-    psubsw mm7, mm2 ; tm765 = t7 - tp65
-
-    paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65
-    pmulhw mm5, mm7 ; tm765*tg_3_16
-
-    paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16
-    paddsw mm6, mm4 ; tm465*tg_3_16
-
-    pmulhw mm3,  [tg_1_16] ; tp765*tg_1_16
-    ;;
-
-    por mm0, [fdct_one_corr] ; correction y1 +0.5
-    paddsw mm5, mm7 ; tm765*tg_3_16
-
-    psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16
-    ;;
-
-    movq [y1+8], mm0 ; 0 ; save y1
-    paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465
-
-    movq [y3+8], mm7 ; 7 ; save y3
-    psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465
-
-    movq [y5+8], mm5 ; 5 ; save y5
-
-    movq [y7+8], mm3 ; 3 ; save y7
-
-;    emms;
-;    }   ; end of forward_dct_col07() 
-    ;  done with dct_row transform
-
-  
-  ; fdct_mmx32_cols() --
-  ; the following subroutine repeats the row-transform operation, 
-  ; except with different shift&round constants.  This version
-  ; does NOT transpose the output again.  Thus the final output
-  ; is transposed with respect to the source.
-  ;
-  ;  The output is stored into blk[], which destroys the original
-  ;  input data.
-	mov INP,  [ebp+8];		;; row 0
-	 mov edi, 0x08;	;x = 8
-
-	lea TABLE,  [tab_frw_01234567]; ; row 0
-	 mov OUT, INP;
-
-	lea round_frw_row,  [fdct_r_row];
-	; for ( x = 8; x > 0; --x )  ; transform one row per iteration
-
-; ---------- loop begin
-  .lp_mmx_fdct_row1:
-    movd mm5,  [INP+12]; ; mm5 = 7 6
-
-    punpcklwd mm5,  [INP+8] ; mm5 =  5 7 4 6
-
-    movq mm2, mm5;     ; mm2 = 5 7 4 6
-    psrlq mm5, 32;     ; mm5 = _ _ 5 7
-
-    movq mm0,  [INP]; ; mm0 = 3 2 1 0
-    punpcklwd mm5, mm2;; mm5 = 4 5 6 7
-
-    movq mm1, mm0;     ; mm1 = 3 2 1 0
-    paddsw mm0, mm5;   ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0)
-
-    psubsw mm1, mm5;   ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4)
-    movq mm2, mm0;     ; mm2 = [ xt3 xt2 xt1 xt0 ]
-
-    ;movq [ xt3xt2xt1xt0 ], mm0;
-    ;movq [ xt7xt6xt5xt4 ], mm1;
-
-    punpcklwd mm0, mm1;; mm0 = [ xt5 xt1 xt4 xt0 ]
-
-    punpckhwd mm2, mm1;; mm2 = [ xt7 xt3 xt6 xt2 ]
-    movq mm1, mm2;     ; mm1
-
-    ;; shuffle bytes around
-
-;  movq mm0,  [INP] ; 0 ; x3 x2 x1 x0
-
-;  movq mm1,  [INP+8] ; 1 ; x7 x6 x5 x4
-    movq mm2, mm0 ; 2 ; x3 x2 x1 x0
-
-    movq mm3,  [TABLE] ; 3 ; w06 w04 w02 w00
-    punpcklwd mm0, mm1 ; x5 x1 x4 x0
-
-    movq mm5, mm0 ; 5 ; x5 x1 x4 x0
-    punpckldq mm0, mm0 ; x4 x0 x4 x0  [ xt2 xt0 xt2 xt0 ]
-
-    movq mm4,  [TABLE+8] ; 4 ; w07 w05 w03 w01
-    punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2
-
-    pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00
-    movq mm6, mm2 ; 6 ; x7 x3 x6 x2
-
-    movq mm1,  [TABLE+32] ; 1 ; w22 w20 w18 w16
-    punpckldq mm2, mm2 ; x6 x2 x6 x2  [ xt3 xt1 xt3 xt1 ]
-
-    pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01
-    punpckhdq mm5, mm5 ; x5 x1 x5 x1  [ xt6 xt4 xt6 xt4 ]
-
-    pmaddwd mm0,  [TABLE+16] ; x4*w14+x0*w12 x4*w10+x0*w08
-    punpckhdq mm6, mm6 ; x7 x3 x7 x3  [ xt7 xt5 xt7 xt5 ]
-
-    movq mm7,  [TABLE+40] ; 7 ; w23 w21 w19 w17
-    pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16
-;mm3 = a1, a0 (y2,y0)
-;mm1 = b1, b0 (y3,y1)
-;mm0 = a3,a2  (y6,y4)
-;mm5 = b3,b2  (y7,y5)
-
-    paddd mm3,  [round_frw_row] ; +rounder (y2,y0)
-    pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17
-
-    pmaddwd mm2,  [TABLE+24] ; x6*w15+x2*w13 x6*w11+x2*w09
-    paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) ; now ( y2, y0)
-
-    pmaddwd mm5,  [TABLE+48] ; x5*w30+x1*w28 x5*w26+x1*w24
-    ;;
-
-    pmaddwd mm6,  [TABLE+56] ; x7*w31+x3*w29 x7*w27+x3*w25
-    paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) ; now ( y3, y1)
-
-    paddd mm0,  [round_frw_row] ; +rounder (y6,y4)
-    psrad mm3, SHIFT_FRW_ROW ; (y2, y0)
-
-    paddd mm1,  [round_frw_row] ; +rounder (y3,y1)
-    paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) ; now (y6, y4)
-
-    paddd mm5,  [round_frw_row] ; +rounder (y7,y5)
-    psrad mm1, SHIFT_FRW_ROW ; y1=a1+b1 y0=a0+b0
-
-    paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) ; now ( y7, y5)
-    psrad mm0, SHIFT_FRW_ROW ;y3=a3+b3 y2=a2+b2
-
-    add OUT, 16;  ; increment row-output address by 1 row
-    psrad mm5, SHIFT_FRW_ROW ; y4=a3-b3 y5=a2-b2
-
-    add INP, 16;  ; increment row-address by 1 row
-    packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0
-
-    packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1
-    movq mm6, mm3;    ; mm0 = y6 y4 y2 y0
-
-    punpcklwd mm3, mm1; ; y3 y2 y1 y0
-    sub edi, 0x01;   ; i = i - 1
-    
-    punpckhwd mm6, mm1; ; y7 y6 y5 y4
-    add TABLE,64;  ; increment to next table
-
-    movq  [OUT-16], mm3 ; 1 ; save y3 y2 y1 y0
-
-    movq  [OUT-8], mm6 ; 7 ; save y7 y6 y5 y4
-
-    cmp edi, 0x00;
-    jg near .lp_mmx_fdct_row1;  ; begin fdct processing on next row
-		;; 
-		;; Tidy up and return
-		;;
-	pop edi
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp			; restore stack pointer
-	emms
-	ret		
-  
\ No newline at end of file
diff --git a/libavcodec/i386/motion_est_mmx.c b/libavcodec/i386/motion_est_mmx.c
new file mode 100644
index 0000000000..35b16b711c
--- /dev/null
+++ b/libavcodec/i386/motion_est_mmx.c
@@ -0,0 +1,244 @@
+/*
+ * MMX optimized motion estimation
+ * Copyright (c) 2001 Gerard Lantau.
