recon.cc

ac3的解码程序

		__asm__ (
			".align	8\n"
			"1:"
		   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   			"movq		(%2),	%%mm2\n"      /* 8 d */
				"pavgusb 	%%mm2, 	%%mm0\n"
				"addl		%3,		%1\n"
				"movq		%%mm0,	(%2)\n"
				"addl		%3,		%2\n"
			"loop		1b\n"
			:
			: "c" (h), "r" (s), "r" (d), "r" (lx2)
		);
	#else
	__asm__ (
  		"movq	     MASK_AND, 	%%mm5\n"
  		"movq	     ADD_1, 	%%mm6\n"
		"1:\t"
  			"movq       (%1),%%mm0\n"
  			"movq       (%2),%%mm1\n"
  			"psrlw		$1,%%mm0\n"
  			"psrlw	    $1,%%mm1\n"
  			"pand	    %%mm5,%%mm0\n"
  			"pand	    %%mm5,%%mm1\n"
  			"paddusb    %%mm1,%%mm0\n"
  			"paddusb    %%mm6,%%mm0\n"
  			"addl       %3,%1\n"
  			"movq       %%mm0,(%2)\n"
  			"decl       %0\n"
  			"leal		(%2, %3), %2\n"
  		"jnz        1b\n"
		:
		: "c" (h), "r" (s), "r" (d), "r" (lx2)
		);
	#endif
#else
  for (int j=0; j<h; j++, s+= lx2, d+= lx2){
    d[0] = (unsigned int)(d[0] + s[0] + 1)>>1;
    d[1] = (unsigned int)(d[1] + s[1] + 1)>>1;
    d[2] = (unsigned int)(d[2] + s[2] + 1)>>1;
    d[3] = (unsigned int)(d[3] + s[3] + 1)>>1;
    d[4] = (unsigned int)(d[4] + s[4] + 1)>>1;
    d[5] = (unsigned int)(d[5] + s[5] + 1)>>1;
    d[6] = (unsigned int)(d[6] + s[6] + 1)>>1;
    d[7] = (unsigned int)(d[7] + s[7] + 1)>>1;
  }
#endif
}

inline void recv(unsigned char *s, unsigned char *d, int lx,int lx2, int h){
#ifdef HAVE_MMX
	#ifdef HAVE_3Dnow
	__asm__(
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		(%4),	%%mm2\n"      /* 8 s +lx */
	   		"movq		8(%1),	%%mm1\n"      /* 8 s */
	   		"movq		8(%4),	%%mm3\n"      /* 8 s +lx **/

			"pavgusb %%mm2, 	%%mm0\n"
			"addl		%3,		%1\n"
			"pavgusb %%mm3, 	%%mm1\n"

			"movq		%%mm0,	(%2)\n"
			"addl		%3,		%4\n"
			"movq		%%mm1,	8(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
	#else
	__asm__ (
  		"movq	     MASK_AND, 	%%mm5\n"
  		"movq	     ADD_1, 	%%mm6\n"
		"1:\t"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		(%4),	%%mm1\n"      /* 8 s +lx */
	   		"movq		8(%1),	%%mm2\n"      /* 8 s */
	   		"movq		8(%4),	%%mm3\n"      /* 8 s +lx **/
  			"psrlw      $1,%%mm0\n"
  			"psrlw      $1,%%mm1\n"
  			"psrlw      $1,%%mm2\n"
  			"psrlw      $1,%%mm3\n"
  			"pand       %%mm5,%%mm0\n"
  			"pand       %%mm5,%%mm1\n"
  			"pand       %%mm5,%%mm2\n"
  			"pand       %%mm5,%%mm3\n"
  			"paddusb    %%mm1,%%mm0\n"
  			"paddusb    %%mm3,%%mm2\n"
  			"paddusb    %%mm6,%%mm0\n"
  			"paddusb    %%mm6,%%mm2\n"
  			"movq       %%mm0,(%2)\n"
  			"addl       %3,%1\n"
  			"movq	    %%mm2, 8(%2)\n"
  			"addl       %3,%4\n"
  			"decl       %0\n"
  			"leal		(%2, %3), %2\n"
  		"jnz        1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
	#endif

