postprocess_template.c

这是著名的TCPMP播放器在WINDWOWS,和WINCE下编译通过的源程序.笔者对其中的LIBMAD库做了针对ARM MPU的优化. 并增加了词幕功能.

		"pxor %%mm2, %%mm5				\n\t" // -line 4 - 1
		PAVGB(%%mm3, %%mm5)
		"paddb %%mm6, %%mm5				\n\t" // (l5-l4)/2
		"psubusb %%mm3, %%mm4				\n\t"
		"psubusb %%mm2, %%mm3				\n\t"
		"por %%mm3, %%mm4				\n\t" // |l4 - l5|
		"psubusb %%mm0, %%mm4				\n\t"
		"pcmpeqb %%mm7, %%mm4				\n\t"
		"pand %%mm4, %%mm5				\n\t" // d/2

//		"paddb %%mm6, %%mm2				\n\t" // line 4 + 0x80
		"paddb %%mm5, %%mm2				\n\t"
//		"psubb %%mm6, %%mm2				\n\t"
		"movq %%mm2, (%0,%1, 4)				\n\t"

		"movq (%%"REG_c"), %%mm2				\n\t"
//		"paddb %%mm6, %%mm2				\n\t" // line 5 + 0x80
		"psubb %%mm5, %%mm2				\n\t"
//		"psubb %%mm6, %%mm2				\n\t"
		"movq %%mm2, (%%"REG_c")				\n\t"

		"paddb %%mm6, %%mm5				\n\t"
		"psrlw $2, %%mm5				\n\t"
		"pand "MANGLE(b3F)", %%mm5			\n\t"
		"psubb "MANGLE(b20)", %%mm5			\n\t" // (l5-l4)/8

		"movq (%%"REG_a", %1, 2), %%mm2			\n\t"
		"paddb %%mm6, %%mm2				\n\t" // line 3 + 0x80
		"paddsb %%mm5, %%mm2				\n\t"
		"psubb %%mm6, %%mm2				\n\t"
		"movq %%mm2, (%%"REG_a", %1, 2)			\n\t"

		"movq (%%"REG_c", %1), %%mm2			\n\t"
		"paddb %%mm6, %%mm2				\n\t" // line 6 + 0x80
		"psubsb %%mm5, %%mm2				\n\t"
		"psubb %%mm6, %%mm2				\n\t"
		"movq %%mm2, (%%"REG_c", %1)			\n\t"

		:
		: "r" (src), "r" ((long)stride)
		: "%"REG_a, "%"REG_c
	);
#else
 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;
	const int l6= stride + l5;
//	const int l7= stride + l6;
//	const int l8= stride + l7;
//	const int l9= stride + l8;
	int x;
	const int QP15= QP + (QP>>2);
	src+= stride*3;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		const int v = (src[x+l5] - src[x+l4]);
		if(ABS(v) < QP15)
		{
			src[x+l3] +=v>>3;
			src[x+l4] +=v>>1;
			src[x+l5] -=v>>1;
			src[x+l6] -=v>>3;

		}
	}

#endif
}
#endif

/**
 * Experimental Filter 1
 * will not damage linear gradients
 * Flat blocks should look like they where passed through the (1,1,2,2,4,2,2,1,1) 9-Tap filter
 * can only smooth blocks at the expected locations (it cant smooth them if they did move)
 * MMX2 version does correct clipping C version doesnt
 */
static inline void RENAME(vertX1Filter)(uint8_t *src, int stride, PPContext *co)
{
#if defined (HAVE_MMX2) || defined (HAVE_3DNOW)
	src+= stride*3;