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
+ *
+ * This program 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 General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ *
+ */
+#include "../dsputil.h"
+#include "mmx.h"
+
+static const unsigned long long int mm_wone __attribute__ ((aligned(8))) = 0x0001000100010001;
+static const unsigned long long int mm_wtwo __attribute__ ((aligned(8))) = 0x0002000200020002;
+
+/* mm7 is accumulator, mm6 is zero */
+static inline void sad_add(const UINT8 *p1, const UINT8 *p2)
+{
+    movq_m2r(*p1, mm0);
+    movq_m2r(*p2, mm1);
+    movq_r2r(mm0, mm2);
+    psubusb_r2r(mm1, mm0);
+    psubusb_r2r(mm2, mm1);
+    por_r2r(mm1, mm0); /* mm0 is absolute value */
+
+    movq_r2r(mm0, mm1);
+    punpcklbw_r2r(mm6, mm0);
+    punpckhbw_r2r(mm6, mm1);
+    paddusw_r2r(mm0, mm7);
+    paddusw_r2r(mm1, mm7);
+}
+
+/* convert mm7 to value */
+static inline int sad_end(void)
+{
+    int res;
+
+    movq_r2r(mm7, mm0);
+    psrlq_i2r(32, mm7);
+    paddusw_r2r(mm0, mm7);
+
+    movq_r2r(mm7, mm0);
+    psrlq_i2r(16, mm7);
+    paddusw_r2r(mm0, mm7);
+    __asm __volatile ("movd %%mm7, %0" : "=a" (res));
+    return res & 0xffff;
+}
+
+int pix_abs16x16_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
+{
+    const UINT8 *p1, *p2;
+
+    h >>= 1;
+    p1 = blk1;
+    p2 = blk2;
+    pxor_r2r(mm7, mm7); /* mm7 is accumulator */
+    pxor_r2r(mm6, mm6); /* mm7 is zero constant */
+    do {
+        sad_add(p1, p2);
+        sad_add(p1 + 8, p2 + 8);
+        p1 += lx;
+        p2 += lx;
+        sad_add(p1, p2);
+        sad_add(p1 + 8, p2 + 8);
+        p1 += lx;
+        p2 += lx;
+    } while (--h);
+    return sad_end();
+}
+
+/* please test it ! */
+static inline void sad_add_sse(const UINT8 *p1, const UINT8 *p2)
+{
+    movq_m2r(*(p1 + 0), mm0);
+    movq_m2r(*(p1 + 8), mm1);
+    psadbw_m2r(*(p2 + 0), mm0);
+    psadbw_m2r(*(p2 + 8), mm1);
+    paddusw_r2r(mm0, mm7);
+    paddusw_r2r(mm1, mm7);
+}
+
+int pix_abs16x16_sse(UINT8 *blk1, UINT8 *blk2, int lx, int h)
+{
+    const UINT8 *p1, *p2;
+
+    h >>= 1;
+    p1 = blk1;
+    p2 = blk2;
+    pxor_r2r(mm7, mm7); /* mm7 is accumulator */
+    do {
+        sad_add_sse(p1, p2);
+        p1 += lx;
+        p2 += lx;
+        sad_add_sse(p1, p2);
+        p1 += lx;
+        p2 += lx;
+    } while (--h);
+    return sad_end();
+}
+
+#define DUMP(reg) { mmx_t tmp; movq_r2m(reg, tmp); printf(#reg "=%016Lx\n", tmp.uq); }
+
+/* mm7 is accumulator, mm6 is zero */
+static inline void sad_add_x2(const UINT8 *p1, const UINT8 *p2, const UINT8 *p3)
+{
+    movq_m2r(*(p2 + 0), mm0);
+    movq_m2r(*(p3 + 0), mm1);
+    movq_r2r(mm0, mm2);
+    movq_r2r(mm1, mm3);
+    punpcklbw_r2r(mm6, mm0); /* extract 4 bytes low */
+    punpcklbw_r2r(mm6, mm1);
+    punpckhbw_r2r(mm6, mm2); /* high */
+    punpckhbw_r2r(mm6, mm3); 
+    paddusw_r2r(mm1, mm0);
+    paddusw_r2r(mm3, mm2);
+    movq_m2r(*(p1 + 0), mm1); /* mm1 : other value */
+    paddusw_r2r(mm5, mm0); /* + 1 */
+    paddusw_r2r(mm5, mm2); /* + 1 */
+    psrlw_i2r(1, mm0);
+    psrlw_i2r(1, mm2);
+    packuswb_r2r(mm2, mm0); /* average is in mm0 */
+
+    movq_r2r(mm1, mm2); 
+    psubusb_r2r(mm0, mm1);
+    psubusb_r2r(mm2, mm0);
+    por_r2r(mm1, mm0); /* mm0 is absolute value */
+
+    movq_r2r(mm0, mm1);
+    punpcklbw_r2r(mm6, mm0);
+    punpckhbw_r2r(mm6, mm1);
+    paddusw_r2r(mm0, mm7);
+    paddusw_r2r(mm1, mm7);
+}
+
+int pix_abs16x16_x2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