#else
  unsigned char *dp,*sp,*sp2;
  sp = s;
  sp2 = s+lx;
  dp = d;
  for (int j=0; j<h; j++){
    dp[0] = (unsigned int)(sp[0]+sp2[0]+1)>>1;
    dp[1] = (unsigned int)(sp[1]+sp2[1]+1)>>1;
    dp[2] = (unsigned int)(sp[2]+sp2[2]+1)>>1;
    dp[3] = (unsigned int)(sp[3]+sp2[3]+1)>>1;
    dp[4] = (unsigned int)(sp[4]+sp2[4]+1)>>1;
    dp[5] = (unsigned int)(sp[5]+sp2[5]+1)>>1;
    dp[6] = (unsigned int)(sp[6]+sp2[6]+1)>>1;
    dp[7] = (unsigned int)(sp[7]+sp2[7]+1)>>1;
    dp[8] = (unsigned int)(sp[8]+sp2[8]+1)>>1;
    dp[9] = (unsigned int)(sp[9]+sp2[9]+1)>>1;
    dp[10] = (unsigned int)(sp[10]+sp2[10]+1)>>1;
    dp[11] = (unsigned int)(sp[11]+sp2[11]+1)>>1;
    dp[12] = (unsigned int)(sp[12]+sp2[12]+1)>>1;
    dp[13] = (unsigned int)(sp[13]+sp2[13]+1)>>1;
    dp[14] = (unsigned int)(sp[14]+sp2[14]+1)>>1;
    dp[15] = (unsigned int)(sp[15]+sp2[15]+1)>>1;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx2;
  }
#endif
}

inline void recvc(unsigned char *s, unsigned char *d, int lx, int lx2, int h){
#ifdef HAVE_MMX
	#ifdef HAVE_3Dnow
	__asm__(
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		(%4),	%%mm2\n"      /* 8 s +lx */
			"addl		%3,		%1\n"
			"pavgusb %%mm2, 	%%mm0\n"
			"addl		%3,		%4\n"
			"movq		%%mm0,	(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
	#else
	__asm__ (
  		"movq	     MASK_AND, 	%%mm5\n"
  		"movq	     ADD_1, 	%%mm6\n"
		"1:\t"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		(%4),	%%mm1\n"      /* 8 s +lx */
  			"psrlw      $1,%%mm0\n"
  			"psrlw      $1,%%mm1\n"
  			"pand       %%mm5,%%mm0\n"
  			"pand       %%mm5,%%mm1\n"
  			"paddusb    %%mm1,%%mm0\n"
  			"addl       %3,%1\n"
  			"paddusb    %%mm6,%%mm0\n"
  			"addl       %3,%4\n"
  			"movq       %%mm0,(%2)\n"
  			"decl       %0\n"
  			"leal		(%2, %3), %2\n"
  		"jnz        1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
	#endif

#else
  unsigned char *dp,*sp,*sp2;

  sp = s;
  sp2 = s+lx;
  dp = d;
  for (int j=0; j<h; j++){
    dp[0] = (unsigned int)(sp[0]+sp2[0]+1)>>1;
    dp[1] = (unsigned int)(sp[1]+sp2[1]+1)>>1;
    dp[2] = (unsigned int)(sp[2]+sp2[2]+1)>>1;
    dp[3] = (unsigned int)(sp[3]+sp2[3]+1)>>1;
    dp[4] = (unsigned int)(sp[4]+sp2[4]+1)>>1;
    dp[5] = (unsigned int)(sp[5]+sp2[5]+1)>>1;
    dp[6] = (unsigned int)(sp[6]+sp2[6]+1)>>1;
    dp[7] = (unsigned int)(sp[7]+sp2[7]+1)>>1;
    sp+= lx2;
    sp2+= lx2;
    dp+= lx2;
  }
#endif
}