	asm volatile(
		"pxor %%mm7, %%mm7				\n\t" // 0
		"lea (%0, %1), %%"REG_a"			\n\t"
		"lea (%%"REG_a", %1, 4), %%"REG_c"		\n\t"
//	0	1	2	3	4	5	6	7	8	9
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1	%0+8%1	ecx+4%1
		"movq (%%"REG_a", %1, 2), %%mm0			\n\t" // line 3
		"movq (%0, %1, 4), %%mm1			\n\t" // line 4
		"movq %%mm1, %%mm2				\n\t" // line 4
		"psubusb %%mm0, %%mm1				\n\t"
		"psubusb %%mm2, %%mm0				\n\t"
		"por %%mm1, %%mm0				\n\t" // |l2 - l3|
		"movq (%%"REG_c"), %%mm3				\n\t" // line 5
		"movq (%%"REG_c", %1), %%mm4			\n\t" // line 6
		"movq %%mm3, %%mm5				\n\t" // line 5
		"psubusb %%mm4, %%mm3				\n\t"
		"psubusb %%mm5, %%mm4				\n\t"
		"por %%mm4, %%mm3				\n\t" // |l5 - l6|
		PAVGB(%%mm3, %%mm0)				      // (|l2 - l3| + |l5 - l6|)/2
		"movq %%mm2, %%mm1				\n\t" // line 4
		"psubusb %%mm5, %%mm2				\n\t"
		"movq %%mm2, %%mm4				\n\t"
		"pcmpeqb %%mm7, %%mm2				\n\t" // (l4 - l5) <= 0 ? -1 : 0
		"psubusb %%mm1, %%mm5				\n\t"
		"por %%mm5, %%mm4				\n\t" // |l4 - l5|
		"psubusb %%mm0, %%mm4		\n\t" //d = MAX(0, |l4-l5| - (|l2-l3| + |l5-l6|)/2)
		"movq %%mm4, %%mm3				\n\t" // d
		"movq %2, %%mm0			\n\t"
                "paddusb %%mm0, %%mm0				\n\t"
		"psubusb %%mm0, %%mm4				\n\t"
		"pcmpeqb %%mm7, %%mm4				\n\t" // d <= QP ? -1 : 0
		"psubusb "MANGLE(b01)", %%mm3			\n\t"
		"pand %%mm4, %%mm3				\n\t" // d <= QP ? d : 0

		PAVGB(%%mm7, %%mm3)				      // d/2
		"movq %%mm3, %%mm1				\n\t" // d/2
		PAVGB(%%mm7, %%mm3)				      // d/4
		PAVGB(%%mm1, %%mm3)				      // 3*d/8

		"movq (%0, %1, 4), %%mm0			\n\t" // line 4
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
		"psubusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%0, %1, 4)			\n\t" // line 4

		"movq (%%"REG_c"), %%mm0			\n\t" // line 5
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
		"paddusb %%mm3, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%"REG_c")			\n\t" // line 5

		PAVGB(%%mm7, %%mm1)				      // d/4

		"movq (%%"REG_a", %1, 2), %%mm0			\n\t" // line 3
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l4-1 : l4
		"psubusb %%mm1, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%"REG_a", %1, 2)			\n\t" // line 3

		"movq (%%"REG_c", %1), %%mm0			\n\t" // line 6
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l5-1 : l5
		"paddusb %%mm1, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%"REG_c", %1)			\n\t" // line 6

		PAVGB(%%mm7, %%mm1)				      // d/8

		"movq (%%"REG_a", %1), %%mm0			\n\t" // line 2
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l2-1 : l2
		"psubusb %%mm1, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%"REG_a", %1)			\n\t" // line 2

		"movq (%%"REG_c", %1, 2), %%mm0			\n\t" // line 7
		"pxor %%mm2, %%mm0				\n\t" //(l4 - l5) <= 0 ? -l7-1 : l7
		"paddusb %%mm1, %%mm0				\n\t"
		"pxor %%mm2, %%mm0				\n\t"
		"movq %%mm0, (%%"REG_c", %1, 2)			\n\t" // line 7

		:
		: "r" (src), "r" ((long)stride), "m" (co->pQPb)
		: "%"REG_a, "%"REG_c
	);
#else

 	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= stride + l3;
	const int l5= stride + l4;
	const int l6= stride + l5;
	const int l7= stride + l6;
//	const int l8= stride + l7;
//	const int l9= stride + l8;
	int x;

	src+= stride*3;
	for(x=0; x<BLOCK_SIZE; x++)
	{
		int a= src[l3] - src[l4];
		int b= src[l4] - src[l5];
		int c= src[l5] - src[l6];

		int d= ABS(b) - ((ABS(a) + ABS(c))>>1);
		d= MAX(d, 0);

		if(d < co->QP*2)
		{
			int v = d * SIGN(-b);

			src[l2] +=v>>3;
			src[l3] +=v>>2;
			src[l4] +=(3*v)>>3;
			src[l5] -=(3*v)>>3;
			src[l6] -=v>>2;
			src[l7] -=v>>3;