+{
+    const UINT8 *p1, *p2;
+
+    p1 = blk1;
+    p2 = blk2;
+    pxor_r2r(mm7, mm7); /* mm7 is accumulator */
+    pxor_r2r(mm6, mm6); /* mm7 is zero constant */
+    movq_m2r(mm_wone, mm5); /* one constant */
+    do {
+        sad_add_x2(p1, p2, p2 + 1);
+        sad_add_x2(p1 + 8, p2 + 8, p2 + 9);
+        p1 += lx;
+        p2 += lx;
+    } while (--h);
+    return sad_end();
+}
+
+int pix_abs16x16_y2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
+{
+    const UINT8 *p1, *p2;
+
+    p1 = blk1;
+    p2 = blk2;
+    pxor_r2r(mm7, mm7); /* mm7 is accumulator */
+    pxor_r2r(mm6, mm6); /* mm7 is zero constant */
+    movq_m2r(mm_wone, mm5); /* one constant */
+    do {
+        sad_add_x2(p1, p2, p2 + lx);
+        sad_add_x2(p1 + 8, p2 + 8, p2 + 8 + lx);
+        p1 += lx;
+        p2 += lx;
+    } while (--h);
+    return sad_end();
+}
+
+/* mm7 is accumulator, mm6 is zero */
+static inline void sad_add_xy2(const UINT8 *p1, const UINT8 *p2, const UINT8 *p3)
+{
+    movq_m2r(*(p2 + 0), mm0);
+    movq_m2r(*(p3 + 0), mm1);
+    movq_r2r(mm0, mm2);
+    movq_r2r(mm1, mm3);
+    punpcklbw_r2r(mm6, mm0); /* extract 4 bytes low */
+    punpcklbw_r2r(mm6, mm1);
+    punpckhbw_r2r(mm6, mm2); /* high */
+    punpckhbw_r2r(mm6, mm3); 
+    paddusw_r2r(mm1, mm0);
+    paddusw_r2r(mm3, mm2);
+
+    movq_m2r(*(p2 + 1), mm1);
+    movq_m2r(*(p3 + 1), mm3);
+    movq_r2r(mm1, mm4);
+    punpcklbw_r2r(mm6, mm1); /* low */
+    punpckhbw_r2r(mm6, mm4); /* high */
+    paddusw_r2r(mm1, mm0);
+    paddusw_r2r(mm4, mm2);
+    movq_r2r(mm3, mm4);
+    punpcklbw_r2r(mm6, mm3); /* low */
+    punpckhbw_r2r(mm6, mm4); /* high */
+    paddusw_r2r(mm3, mm0);
+    paddusw_r2r(mm4, mm2);
+    
+    movq_m2r(*(p1 + 0), mm1); /* mm1 : other value */
+    paddusw_r2r(mm5, mm0); /* + 2 */
+    paddusw_r2r(mm5, mm2); /* + 2 */
+    psrlw_i2r(2, mm0);
+    psrlw_i2r(2, mm2);
+    packuswb_r2r(mm2, mm0); /* average is in mm0 */
+
+    movq_r2r(mm1, mm2); 
+    psubusb_r2r(mm0, mm1);
+    psubusb_r2r(mm2, mm0);
+    por_r2r(mm1, mm0); /* mm0 is absolute value */
+
+    movq_r2r(mm0, mm1);
+    punpcklbw_r2r(mm6, mm0);
+    punpckhbw_r2r(mm6, mm1);
+    paddusw_r2r(mm0, mm7);
+    paddusw_r2r(mm1, mm7);
+}
+
+int pix_abs16x16_xy2_mmx(UINT8 *blk1, UINT8 *blk2, int lx, int h)
+{
+    const UINT8 *p1, *p2, *p3;
+
+    p1 = blk1;
+    p2 = blk2;
+    p3 = blk2 + lx;
+    pxor_r2r(mm7, mm7); /* mm7 is accumulator */
+    pxor_r2r(mm6, mm6); /* mm7 is zero constant */
+    movq_m2r(mm_wtwo, mm5); /* one constant */
+    do {
+        sad_add_xy2(p1, p2, p2 + lx);
+        sad_add_xy2(p1 + 8, p2 + 8, p2 + 8 + lx);
+        p1 += lx;
+        p2 += lx;
+    } while (--h);
+    return sad_end();
+}
diff --git a/libavcodec/i386/sad_mmx.s b/libavcodec/i386/sad_mmx.s
deleted file mode 100644
index b512639c25..0000000000
--- a/libavcodec/i386/sad_mmx.s
+++ /dev/null
@@ -1,798 +0,0 @@
-;  MMX/SSE optimized routines for SAD of 16*16 macroblocks
-;	Copyright (C) Juan J. Sierralta P. <juanjo@atmlab.utfsm.cl>
-;
-;  dist1_* Original Copyright (C) 2000 Chris Atenasio <chris@crud.net>
-;  Enhancements and rest Copyright (C) 2000 Andrew Stevens <as@comlab.ox.ac.uk>
-
-;
-;  This program is free software; you can redistribute it and/or
-;  modify it under the terms of the GNU General Public License
-;  as published by the Free Software Foundation; either version 2
-;  of the License, or (at your option) any later version.