#ifdef HAVE_3Dnow
inline void recva(unsigned char *s, unsigned char *d, int lx, int lx2, int h)
{
	__asm__(
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		8(%1),	%%mm1\n"      /* 8 s */
	   		"movq		(%4),	%%mm2\n"      /* 8 s +lx */
	   		"movq		8(%4),	%%mm3\n"      /* 8 s +lx **/

			"pavgusb %%mm2, 	%%mm0\n"
			"addl		%3,		%1\n"
			"pavgusb %%mm3, 	%%mm1\n"

	   		"movq		(%2),	%%mm2\n"      /* 8 d */
	   		"movq		8(%2),	%%mm3\n"      /* 8 d */
			"pavgusb %%mm2, 	%%mm0\n"
			"addl		%3,		%4\n"
			"pavgusb %%mm3, 	%%mm1\n"

			"movq		%%mm0,	(%2)\n"
			"movq		%%mm1,	8(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
}

inline void recvac(unsigned char *s, unsigned char *d, int lx,int lx2, int h)
{
	__asm__(
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		(%4),	%%mm2\n"      /* 8 s +lx */
			"addl		%3,		%1\n"
			"pavgusb %%mm2, 	%%mm0\n"
	   		"movq		(%2),	%%mm3\n"      /* 8 d */
			"addl		%3,		%4\n"
			"pavgusb %%mm3, 	%%mm0\n"
			"movq		%%mm0,	(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
}

inline void rech(unsigned char *s, unsigned char *d, int lx2, int h)
{
	__asm__ (
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		8(%1),	%%mm1\n"      /* 8 s */
	   		"movq		1(%1),	%%mm2\n"      /* 8 s */
	   		"movq		9(%1),	%%mm3\n"      /* 8 s */

			"pavgusb 	%%mm2, 	%%mm0\n"
			"addl		%3,		%1\n"
			"pavgusb 	%%mm3, 	%%mm1\n"

			"movq		%%mm0,	(%2)\n"
			"movq		%%mm1,	8(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2)
	);
}

inline void rechc(unsigned char *s,unsigned char *d, int lx2, int h){
	__asm__ (
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		1(%1),	%%mm2\n"      /* 8 s +1 */
			"addl		%3,		%1\n"
			"pavgusb 	%%mm2, 	%%mm0\n"
			"movq		%%mm0,	(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2)
	);
}

inline void recha(unsigned char *s, unsigned char *d,int lx2, int h)
{
	__asm__ (
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		8(%1),	%%mm1\n"      /* 8 s */
	   		"movq		1(%1),	%%mm2\n"      /* 8 s */
	   		"movq		9(%1),	%%mm3\n"      /* 8 s */

			"pavgusb 	%%mm2, 	%%mm0\n"
			"addl		%3,		%1\n"
			"pavgusb 	%%mm3, 	%%mm1\n"

	   		"movq		(%2),	%%mm2\n"      /* 8 d */
	   		"movq		8(%2),	%%mm3\n"      /* 8 d */
			"pavgusb 	%%mm2, 	%%mm0\n"
			"pavgusb 	%%mm3, 	%%mm1\n"

			"movq		%%mm0,	(%2)\n"
			"movq		%%mm1,	8(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2)
	);
}

inline void rechac(unsigned char *s,unsigned char  *d, int lx2, int h)
{
	__asm__ (
		".align	8\n"
		"1:"
	   		"movq		(%1),	%%mm0\n"      /* 8 s */
	   		"movq		1(%1),	%%mm2\n"      /* 8 s */

			"addl		%3,		%1\n"
			"pavgusb 	%%mm2, 	%%mm0\n"

	   		"movq		(%2),	%%mm1\n"      /* 8 d */
			"pavgusb 	%%mm1, 	%%mm0\n"