		}
		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];
	const int l1= stride;
	const int l2= stride + l1;
	const int l3= stride + l2;
	const int l4= (int)tmp - (int)src - stride*3;
	const int l5= (int)tmp - (int)src - stride*3 + 8;
	const int l6= stride*3 + l3;
	const int l7= stride + l6;
	const int l8= stride + l7;

	memcpy(tmp, src+stride*7, 8);
	memcpy(tmp+8, src+stride*8, 8);
*/
	src+= stride*4;
	asm volatile(

#if 0 //sligtly more accurate and slightly slower
		"pxor %%mm7, %%mm7				\n\t" // 0
		"lea (%0, %1), %%"REG_a"			\n\t"
		"lea (%%"REG_a", %1, 4), %%"REG_c"		\n\t"
//	0	1	2	3	4	5	6	7
//	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ecx+%1	ecx+2%1
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1


		"movq (%0, %1, 2), %%mm0			\n\t" // l2
		"movq (%0), %%mm1				\n\t" // l0
		"movq %%mm0, %%mm2				\n\t" // l2
		PAVGB(%%mm7, %%mm0)				      // ~l2/2
		PAVGB(%%mm1, %%mm0)				      // ~(l2 + 2l0)/4
		PAVGB(%%mm2, %%mm0)				      // ~(5l2 + 2l0)/8

		"movq (%%"REG_a"), %%mm1			\n\t" // l1
		"movq (%%"REG_a", %1, 2), %%mm3			\n\t" // l3
		"movq %%mm1, %%mm4				\n\t" // l1
		PAVGB(%%mm7, %%mm1)				      // ~l1/2
		PAVGB(%%mm3, %%mm1)				      // ~(l1 + 2l3)/4
		PAVGB(%%mm4, %%mm1)				      // ~(5l1 + 2l3)/8

		"movq %%mm0, %%mm4				\n\t" // ~(5l2 + 2l0)/8
		"psubusb %%mm1, %%mm0				\n\t"
		"psubusb %%mm4, %%mm1				\n\t"
		"por %%mm0, %%mm1				\n\t" // ~|2l0 - 5l1 + 5l2 - 2l3|/8
// mm1= |lenergy|, mm2= l2, mm3= l3, mm7=0

		"movq (%0, %1, 4), %%mm0			\n\t" // l4
		"movq %%mm0, %%mm4				\n\t" // l4
		PAVGB(%%mm7, %%mm0)				      // ~l4/2
		PAVGB(%%mm2, %%mm0)				      // ~(l4 + 2l2)/4
		PAVGB(%%mm4, %%mm0)				      // ~(5l4 + 2l2)/8

		"movq (%%"REG_c"), %%mm2			\n\t" // l5
		"movq %%mm3, %%mm5				\n\t" // l3
		PAVGB(%%mm7, %%mm3)				      // ~l3/2
		PAVGB(%%mm2, %%mm3)				      // ~(l3 + 2l5)/4
		PAVGB(%%mm5, %%mm3)				      // ~(5l3 + 2l5)/8

		"movq %%mm0, %%mm6				\n\t" // ~(5l4 + 2l2)/8
		"psubusb %%mm3, %%mm0				\n\t"
		"psubusb %%mm6, %%mm3				\n\t"
		"por %%mm0, %%mm3				\n\t" // ~|2l2 - 5l3 + 5l4 - 2l5|/8
		"pcmpeqb %%mm7, %%mm0				\n\t" // SIGN(2l2 - 5l3 + 5l4 - 2l5)
// mm0= SIGN(menergy), mm1= |lenergy|, mm2= l5, mm3= |menergy|, mm4=l4, mm5= l3, mm7=0

		"movq (%%"REG_c", %1), %%mm6			\n\t" // l6
		"movq %%mm6, %%mm5				\n\t" // l6
		PAVGB(%%mm7, %%mm6)				      // ~l6/2
		PAVGB(%%mm4, %%mm6)				      // ~(l6 + 2l4)/4
		PAVGB(%%mm5, %%mm6)				      // ~(5l6 + 2l4)/8