-;
-;  This program 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 General Public License for more details.
-;
-;  You should have received a copy of the GNU General Public License
-;  along with this program; if not, write to the Free Software
-;  Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111-1307, USA.
-;
-
-global pix_abs16x16_mmx
-
-; int  pix_abs16x16_mmx(unsigned char *pix1,unsigned char *pix2, int lx, int h);
-; esi = p1 (init:		blk1)
-; edi = p2 (init:		blk2)
-; ecx = rowsleft (init:	 h)
-; edx = lx;
-
-; mm0 = distance accumulators (4 words)
-; mm1 = distance accumulators (4 words) 
-; mm2 = temp 
-; mm3 = temp
-; mm4 = temp
-; mm5 = temp 
-; mm6 = 0
-; mm7 = temp
-
-
-align 32
-pix_abs16x16_mmx:
-	push				ebp					; save frame pointer
-	mov				ebp, esp
-
-	push				ebx					; Saves registers (called saves convention in
-	push				ecx					; x86 GCC it seems)
-	push				edx					; 
-	push				esi
-	push				edi
-		
-	pxor				mm0, mm0				; zero acculumators
-	pxor				mm1, mm1
-	pxor				mm6, mm6
-	mov				esi, [ebp+8]		; get pix1
-	mov				edi, [ebp+12]		; get pix2
-	mov				edx, [ebp+16]		; get lx
-	mov				ecx, [ebp+20]		; get rowsleft
-	jmp				.nextrow
-align 32
-
-.nextrow:
-	; First 8 bytes of the row
-	
-	movq				mm4, [edi]	; load first 8 bytes of pix2 row 
-	movq				mm5, [esi]	; load first 8 bytes of pix1 row
-	movq				mm3, mm4		; mm4 := abs(mm4-mm5)
-	movq				mm2,[esi+8]	; load last 8 bytes of pix1 row	
-	psubusb			mm4, mm5
-	movq				mm7,[edi+8]	; load last 8 bytes of pix2 row 		
-	psubusb			mm5, mm3
-	por				mm4, mm5
-	
-	; Last 8 bytes of the row	
-
-	movq				mm3, mm7		; mm7 := abs(mm7-mm2)
-	psubusb			mm7, mm2
-	psubusb			mm2, mm3
-	por				mm7, mm2
-	
-	; Now mm4 and mm7 have 16 absdiffs to add
-	
-	; First 8 bytes of the row2
-	
-	
-	add				edi, edx
-	movq				mm2, [edi]	; load first 8 bytes of pix2 row 
-	add				esi, edx
-	movq				mm5, [esi]	; load first 8 bytes of pix1 row
-	
-	
-	
-	movq				mm3, mm2		; mm2 := abs(mm2-mm5)
-	psubusb			mm2, mm5
-	movq				mm6,[esi+8]	; load last 8 bytes of pix1 row
-	psubusb			mm5, mm3
-	por				mm2, mm5
-		
-	; Last 8 bytes of the row2
-	
-	movq				mm5,[edi+8]	; load last 8 bytes of pix2 row 
-	
-	
-	movq				mm3, mm5		; mm5 := abs(mm5-mm6)
-	psubusb			mm5, mm6
-	psubusb			mm6, mm3
-	por				mm5, mm6
-
-	; Now mm2, mm4, mm5, mm7 have 32 absdiffs
-	
-	movq				mm3, mm7
-	
-	pxor				mm6, mm6		; Zero mm6
-		
-	punpcklbw		mm3, mm6		; Unpack to words and add
-	punpckhbw		mm7, mm6
-	paddusw			mm7, mm3
-	
-	movq				mm3, mm5
-	
-	punpcklbw		mm3, mm6		; Unpack to words and add
-	punpckhbw		mm5, mm6
-	paddusw			mm5, mm3
-		
-	paddusw			mm0, mm7		; Add to the acumulator (mm0)
-	paddusw			mm1, mm5		; Add to the acumulator (mm1)
-		
-	movq				mm3, mm4
-	
-	punpcklbw		mm3, mm6		; Unpack to words and add
-	punpckhbw		mm4, mm6
-	movq				mm5, mm2
-	paddusw			mm4, mm3
-
-	
-	
-	punpcklbw		mm5, mm6		; Unpack to words and add
-	punpckhbw		mm2, mm6
-	paddusw			mm2, mm5
-
-	; Loop termination
-	
-	add				esi, edx		; update pointers to next row
-	paddusw			mm0, mm4		; Add to the acumulator (mm0)
-	add				edi, edx		
-	sub				ecx,2
-	paddusw			mm1, mm2		; Add to the acumulator (mm1)
-	test				ecx, ecx		; check rowsleft
-	jnz				near .nextrow
-	
-	paddusw			mm0, mm1
-	movq				mm2, mm0		; Copy mm0 to mm2
-	psrlq				mm2, 32
-	paddusw			mm0, mm2		; Add 
-	movq				mm3, mm0
-	psrlq				mm3, 16
-	paddusw			mm0, mm3
-	movd				eax, mm0		; Store return value
-	and				eax, 0xffff
-
-	pop edi
-	pop esi	
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp							; restore stack pointer
-
-	;emms								; clear mmx registers
-	ret								; return
-
-global pix_abs16x16_sse
-
-; int  pix_abs16x16_mmx(unsigned char *pix1,unsigned char *pix2, int lx, int h);
-; esi = p1 (init:		blk1)
-; edi = p2 (init:		blk2)
-; ecx = rowsleft (init:	 h)
-; edx = lx;
-
-; mm0 = distance accumulators (4 words)
-; mm1 = distance accumulators (4 words) 
-; mm2 = temp 
-; mm3 = temp
-; mm4 = temp
-; mm5 = temp 
-; mm6 = temp
-; mm7 = temp
-
-
-align 32
-pix_abs16x16_sse:
-	push				ebp					; save frame pointer
-	mov				ebp, esp
-
-	push				ebx					; Saves registers (called saves convention in
-	push				ecx					; x86 GCC it seems)
-	push				edx					; 
-	push				esi
-	push				edi
-		
-	pxor				mm0, mm0				; zero acculumators
-	pxor				mm1, mm1
-	mov				esi, [ebp+8]		; get pix1
-	mov				edi, [ebp+12]		; get pix2
-	mov				edx, [ebp+16]		; get lx
-	mov				ecx, [ebp+20]		; get rowsleft
-	jmp				.next4row
-align 32
-
-.next4row:
-	; First row
-	
-	movq				mm4, [edi]		; load first 8 bytes of pix2 row 
-	movq				mm5, [edi+8]	; load last 8 bytes of pix2 row
-	psadbw			mm4, [esi]		; SAD of first 8 bytes
-	psadbw			mm5, [esi+8]	; SAD of last 8 bytes
-	paddw				mm0, mm4			; Add to acumulators
-	paddw				mm1, mm5
-		
-	; Second row	
-
-	add				edi, edx;
-	add				esi, edx;
-	
-	movq				mm6, [edi]		; load first 8 bytes of pix2 row 
-	movq				mm7, [edi+8]	; load last 8 bytes of pix2 row
-	psadbw			mm6, [esi]		; SAD of first 8 bytes
-	psadbw			mm7, [esi+8]	; SAD of last 8 bytes
-	paddw				mm0, mm6			; Add to acumulators
-	paddw				mm1, mm7
-		
-	; Third row
-	
-	add				edi, edx;
-	add				esi, edx;
-	
-	movq				mm4, [edi]		; load first 8 bytes of pix2 row 
-	movq				mm5, [edi+8]	; load last 8 bytes of pix2 row
-	psadbw			mm4, [esi]		; SAD of first 8 bytes
-	psadbw			mm5, [esi+8]	; SAD of last 8 bytes
-	paddw				mm0, mm4			; Add to acumulators
-	paddw				mm1, mm5
-		
-	; Fourth row	
-
-	add				edi, edx;
-	add				esi, edx;
-	
-	movq				mm6, [edi]		; load first 8 bytes of pix2 row 
-	movq				mm7, [edi+8]	; load last 8 bytes of pix2 row
-	psadbw			mm6, [esi]		; SAD of first 8 bytes
-	psadbw			mm7, [esi+8]	; SAD of last 8 bytes
-	paddw				mm0, mm6			; Add to acumulators
-	paddw				mm1, mm7
-	
-	; Loop termination
-	
-	add				esi, edx		; update pointers to next row
-	add				edi, edx		
-	sub				ecx,4
-	test				ecx, ecx		; check rowsleft
-	jnz				near .next4row
-	
-	paddd				mm0, mm1		; Sum acumulators
-	movd				eax, mm0		; Store return value
-
-	pop edi
-	pop esi	
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp							; restore stack pointer
-
-	;emms								; clear mmx registers
-	ret								; return
-		
-global pix_abs16x16_x2_mmx
-
-; int  pix_abs16x16_x2_mmx(unsigned char *pix1,unsigned char *pix2, int lx, int h);
-; esi = p1 (init:		blk1)
-; edi = p2 (init:		blk2)
-; ecx = rowsleft (init:	 h)
-; edx = lx;
-
-; mm0 = distance accumulators (4 words)
-; mm1 = distance accumulators (4 words) 
-; mm2 = temp 
-; mm3 = temp
-; mm4 = temp
-; mm5 = temp 
-; mm6 = 0
-; mm7 = temp
-
-
-align 32
-pix_abs16x16_x2_mmx:
-	push				ebp					; save frame pointer
-	mov				ebp, esp
-
-	push				ebx					; Saves registers (called saves convention in
-	push				ecx					; x86 GCC it seems)
-	push				edx					; 
-	push				esi
-	push				edi
-		
-	pxor				mm0, mm0				; zero acculumators
-	pxor				mm1, mm1
-	pxor				mm6, mm6
-	mov				esi, [ebp+8]		; get pix1
-	mov				edi, [ebp+12]		; get pix2
-	mov				edx, [ebp+16]		; get lx
-	mov				ecx, [ebp+20]		; get rowsleft
-	jmp				.nextrow_x2
-align 32
-
-.nextrow_x2:
-	; First 8 bytes of the row
-	
-	movq				mm4, [edi]			; load first 8 bytes of pix2 row 
-	movq				mm5, [edi+1]		; load bytes 1-8 of pix2 row
-
-	movq				mm2, mm4		; copy mm4 on mm2
-	movq				mm3, mm5		; copy mm5 on mm3
-	punpcklbw		mm4, mm6		; first 4 bytes of [edi] on mm4
-	punpcklbw		mm5, mm6		; first 4 bytes of [edi+1] on mm5
-	paddusw			mm4, mm5		; mm4 := first 4 bytes interpolated in words
-	psrlw				mm4, 1
-
-	punpckhbw		mm2, mm6		; last 4 bytes of [edi] on mm2
-	punpckhbw		mm3, mm6		; last 4 bytes of [edi+1] on mm3
-	paddusw			mm2, mm3		; mm2 := last 4 bytes interpolated in words
-	psrlw				mm2, 1
-	
-	packuswb			mm4, mm2 	; pack 8 bytes interpolated on mm4
-	movq				mm5,[esi]	; load first 8 bytes of pix1 row
-	
-	movq				mm3, mm4		; mm4 := abs(mm4-mm5)
-	psubusb			mm4, mm5
-	psubusb			mm5, mm3
-	por				mm4, mm5
-		
-	; Last 8 bytes of the row	
-
-	movq	mm7, [edi+8]			; load last 8 bytes of pix2 row 
-	movq	mm5, [edi+9]			; load bytes 10-17 of pix2 row
