predframe = forward_reference_frame;

				form_prediction(predframe, motion_vertical_field_select[1][0], current_frame, 
					0, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 8, bx, by+8,
					PMV[1][0][0], PMV[1][0][1], stw);
			}
			else if (motion_type==MC_DMV)
			{
				if (Second_Field)
					predframe = backward_reference_frame;
				else
					predframe = forward_reference_frame;

				/* calculate derived motion vectors */
				Dual_Prime_Arithmetic(DMV, dmvector, PMV[0][0][0], PMV[0][0][1]);

				/* predict from field of same parity */
				form_prediction(forward_reference_frame, currentfield, current_frame, 0, 
					Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 16, bx, by,
					PMV[0][0][0], PMV[0][0][1], 0);

				/* predict from field of opposite parity */
				form_prediction(predframe, !currentfield, current_frame, 0,
					Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 16, bx, by,
					DMV[0][0], DMV[0][1], 1);
			}
			else
				Fault_Flag = 5;
		}

		stw = 1;
	}

	if (macroblock_type & MACROBLOCK_MOTION_BACKWARD)
	{
		if (picture_structure==FRAME_PICTURE)
		{
			if (motion_type==MC_FRAME)
			{
				/* frame-based prediction */
				form_prediction(backward_reference_frame, 0, current_frame, 0,
					Coded_Picture_Width, Coded_Picture_Width<<1, 16, 8, bx, by,
					PMV[0][1][0], PMV[0][1][1], stw);

				form_prediction(backward_reference_frame, 1, current_frame, 1,
					Coded_Picture_Width, Coded_Picture_Width<<1, 16, 8, bx, by,
					PMV[0][1][0], PMV[0][1][1], stw);
			}
			else /* field-based prediction */
			{
				/* top field prediction */
				form_prediction(backward_reference_frame, motion_vertical_field_select[0][1], 
					current_frame, 0, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 8,
					bx, by>>1, PMV[0][1][0], PMV[0][1][1]>>1, stw);

				/* bottom field prediction */
				form_prediction(backward_reference_frame, motion_vertical_field_select[1][1], 
					current_frame, 1, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 8,
					bx, by>>1, PMV[1][1][0], PMV[1][1][1]>>1, stw);
			}
		}
		else
		{
			/* field picture */
			if (motion_type==MC_FIELD)
			{
				/* field-based prediction */
				form_prediction(backward_reference_frame, motion_vertical_field_select[0][1], 
					current_frame, 0, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 16,
					bx, by, PMV[0][1][0], PMV[0][1][1], stw);
			}
			else if (motion_type==MC_16X8)
			{
				form_prediction(backward_reference_frame, motion_vertical_field_select[0][1],
					current_frame, 0, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 8,
					bx, by, PMV[0][1][0], PMV[0][1][1], stw);

				form_prediction(backward_reference_frame, motion_vertical_field_select[1][1],
					current_frame, 0, Coded_Picture_Width<<1, Coded_Picture_Width<<1, 16, 8,
					bx, by+8, PMV[1][1][0], PMV[1][1][1], stw);
			}
			else
				Fault_Flag = 5;
		}
	}
}

static void form_prediction(unsigned char *src[], int sfield, unsigned char *dst[],
							int dfield, int lx, int lx2, int w, int h, int x, int y,
							int dx, int dy, int average_flag)
{
	form_component_prediction(src[0]+(sfield?lx2>>1:0), dst[0]+(dfield?lx2>>1:0),
		lx, lx2, w, h, x, y, dx, dy, average_flag);

	if (chroma_format!=CHROMA444)
	{
		lx>>=1; lx2>>=1; w>>=1; x>>=1; dx/=2;
	}

	if (chroma_format==CHROMA420)
	{
		h>>=1; y>>=1; dy/=2;
	}

	/* Cb */
	form_component_prediction(src[1]+(sfield?lx2>>1:0), dst[1]+(dfield?lx2>>1:0),
		lx, lx2, w, h, x, y, dx, dy, average_flag);