		"movq (%%"REG_c", %1, 2), %%mm5			\n\t" // l7
		"movq %%mm2, %%mm4				\n\t" // l5
		PAVGB(%%mm7, %%mm2)				      // ~l5/2
		PAVGB(%%mm5, %%mm2)				      // ~(l5 + 2l7)/4
		PAVGB(%%mm4, %%mm2)				      // ~(5l5 + 2l7)/8

		"movq %%mm6, %%mm4				\n\t" // ~(5l6 + 2l4)/8
		"psubusb %%mm2, %%mm6				\n\t"
		"psubusb %%mm4, %%mm2				\n\t"
		"por %%mm6, %%mm2				\n\t" // ~|2l4 - 5l5 + 5l6 - 2l7|/8
// mm0= SIGN(menergy), mm1= |lenergy|/8, mm2= |renergy|/8, mm3= |menergy|/8, mm7=0


		PMINUB(%%mm2, %%mm1, %%mm4)			      // MIN(|lenergy|,|renergy|)/8
		"movq %2, %%mm4					\n\t" // QP //FIXME QP+1 ?
		"paddusb "MANGLE(b01)", %%mm4			\n\t"
		"pcmpgtb %%mm3, %%mm4				\n\t" // |menergy|/8 < QP
		"psubusb %%mm1, %%mm3				\n\t" // d=|menergy|/8-MIN(|lenergy|,|renergy|)/8
		"pand %%mm4, %%mm3				\n\t"

		"movq %%mm3, %%mm1				\n\t"
//		"psubusb "MANGLE(b01)", %%mm3			\n\t"
		PAVGB(%%mm7, %%mm3)
		PAVGB(%%mm7, %%mm3)
		"paddusb %%mm1, %%mm3				\n\t"
//		"paddusb "MANGLE(b01)", %%mm3			\n\t"

		"movq (%%"REG_a", %1, 2), %%mm6			\n\t" //l3
		"movq (%0, %1, 4), %%mm5			\n\t" //l4
		"movq (%0, %1, 4), %%mm4			\n\t" //l4
		"psubusb %%mm6, %%mm5				\n\t"
		"psubusb %%mm4, %%mm6				\n\t"
		"por %%mm6, %%mm5				\n\t" // |l3-l4|
		"pcmpeqb %%mm7, %%mm6				\n\t" // SIGN(l3-l4)
		"pxor %%mm6, %%mm0				\n\t"
		"pand %%mm0, %%mm3				\n\t"
		PMINUB(%%mm5, %%mm3, %%mm0)

		"psubusb "MANGLE(b01)", %%mm3			\n\t"
		PAVGB(%%mm7, %%mm3)

		"movq (%%"REG_a", %1, 2), %%mm0			\n\t"
		"movq (%0, %1, 4), %%mm2			\n\t"
		"pxor %%mm6, %%mm0				\n\t"
		"pxor %%mm6, %%mm2				\n\t"
		"psubb %%mm3, %%mm0				\n\t"
		"paddb %%mm3, %%mm2				\n\t"
		"pxor %%mm6, %%mm0				\n\t"
		"pxor %%mm6, %%mm2				\n\t"
		"movq %%mm0, (%%"REG_a", %1, 2)			\n\t"
		"movq %%mm2, (%0, %1, 4)			\n\t"
#endif

		"lea (%0, %1), %%"REG_a"			\n\t"
		"pcmpeqb %%mm6, %%mm6				\n\t" // -1
//	0	1	2	3	4	5	6	7
//	%0	%0+%1	%0+2%1	eax+2%1	%0+4%1	eax+4%1	ecx+%1	ecx+2%1
//	%0	eax	eax+%1	eax+2%1	%0+4%1	ecx	ecx+%1	ecx+2%1


		"movq (%%"REG_a", %1, 2), %%mm1			\n\t" // l3
		"movq (%0, %1, 4), %%mm0			\n\t" // l4
		"pxor %%mm6, %%mm1				\n\t" // -l3-1
		PAVGB(%%mm1, %%mm0)				      // -q+128 = (l4-l3+256)/2
// mm1=-l3-1, mm0=128-q