-
-	movq				mm2, mm7		; copy mm7 on mm2
-	movq				mm3, mm5		; copy mm5 on mm3
-	punpcklbw		mm7, mm6		; first 4 bytes of [edi+8] on mm7
-	punpcklbw		mm5, mm6		; first 4 bytes of [edi+9] on mm5
-	paddusw			mm7, mm5		; mm1 := first 4 bytes interpolated in words
-	psrlw				mm7, 1
-
-	punpckhbw		mm2, mm6		; last 4 bytes of [edi] on mm2
-	punpckhbw		mm3, mm6		; last 4 bytes of [edi+1] on mm3
-	paddusw			mm2, mm3		; mm2 := last 4 bytes interpolated in words
-	psrlw				mm2, 1
-	
-	packuswb			mm7, mm2 	; pack 8 bytes interpolated on mm1
-	movq				mm5,[esi+8]	; load last 8 bytes of pix1 row
-	
-	movq				mm3, mm7		; mm7 := abs(mm1-mm5)
-	psubusb			mm7, mm5
-	psubusb			mm5, mm3
-	por				mm7, mm5
-	
-	; Now mm4 and mm7 have 16 absdiffs to add
-	
-	movq				mm3, mm4		; Make copies of these bytes
-	movq				mm2, mm7
-		
-	punpcklbw		mm4, mm6		; Unpack to words and add
-	punpcklbw		mm7, mm6
-	paddusw			mm4, mm7
-	paddusw			mm0, mm4		; Add to the acumulator (mm0)
-	
-	punpckhbw		mm3, mm6		; Unpack to words and add
-	punpckhbw		mm2, mm6
-	paddusw			mm3, mm2
-	paddusw			mm1, mm3		; Add to the acumulator (mm1)
-	
-	; Loop termination
-	
-	add				esi, edx		; update pointers to next row
-	add				edi, edx		
-			
-	sub				ecx,1
-	test				ecx, ecx		; check rowsleft
-	jnz				near .nextrow_x2
-	
-	paddusw			mm0, mm1 
-	
-	movq				mm1, mm0		; Copy mm0 to mm1
-	psrlq				mm1, 32
-	paddusw			mm0, mm1		; Add 
-	movq				mm2, mm0
-	psrlq				mm2, 16
-	paddusw			mm0, mm2
-	movd				eax, mm0		; Store return value
-	and				eax, 0xffff
-
-	pop edi
-	pop esi	
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp							; restore stack pointer
-
-	emms								; clear mmx registers
-	ret								; return
-	
-global pix_abs16x16_y2_mmx
-
-; int  pix_abs16x16_y2_mmx(unsigned char *pix1,unsigned char *pix2, int lx, int h);
-; esi = p1 (init:		blk1)
-; edi = p2 (init:		blk2)
-; ebx = p2 + lx
-; ecx = rowsleft (init:	 h)
-; edx = lx;
-
-; mm0 = distance accumulators (4 words)
-; mm1 = distance accumulators (4 words) 
-; mm2 = temp 
-; mm3 = temp
-; mm4 = temp
-; mm5 = temp 
-; mm6 = 0
-; mm7 = temp
-
-
-align 32
-pix_abs16x16_y2_mmx:
-	push				ebp					; save frame pointer
-	mov				ebp, esp
-
-	push				ebx					; Saves registers (called saves convention in
-	push				ecx					; x86 GCC it seems)
-	push				edx					; 
-	push				esi
-	push				edi
-		
-	pxor				mm0, mm0				; zero acculumators
-	pxor				mm1, mm1
-	pxor				mm6, mm6
-	mov				esi, [ebp+8]		; get pix1
-	mov				edi, [ebp+12]		; get pix2
-	mov				edx, [ebp+16]		; get lx
-	mov				ecx, [ebp+20]		; get rowsleft
-	mov				ebx, edi
-	add				ebx, edx
-	jmp				.nextrow_y2
-align 32
-
-.nextrow_y2:
-	; First 8 bytes of the row
-	
-	movq				mm4, [edi]			; load first 8 bytes of pix2 row 
-	movq				mm5, [ebx]			; load bytes 1-8 of pix2 row
-
-	movq				mm2, mm4		; copy mm4 on mm2
-	movq				mm3, mm5		; copy mm5 on mm3
-	punpcklbw		mm4, mm6		; first 4 bytes of [edi] on mm4
-	punpcklbw		mm5, mm6		; first 4 bytes of [ebx] on mm5
-	paddusw			mm4, mm5		; mm4 := first 4 bytes interpolated in words
-	psrlw				mm4, 1
-
-	punpckhbw		mm2, mm6		; last 4 bytes of [edi] on mm2
-	punpckhbw		mm3, mm6		; last 4 bytes of [edi+1] on mm3
-	paddusw			mm2, mm3		; mm2 := last 4 bytes interpolated in words
-	psrlw				mm2, 1
-	
-	packuswb			mm4, mm2 	; pack 8 bytes interpolated on mm4
-	movq				mm5,[esi]	; load first 8 bytes of pix1 row
-	
-	movq				mm3, mm4		; mm4 := abs(mm4-mm5)
-	psubusb			mm4, mm5
-	psubusb			mm5, mm3
-	por				mm4, mm5
-		
-	; Last 8 bytes of the row	
-
-	movq	mm7, [edi+8]			; load last 8 bytes of pix2 row 
-	movq	mm5, [ebx+8]			; load bytes 10-17 of pix2 row
-
-	movq				mm2, mm7		; copy mm7 on mm2
-	movq				mm3, mm5		; copy mm5 on mm3
-	punpcklbw		mm7, mm6		; first 4 bytes of [edi+8] on mm7
-	punpcklbw		mm5, mm6		; first 4 bytes of [ebx+8] on mm5
-	paddusw			mm7, mm5		; mm1 := first 4 bytes interpolated in words
-	psrlw				mm7, 1
-
-	punpckhbw		mm2, mm6		; last 4 bytes of [edi+8] on mm2
-	punpckhbw		mm3, mm6		; last 4 bytes of [ebx+8] on mm3