			"movq		%%mm0,	(%2)\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      :
      : "c" (h), "r" (s), "r" (d), "r" (lx2)
	);
}
inline void rec4(unsigned char *s, unsigned char *d, int lx, int lx2, int h)
{
	__asm__ __volatile__(
	   	"movq		(%1),	%%mm0\n"  /* 8 s */
	   	"movq		8(%1),	%%mm1\n"  /* 8 s */
	   	"movq		1(%1),	%%mm2\n"  /* 8 s +1*/
	   	"movq		9(%1),	%%mm3\n"  /* 8 s +1*/
		".align 8\n"
		"1:"
	   		"movq		(%4),	%%mm4\n"  /* 8 s+lx */
			"pavgusb 	%%mm2, 	%%mm0\n"
	   		"movq		8(%4),	%%mm5\n"  /* 8 s+lx */
			"pavgusb 	%%mm3, 	%%mm1\n"

	   		"movq		1(%4),	%%mm6\n"  /* 8 s+lx +1*/
			"pavgusb 	%%mm4, 	%%mm0\n"
	   		"movq		9(%4),	%%mm7\n"  /* 8 s+lx +1*/
			"pavgusb 	%%mm5, 	%%mm1\n"

			"pavgusb 	%%mm6, 	%%mm0\n"
			"addl		%3,		%4\n"
			"pavgusb 	%%mm7, 	%%mm1\n"
			"movq		%%mm0,	(%2)\n"
			"movq		%%mm6,	%%mm2\n"
			"movq		%%mm7,	%%mm3\n"
			"movq		%%mm1,	8(%2)\n"
			"movq		%%mm4,	%%mm0\n"
			"movq		%%mm5,	%%mm1\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      	:
		: "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
}

inline void rec4c(unsigned char *s,unsigned char *d, int lx, int lx2, int h)
{
	__asm__ __volatile__(
	   	"movq		(%1),	%%mm0\n"  /* 8 s */
	   	"movq		1(%1),	%%mm2\n"  /* 8 s +1*/
		".align 8\n"
		"1:"
	   		"movq		(%4),	%%mm4\n"  /* 8 s+lx */
			"pavgusb 	%%mm2, 	%%mm0\n"

	   		"movq		1(%4),	%%mm6\n"  /* 8 s+lx +1*/
			"pavgusb 	%%mm4, 	%%mm0\n"

			"addl		%3,		%4\n"
			"pavgusb 	%%mm6, 	%%mm0\n"
			"movq		%%mm0,	(%2)\n"
			"movq		%%mm6,	%%mm2\n"
			"movq		%%mm4,	%%mm0\n"
			"addl		%3,		%2\n"
		"loop		1b\n"
      	:
		: "c" (h), "r" (s), "r" (d), "r" (lx2), "r" (s +lx)
	);
}

inline void rec4a(unsigned char *s,unsigned char *d, int lx, int lx2, int h)
{
	__asm__ __volatile__(
	   	"movq		(%1),	%%mm0\n"  /* 8 s */
	   	"movq		8(%1),	%%mm1\n"  /* 8 s */
	   	"movq		1(%1),	%%mm2\n"  /* 8 s +1*/
	   	"movq		9(%1),	%%mm3\n"  /* 8 s +1*/
		".align 8\n"
		"1:"
	   		"movq		(%4),	%%mm4\n"  /* 8 s+lx */
			"pavgusb 	%%mm2, 	%%mm0\n"
	   		"movq		8(%4),	%%mm5\n"  /* 8 s+lx */
			"pavgusb 	%%mm3, 	%%mm1\n"

	   		"movq		1(%4),	%%mm6\n"  /* 8 s+lx +1*/
			"pavgusb 	%%mm4, 	%%mm0\n"
	   		"movq		9(%4),	%%mm7\n"  /* 8 s+lx +1*/
			"pavgusb 	%%mm5, 	%%mm1\n"
			"movq		(%2),	%%mm2\n"
			"pavgusb 	%%mm6, 	%%mm0\n"
			"movq		8(%2),	%%mm3\n"

			"pavgusb 	%%mm2, 	%%mm0\n"