		"movq (%%"REG_a", %1, 4), %%mm2			\n\t" // l5
		"movq (%%"REG_a", %1), %%mm3			\n\t" // l2
		"pxor %%mm6, %%mm2				\n\t" // -l5-1
		"movq %%mm2, %%mm5				\n\t" // -l5-1
		"movq "MANGLE(b80)", %%mm4			\n\t" // 128
		"lea (%%"REG_a", %1, 4), %%"REG_c"		\n\t"
		PAVGB(%%mm3, %%mm2)				      // (l2-l5+256)/2
		PAVGB(%%mm0, %%mm4)				      // ~(l4-l3)/4 + 128
		PAVGB(%%mm2, %%mm4)				      // ~(l2-l5)/4 +(l4-l3)/8 + 128
		PAVGB(%%mm0, %%mm4)				      // ~(l2-l5)/8 +5(l4-l3)/16 + 128
// mm1=-l3-1, mm0=128-q, mm3=l2, mm4=menergy/16 + 128, mm5= -l5-1

		"movq (%%"REG_a"), %%mm2			\n\t" // l1
		"pxor %%mm6, %%mm2				\n\t" // -l1-1
		PAVGB(%%mm3, %%mm2)				      // (l2-l1+256)/2
		PAVGB((%0), %%mm1)				      // (l0-l3+256)/2
		"movq "MANGLE(b80)", %%mm3			\n\t" // 128
		PAVGB(%%mm2, %%mm3)				      // ~(l2-l1)/4 + 128
		PAVGB(%%mm1, %%mm3)				      // ~(l0-l3)/4 +(l2-l1)/8 + 128
		PAVGB(%%mm2, %%mm3)				      // ~(l0-l3)/8 +5(l2-l1)/16 + 128
// mm0=128-q, mm3=lenergy/16 + 128, mm4= menergy/16 + 128, mm5= -l5-1

		PAVGB((%%REGc, %1), %%mm5)			      // (l6-l5+256)/2
		"movq (%%"REG_c", %1, 2), %%mm1			\n\t" // l7
		"pxor %%mm6, %%mm1				\n\t" // -l7-1
		PAVGB((%0, %1, 4), %%mm1)			      // (l4-l7+256)/2
		"movq "MANGLE(b80)", %%mm2			\n\t" // 128
		PAVGB(%%mm5, %%mm2)				      // ~(l6-l5)/4 + 128
		PAVGB(%%mm1, %%mm2)				      // ~(l4-l7)/4 +(l6-l5)/8 + 128
		PAVGB(%%mm5, %%mm2)				      // ~(l4-l7)/8 +5(l6-l5)/16 + 128
// mm0=128-q, mm2=renergy/16 + 128, mm3=lenergy/16 + 128, mm4= menergy/16 + 128

		"movq "MANGLE(b00)", %%mm1			\n\t" // 0
		"movq "MANGLE(b00)", %%mm5			\n\t" // 0
		"psubb %%mm2, %%mm1				\n\t" // 128 - renergy/16
		"psubb %%mm3, %%mm5				\n\t" // 128 - lenergy/16
		PMAXUB(%%mm1, %%mm2)				      // 128 + |renergy/16|
 		PMAXUB(%%mm5, %%mm3)				      // 128 + |lenergy/16|
		PMINUB(%%mm2, %%mm3, %%mm1)			      // 128 + MIN(|lenergy|,|renergy|)/16

// mm0=128-q, mm3=128 + MIN(|lenergy|,|renergy|)/16, mm4= menergy/16 + 128

		"movq "MANGLE(b00)", %%mm7			\n\t" // 0
		"movq %2, %%mm2					\n\t" // QP
		PAVGB(%%mm6, %%mm2)				      // 128 + QP/2
		"psubb %%mm6, %%mm2				\n\t"

		"movq %%mm4, %%mm1				\n\t"
		"pcmpgtb %%mm7, %%mm1				\n\t" // SIGN(menergy)
		"pxor %%mm1, %%mm4				\n\t"
		"psubb %%mm1, %%mm4				\n\t" // 128 + |menergy|/16
		"pcmpgtb %%mm4, %%mm2				\n\t" // |menergy|/16 < QP/2
		"psubusb %%mm3, %%mm4				\n\t" //d=|menergy|/16 - MIN(|lenergy|,|renergy|)/16
// mm0=128-q, mm1= SIGN(menergy), mm2= |menergy|/16 < QP/2, mm4= d/16