-	paddusw			mm2, mm3		; mm2 := last 4 bytes interpolated in words
-	psrlw				mm2, 1
-	
-	packuswb			mm7, mm2 	; pack 8 bytes interpolated on mm1
-	movq				mm5,[esi+8]	; load last 8 bytes of pix1 row
-	
-	movq				mm3, mm7		; mm7 := abs(mm1-mm5)
-	psubusb			mm7, mm5
-	psubusb			mm5, mm3
-	por				mm7, mm5
-	
-	; Now mm4 and mm7 have 16 absdiffs to add
-	
-	movq				mm3, mm4		; Make copies of these bytes
-	movq				mm2, mm7
-		
-	punpcklbw		mm4, mm6		; Unpack to words and add
-	punpcklbw		mm7, mm6
-	paddusw			mm4, mm7
-	paddusw			mm0, mm4		; Add to the acumulator (mm0)
-	
-	punpckhbw		mm3, mm6		; Unpack to words and add
-	punpckhbw		mm2, mm6
-	paddusw			mm3, mm2
-	paddusw			mm1, mm3		; Add to the acumulator (mm1)
-	
-	; Loop termination
-	
-	add				esi, edx		; update pointers to next row
-	add				edi, edx		
-	add				ebx, edx		
-	sub				ecx,1
-	test				ecx, ecx		; check rowsleft
-	jnz				near .nextrow_y2
-	
-	paddusw			mm0, mm1 
-	
-	movq				mm1, mm0		; Copy mm0 to mm1
-	psrlq				mm1, 32
-	paddusw			mm0, mm1		; Add 
-	movq				mm2, mm0
-	psrlq				mm2, 16
-	paddusw			mm0, mm2
-	movd				eax, mm0		; Store return value
-	and				eax, 0xffff
-
-	pop edi
-	pop esi	
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp							; restore stack pointer
-
-	emms								; clear mmx registers
-	ret								; return
-	
-global pix_abs16x16_xy2_mmx
-
-; int pix_abs16x16_xy2_mmx(unsigned char *p1,unsigned char *p2,int lx,int h);
-
-; esi = p1 (init:		blk1)
-; edi = p2 (init:		blk2)
-; ebx = p1+lx
-; ecx = rowsleft (init:	 h)
-; edx = lx;
-
-; mm0 = distance accumulators (4 words)
-; mm1 = bytes p2
-; mm2 = bytes p1
-; mm3 = bytes p1+lx
-; I'd love to find someplace to stash p1+1 and p1+lx+1's bytes
-; but I don't think thats going to happen in iA32-land...
-; mm4 = temp 4 bytes in words interpolating p1, p1+1
-; mm5 = temp 4 bytes in words from p2
-; mm6 = temp comparison bit mask p1,p2
-; mm7 = temp comparison bit mask p2,p1
-
-
-align 32
-pix_abs16x16_xy2_mmx:
-	push ebp		; save stack pointer
-	mov ebp, esp	; so that we can do this
-
-	push ebx		; Saves registers (called saves convention in
-	push ecx		; x86 GCC it seems)
-	push edx		; 
-	push esi
-	push edi
-		
-	pxor mm0, mm0				; zero acculumators
-
-	mov esi, [ebp+12]			; get p1
-	mov edi, [ebp+8]			; get p2
-	mov edx, [ebp+16]			; get lx
-	mov ecx, [ebp+20]			; rowsleft := h
-	mov ebx, esi
-    add ebx, edx		
-	jmp .nextrowmm11					; snap to it
-align 32
-.nextrowmm11:
-
-		;; 
-		;; First 8 bytes of row
-		;; 
-		
-		;; First 4 bytes of 8
-
-	movq mm4, [esi]             ; mm4 := first 4 bytes p1
-	pxor mm7, mm7
-	movq mm2, mm4				;  mm2 records all 8 bytes
-	punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
-	
-	movq mm6, [ebx]			    ;  mm6 := first 4 bytes p1+lx
-	movq mm3, mm6               ;  mm3 records all 8 bytes
-	punpcklbw mm6, mm7
-	paddw mm4, mm6              
-
-
-	movq mm5, [esi+1]			; mm5 := first 4 bytes p1+1
-	punpcklbw mm5, mm7            ;  First 4 bytes p1 in Words...
-	paddw mm4, mm5		
-	movq mm6, [ebx+1]           ;  mm6 := first 4 bytes p1+lx+1
-	punpcklbw mm6, mm7
-	paddw mm4, mm6
-
-	psrlw mm4, 2	            ; mm4 := First 4 bytes interpolated in words
-		
-	movq mm5, [edi]				; mm5:=first 4 bytes of p2 in words
-	movq mm1, mm5
-	punpcklbw mm5, mm7
-			
-	movq  mm7,mm4
-	pcmpgtw mm7,mm5		; mm7 := [i : W0..3,mm4>mm5]
-
-	movq  mm6,mm4		; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
- 	psubw mm6,mm5
-	pand  mm6, mm7
-
-	paddw mm0, mm6				; Add to accumulator
-
-	movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
-	pcmpgtw mm6,mm4	    
- 	psubw mm5,mm4		; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
-	pand  mm5, mm6		
-
-	paddw mm0, mm5				; Add to accumulator
-
-		;; Second 4 bytes of 8
-	
-	movq mm4, mm2		    ; mm4 := Second 4 bytes p1 in words
-	pxor  mm7, mm7
-	punpckhbw mm4, mm7			
-	movq mm6, mm3			; mm6 := Second 4 bytes p1+1 in words  
-	punpckhbw mm6, mm7
-	paddw mm4, mm6          
-
-	movq mm5, [esi+1]			; mm5 := first 4 bytes p1+1
-	punpckhbw mm5, mm7          ;  First 4 bytes p1 in Words...
