comparison libpostproc/postprocess_template.c @ 2036:6a6c678517b3 libavcodec

altivec optimizations and horizontal filter fix by (Romain Dolbeau <dolbeau at irisa dot fr>)

2035:e1b69326ae36

/**
 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
 */
static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*3;
	asm volatile(	//"movv %0 %1 %2\n\t"

		src++;
	}
#endif
}

#if 0
/**
 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
 * values are correctly clipped (MMX2)

		src++;
	}
#endif
}

static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
/*
	uint8_t tmp[16];

		}
		src++;
	}
#endif
}

static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"pxor %%mm6, %%mm6				\n\t"

//		src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
	}
#endif
#endif
}

/**
 * Deinterlaces the given block by linearly interpolating every second line.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

#else
				if(mode & H_X1_FILTER)
					horizX1Filter(dstBlock-4, stride, QP);
				else if(mode & H_DEBLOCK)
				{
					if( isHorizDC(dstBlock-4, stride, &c))
					{
						if(isHorizMinMaxOk(dstBlock-4, stride, QP))
							doHorizLowPass(dstBlock-4, stride, QP);
					}
					else
						doHorizDefFilter(dstBlock-4, stride, QP);
				}
#endif
				if(mode & DERING)
				{
				//FIXME filter first line

2036:6a6c678517b3

/**
 * Do a vertical low pass filter on the 8x16 block (only write to the 8x8 block in the middle)
 * using the 9-Tap Filter (1,1,2,2,4,2,2,1,1)/16
 */
#ifndef HAVE_ALTIVEC
static inline void RENAME(doVertLowPass)(uint8_t *src, int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*3;
	asm volatile(	//"movv %0 %1 %2\n\t"

		src++;
	}
#endif
}
#endif //HAVE_ALTIVEC

#if 0
/**
 * Experimental implementation of the filter (Algorithm 1) described in a paper from Ramkishor & Karandikar
 * values are correctly clipped (MMX2)

		src++;
	}
#endif
}

#ifndef HAVE_ALTIVEC
static inline void RENAME(doVertDefFilter)(uint8_t src[], int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
/*
	uint8_t tmp[16];

		}
		src++;
	}
#endif
}
#endif //HAVE_ALTIVEC

#ifndef HAVE_ALTIVEC
static inline void RENAME(dering)(uint8_t src[], int stride, PPContext *c)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	asm volatile(
		"pxor %%mm6, %%mm6				\n\t"

//		src[0] = src[7]=src[stride*7]=src[stride*7 + 7]=255;
	}
#endif
#endif
}
#endif //HAVE_ALTIVEC

/**
 * Deinterlaces the given block by linearly interpolating every second line.
 * will be called for every 8x8 block and can read & write from line 4-15
 * lines 0-3 have been passed through the deblock / dering filters allready, but can be read too

#else
				if(mode & H_X1_FILTER)
					horizX1Filter(dstBlock-4, stride, QP);
				else if(mode & H_DEBLOCK)
				{
					const int t= RENAME(horizClassify)(dstBlock-4, stride, &c);

					if(t==1)
						RENAME(doHorizLowPass)(dstBlock-4, stride, &c);
					else if(t==2)
						RENAME(doHorizDefFilter)(dstBlock-4, stride, &c);
				}
#endif
				if(mode & DERING)
				{
				//FIXME filter first line