block.cpp

avs decoder sources, added a command line sample

				  + 5 * *(pbRefY + 
				          max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				          max(0, min(iImgWidth - 1, refX + j + 1)))
				  - *(pbRefY + 
				      max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				      max(0, min(iImgWidth - 1, refX + j + 2)));

				int Epie = 
				  *(pbRefY + 
				    max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				    max(0, min(iImgWidth - 1, refX + j + 1))) * 8;

				pPred[i * 8 + j] = 
					Clip((eepie + 7 * Dpie + 7 * bpie + Epie + 64) >> 7);
			}
		}
	}
	else if (dy == 0 && dx == 3) { //c
		for (i = 0; i < 8; i++) {
			for (j = 0; j < 8; j++) {
				int Dpie = 
				  *(pbRefY + 
				    max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				    max(0, min(iImgWidth - 1, refX + j))) * 8;

				int Epie = 
				  *(pbRefY + 
				    max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				    max(0, min(iImgWidth - 1, refX + j + 1))) * 8;

				int bpie = 
				  - *(pbRefY + 
				      max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				      max(0, min(iImgWidth - 1, refX + j - 1))) 
				  + 5 * *(pbRefY + 
				          max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				         max(0, min(iImgWidth - 1, refX + j))) 
				  + 5 * *(pbRefY + 
				          max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				          max(0, min(iImgWidth - 1, refX + j + 1)))
				  - *(pbRefY + 
				      max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				      max(0, min(iImgWidth - 1, refX + j + 2)));
				int ccpie = 
				  - *(pbRefY + 
				      max(0, min(iImgHeight - 1, refY + i)) * iImgWidth + 
				      max(0, min(iImgWidth - 1, refX + j))) 
				  + 5* *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))) 
					+ 5* *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2)))
					-    *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j+3)));
				pPred[i*8+j] = Clip((Dpie + 7*bpie + 7*Epie + ccpie + 64)>>7);
			}
		}
	}
	else if(dy==1 && dx==0) //d
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int ffpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-2))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)));

				int Dpie = *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j))) * 8;
				int hpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j)));
				int Hpie = *(pbRefY + max(0,min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j))) * 8; 
				int dpie = ffpie + 7*Dpie + 7*hpie + Hpie;
				pPred[i*8+j] = Clip((dpie + 64)>>7);        
			}
		}
	}
	else if(dy==1 && dx==1) //e
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int aapie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int bpie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int spie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int ddpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int jpie = -aapie + 5*bpie + 5*spie - ddpie;

				int Dpiepie = *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j))) * 64;

				pPred[i*8+j] = Clip((Dpiepie+jpie+64)>>7);
			}
		}
	}
	else if(dy==1 && dx==2) //f
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int xxpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int aapie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int bpie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int spie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int ddpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int hhpie = -xxpie + 5*aapie + 5*bpie - spie;
				int jpie = -aapie + 5*bpie + 5*spie - ddpie;

				pPred[i*8+j] = Clip((hhpie+7*bpie*8+7*jpie+spie*8+512)>>10);      
			}
		}
	}
	else if(dy==1 && dx==3) //g
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int aapie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int bpie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int spie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int ddpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int jpie = -aapie + 5*bpie + 5*spie - ddpie;
				int Epiepie = *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))) * 64;
				pPred[i*8+j] = Clip((Epiepie+jpie+64)>>7);
			}
		}
	}
	else if(dy==2 && dx==1) //i
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int aapie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int bpie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));

				int spie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int ddpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2))));
				int jpie = -aapie + 5*bpie + 5*spie - ddpie;

				int aaxxpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-2))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))));
				int eepie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-2))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))));

				int sxxpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-2))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))));
				int ddxxpie = 
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-2))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j  ))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1))));
				int ggpie = -aaxxpie + 5*eepie + 5*sxxpie - ddxxpie;
				int hpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j)));
				int mpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)));
				pPred[i*8+j] = Clip((ggpie+7*hpie*8+7*jpie+mpie*8+512)>>10);
			}
		}
	}
	else if(dy==2 && dx==3) //k
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int bbpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j-1)));
				int hpie =  
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j)));
				int mpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+1)));
				int ccpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+2)));
				int jpie = -bbpie + 5*hpie + 5*mpie - ccpie;

				int xxpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+3)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j+3)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j+3)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j+3)));
				int kkpie = -hpie + 5*mpie + 5*ccpie - xxpie;
				pPred[i*8+j] = Clip((hpie*8+7*jpie+7*mpie*8+kkpie+512)>>10);

			}
		}
	}
	else if(dy==3 && dx==0) //n
	{
		for(i=0; i<8; i++)
		{
			for(j=0; j<8; j++)
			{
				int Dpie = *(pbRefY + max(0,min(iImgHeight-1, refY+i))*iImgWidth + max(0, min(iImgWidth-1, refX+j))) * 8;
				int hpie =  
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i-1))*iImgWidth + max(0, min(iImgWidth-1, refX+j))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j))))
					+ 5* (*(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j))))
					-    (*(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j))));
				int Hpie = *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j))) * 8;  
				int xxpie = 
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i  ))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+1))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					+ 5* *(pbRefY + max(0, min(iImgHeight-1, refY+i+2))*iImgWidth + max(0, min(iImgWidth-1, refX+j)))
					-    *(pbRefY + max(0, min(iImgHeight-1, refY+i+3))*iImgWidth + max(0, min(iImgWidth-1, refX+j)));