-	paddw mm4, mm5
-	movq mm6, [ebx+1]           ;  mm6 := first 4 bytes p1+lx+1
-	punpckhbw mm6, mm7
-	paddw mm4, mm6
-
-	psrlw mm4, 2	            ; mm4 := First 4 bytes interpolated in words
-		
-	movq mm5, mm1			; mm5:= second 4 bytes of p2 in words
-	punpckhbw mm5, mm7
-			
-	movq  mm7,mm4
-	pcmpgtw mm7,mm5		; mm7 := [i : W0..3,mm4>mm5]
-
-	movq  mm6,mm4		; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
- 	psubw mm6,mm5
-	pand  mm6, mm7
-
-	paddw mm0, mm6				; Add to accumulator
-
-	movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
-	pcmpgtw mm6,mm4	    
- 	psubw mm5,mm4		; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
-	pand  mm5, mm6		
-
-	paddw mm0, mm5				; Add to accumulator
-
-
- 		;;
-		;; Second 8 bytes of row
-		;; 
-		;; First 4 bytes of 8
-
-	movq mm4, [esi+8]             ; mm4 := first 4 bytes p1+8
-	pxor mm7, mm7
-	movq mm2, mm4				;  mm2 records all 8 bytes
-	punpcklbw mm4, mm7            ;  First 4 bytes p1 in Words...
-	
-	movq mm6, [ebx+8]			    ;  mm6 := first 4 bytes p1+lx+8
-	movq mm3, mm6               ;  mm3 records all 8 bytes
-	punpcklbw mm6, mm7
-	paddw mm4, mm6              
-
-
-	movq mm5, [esi+9]			; mm5 := first 4 bytes p1+9
-	punpcklbw mm5, mm7            ;  First 4 bytes p1 in Words...
-	paddw mm4, mm5
-	movq mm6, [ebx+9]           ;  mm6 := first 4 bytes p1+lx+9
-	punpcklbw mm6, mm7
-	paddw mm4, mm6
-
-	psrlw mm4, 2	            ; mm4 := First 4 bytes interpolated in words
-		
-	movq mm5, [edi+8]				; mm5:=first 4 bytes of p2+8 in words
-	movq mm1, mm5
-	punpcklbw mm5, mm7
-			
-	movq  mm7,mm4
-	pcmpgtw mm7,mm5		; mm7 := [i : W0..3,mm4>mm5]
-
-	movq  mm6,mm4		; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
- 	psubw mm6,mm5
-	pand  mm6, mm7
-
-	paddw mm0, mm6				; Add to accumulator
-
-	movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
-	pcmpgtw mm6,mm4	    
- 	psubw mm5,mm4		; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
-	pand  mm5, mm6		
-
-	paddw mm0, mm5				; Add to accumulator
-
-		;; Second 4 bytes of 8
-	
-	movq mm4, mm2		    ; mm4 := Second 4 bytes p1 in words
-	pxor  mm7, mm7
-	punpckhbw mm4, mm7			
-	movq mm6, mm3			; mm6 := Second 4 bytes p1+1 in words  
-	punpckhbw mm6, mm7
-	paddw mm4, mm6          
-
-	movq mm5, [esi+9]			; mm5 := first 4 bytes p1+1
-	punpckhbw mm5, mm7          ;  First 4 bytes p1 in Words...
-	paddw mm4, mm5	
-	movq mm6, [ebx+9]           ;  mm6 := first 4 bytes p1+lx+1
-	punpckhbw mm6, mm7
-	paddw mm4, mm6
-		
-	psrlw mm4, 2	            ; mm4 := First 4 bytes interpolated in words
-
-	movq mm5, mm1			; mm5:= second 4 bytes of p2 in words
-	punpckhbw mm5, mm7
-			
-	movq  mm7,mm4
-	pcmpgtw mm7,mm5		; mm7 := [i : W0..3,mm4>mm5]
-
-	movq  mm6,mm4		; mm6 := [i : W0..3, (mm4-mm5)*(mm4-mm5 > 0)]
- 	psubw mm6,mm5
-	pand  mm6, mm7
-
-	paddw mm0, mm6				; Add to accumulator
-
-	movq  mm6,mm5       ; mm6 := [i : W0..3,mm5>mm4]
-	pcmpgtw mm6,mm4	    
- 	psubw mm5,mm4		; mm5 := [i : B0..7, (mm5-mm4)*(mm5-mm4 > 0)]
-	pand  mm5, mm6		
-
-	paddw mm0, mm5				; Add to accumulator
-
-
-		;;
-		;;	Loop termination condition... and stepping
-		;;		
-
-	add esi, edx		; update pointer to next row
-	add edi, edx		; ditto
-	add ebx, edx
-
-	sub  ecx,1
-	test ecx, ecx		; check rowsleft
-	jnz near .nextrowmm11
-		
-		;; Sum the Accumulators
-	movq  mm4, mm0
-	psrlq mm4, 32
-	paddw mm0, mm4
-	movq  mm6, mm0
-	psrlq mm6, 16
-	paddw mm0, mm6
-	movd eax, mm0		; store return value
-	and  eax, 0xffff
-		
-	pop edi
-	pop esi	
-	pop edx			
-	pop ecx			
-	pop ebx			
-
-	pop ebp			; restore stack pointer
-
-	emms			; clear mmx registers
-	ret			; we now return you to your regular programming
-
-