	/* Cr */
	form_component_prediction(src[2]+(sfield?lx2>>1:0), dst[2]+(dfield?lx2>>1:0),
		lx, lx2, w, h, x, y, dx, dy, average_flag);
}

/* ISO/IEC 13818-2 section 7.6.4: Forming predictions */
static void form_component_prediction(unsigned char *src, unsigned char *dst,
										  int lx, int lx2, int w, int h, int x, int y,
										  int dx, int dy, int average_flag)
{
	static const __int64 mmmask_0001 = 0x0001000100010001;
	static const __int64 mmmask_0002 = 0x0002000200020002;
	static const __int64 mmmask_0003 = 0x0003000300030003;
	static const __int64 mmmask_0006 = 0x0006000600060006;

	unsigned char *s = src + lx * (y + (dy>>1)) + x + (dx>>1);
	unsigned char *d = dst + lx * y + x;
	int flag = (average_flag<<2) + ((dx & 1)<<1) + (dy & 1);

	switch (flag)
	{
		case 0:
			// d[i] = s[i];
			__asm
			{
				mov			eax, [s]
				mov			ebx, [d]
				mov			esi, 0x00
				mov			edi, [h]
mc0:
				movq		mm1, [eax+esi]
				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc0

				add			eax, [lx2]
				add			ebx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc0
			}
			break;

		case 1:
			// d[i] = (s[i]+s[i+lx]+1)>>1;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0001]
				mov			eax, [s]
				mov			ebx, [d]
				mov			ecx, eax
				add			ecx, [lx]
				mov			esi, 0x00
				mov			edi, [h]
mc1:
				movq		mm1, [eax+esi]
				movq		mm2, [ecx+esi]

				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0
				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				paddsw		mm1, mm7
				paddsw		mm3, mm7

				psrlw		mm1, 1
				psrlw		mm3, 1

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc1

				add			eax, [lx2]
				add			ebx, [lx2]
				add			ecx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc1
			}
			break;

		case 2:
			// d[i] = (s[i]+s[i+1]+1)>>1;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0001]
				mov			eax, [s]
				mov			ebx, [d]
				mov			esi, 0x00
				mov			edi, [h]
mc2:
				movq		mm1, [eax+esi]
				movq		mm2, [eax+esi+1]

				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				paddsw		mm1, mm7
				paddsw		mm3, mm7

				psrlw		mm1, 1
				psrlw		mm3, 1

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc2

				add			eax, [lx2]
				add			ebx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc2
			}
			break;

		case 3:
			// d[i] = (s[i]+s[i+1]+s[i+lx]+s[i+lx+1]+2)>>2;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0002]
				mov			eax, [s]
				mov			ebx, [d]
				mov			ecx, eax
				add			ecx, [lx]
				mov			esi, 0x00
				mov			edi, [h]
mc3:
				movq		mm1, [eax+esi]
				movq		mm2, [eax+esi+1]
				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				movq		mm5, [ecx+esi]
				paddsw		mm1, mm7

				movq		mm6, [ecx+esi+1]
				paddsw		mm3, mm7

				movq		mm2, mm5
				movq		mm4, mm6

				punpcklbw	mm2, mm0
				punpckhbw	mm5, mm0

				punpcklbw	mm4, mm0
				punpckhbw	mm6, mm0
				
				paddsw		mm2, mm4
				paddsw		mm5, mm6

				paddsw		mm1, mm2
				paddsw		mm3, mm5

				psrlw		mm1, 2
				psrlw		mm3, 2

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc3

				add			eax, [lx2]
				add			ebx, [lx2]
				add			ecx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc3
			}
			break;

		case 4:
			// d[i] = (s[i]+d[i]+1)>>1;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0001]
				mov			eax, [s]
				mov			ebx, [d]
				mov			esi, 0x00
				mov			edi, [h]
mc4:
				movq		mm1, [eax+esi]
				movq		mm2, [ebx+esi]
				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				paddsw		mm1, mm7
				paddsw		mm3, mm7

				psrlw		mm1, 1
				psrlw		mm3, 1

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc4

				add			eax, [lx2]
				add			ebx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc4
			}
			break;

		case 5:
			// d[i] = ((d[i]<<1) + s[i]+s[i+lx] + 3)>>2;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0003]
				mov			eax, [s]
				mov			ebx, [d]
				mov			ecx, eax
				add			ecx, [lx]
				mov			esi, 0x00
				mov			edi, [h]
mc5:
				movq		mm1, [eax+esi]
				movq		mm2, [ecx+esi]
				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				movq		mm5, [ebx+esi]

				paddsw		mm1, mm7
				paddsw		mm3, mm7

				movq		mm6, mm5
				punpcklbw	mm5, mm0
				punpckhbw	mm6, mm0

				psllw		mm5, 1
				psllw		mm6, 1

				paddsw		mm1, mm5
				paddsw		mm3, mm6

				psrlw		mm1, 2
				psrlw		mm3, 2

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc5

				add			eax, [lx2]
				add			ebx, [lx2]
				add			ecx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc5
			}
			break;

		case 6:
			// d[i] = ((d[i]<<1) + s[i]+s[i+1] + 3) >> 2;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0003]
				mov			eax, [s]
				mov			ebx, [d]
				mov			esi, 0x00
				mov			edi, [h]
mc6:
				movq		mm1, [eax+esi]
				movq		mm2, [eax+esi+1]
				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				movq		mm5, [ebx+esi]

				paddsw		mm1, mm7
				paddsw		mm3, mm7

				movq		mm6, mm5
				punpcklbw	mm5, mm0
				punpckhbw	mm6, mm0

				psllw		mm5, 1
				psllw		mm6, 1

				paddsw		mm1, mm5
				paddsw		mm3, mm6

				psrlw		mm1, 2
				psrlw		mm3, 2

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc6

				add			eax, [lx2]
				add			ebx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc6
			}
			break;

		case 7:
			// d[i] = ((d[i]<<2) + s[i]+s[i+1]+s[i+lx]+s[i+lx+1] + 6)>>3;
			__asm
			{
				pxor		mm0, mm0
				movq		mm7, [mmmask_0006]
				mov			eax, [s]
				mov			ebx, [d]
				mov			ecx, eax
				add			ecx, [lx]
				mov			esi, 0x00
				mov			edi, [h]
mc7:
				movq		mm1, [eax+esi]
				movq		mm2, [eax+esi+1]
				movq		mm3, mm1
				movq		mm4, mm2

				punpcklbw	mm1, mm0
				punpckhbw	mm3, mm0

				punpcklbw	mm2, mm0
				punpckhbw	mm4, mm0

				paddsw		mm1, mm2
				paddsw		mm3, mm4

				movq		mm5, [ecx+esi]
				paddsw		mm1, mm7

				movq		mm6, [ecx+esi+1]
				paddsw		mm3, mm7

				movq		mm2, mm5
				movq		mm4, mm6

				punpcklbw	mm2, mm0
				punpckhbw	mm5, mm0

				punpcklbw	mm4, mm0
				punpckhbw	mm6, mm0
				
				paddsw		mm2, mm4
				paddsw		mm5, mm6

				paddsw		mm1, mm2
				paddsw		mm3, mm5

				movq		mm6, [ebx+esi]

				movq		mm4, mm6
				punpcklbw	mm4, mm0
				punpckhbw	mm6, mm0

				psllw		mm4, 2
				psllw		mm6, 2

				paddsw		mm1, mm4
				paddsw		mm3, mm6

				psrlw		mm1, 3
				psrlw		mm3, 3

				packuswb	mm1, mm0
				packuswb	mm3, mm0

				psllq		mm3, 32
				por			mm1, mm3

				add			esi, 0x08
				cmp			esi, [w]
				movq		[ebx+esi-8], mm1
				jl			mc7

				add			eax, [lx2]
				add			ebx, [lx2]
				add			ecx, [lx2]
				sub			edi, 0x01
				mov			esi, 0x00
				cmp			edi, 0x00
				jg			mc7
			}
			break;
	}
}