dsputil_mmx.c

mpeg4 video codec mpeg4 video codec

      "movl %4,%%ecx\n"
      "pxor %%mm0,%%mm0\n"	/* mm0 = 0 */
      "pxor %%mm7,%%mm7\n"	/* mm7 holds the sum */
      "1:\n"
      "movq (%0),%%mm1\n"	/* mm1 = pix1[0-7] */
      "movq (%1),%%mm2\n"	/* mm2 = pix2[0-7] */
      "movq 8(%0),%%mm3\n"	/* mm3 = pix1[8-15] */
      "movq 8(%1),%%mm4\n"	/* mm4 = pix2[8-15] */

      /* todo: mm1-mm2, mm3-mm4 */
      /* algo: substract mm1 from mm2 with saturation and vice versa */
      /*       OR the results to get absolute difference */
      "movq %%mm1,%%mm5\n"
      "movq %%mm3,%%mm6\n"
      "psubusb %%mm2,%%mm1\n"
      "psubusb %%mm4,%%mm3\n"
      "psubusb %%mm5,%%mm2\n"
      "psubusb %%mm6,%%mm4\n"

      "por %%mm1,%%mm2\n"
      "por %%mm3,%%mm4\n"

      /* now convert to 16-bit vectors so we can square them */
      "movq %%mm2,%%mm1\n"
      "movq %%mm4,%%mm3\n"

      "punpckhbw %%mm0,%%mm2\n"
      "punpckhbw %%mm0,%%mm4\n"
      "punpcklbw %%mm0,%%mm1\n"	/* mm1 now spread over (mm1,mm2) */
      "punpcklbw %%mm0,%%mm3\n"	/* mm4 now spread over (mm3,mm4) */

      "pmaddwd %%mm2,%%mm2\n"
      "pmaddwd %%mm4,%%mm4\n"
      "pmaddwd %%mm1,%%mm1\n"
      "pmaddwd %%mm3,%%mm3\n"

      "add %3,%0\n"
      "add %3,%1\n"

      "paddd %%mm2,%%mm1\n"
      "paddd %%mm4,%%mm3\n"
      "paddd %%mm1,%%mm7\n"
      "paddd %%mm3,%%mm7\n"

      "decl %%ecx\n"
      "jnz 1b\n"

      "movq %%mm7,%%mm1\n"
      "psrlq $32, %%mm7\n"	/* shift hi dword to lo */
      "paddd %%mm7,%%mm1\n"
      "movd %%mm1,%2\n"
      : "+r" (pix1), "+r" (pix2), "=r"(tmp) 
      : "r" ((long)line_size) , "m" (h)
      : "%ecx");
    return tmp;
}

static int sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    int tmp;
  asm volatile (
      "shr $1,%2\n"
      "pxor %%xmm0,%%xmm0\n"	/* mm0 = 0 */
      "pxor %%xmm7,%%xmm7\n"	/* mm7 holds the sum */
      "1:\n"
      "movdqu (%0),%%xmm1\n"	/* mm1 = pix1[0][0-15] */
      "movdqu (%1),%%xmm2\n"	/* mm2 = pix2[0][0-15] */
      "movdqu (%0,%4),%%xmm3\n"	/* mm3 = pix1[1][0-15] */
      "movdqu (%1,%4),%%xmm4\n"	/* mm4 = pix2[1][0-15] */

      /* todo: mm1-mm2, mm3-mm4 */
      /* algo: substract mm1 from mm2 with saturation and vice versa */
      /*       OR the results to get absolute difference */
      "movdqa %%xmm1,%%xmm5\n"
      "movdqa %%xmm3,%%xmm6\n"
      "psubusb %%xmm2,%%xmm1\n"
      "psubusb %%xmm4,%%xmm3\n"
      "psubusb %%xmm5,%%xmm2\n"
      "psubusb %%xmm6,%%xmm4\n"

      "por %%xmm1,%%xmm2\n"
      "por %%xmm3,%%xmm4\n"

      /* now convert to 16-bit vectors so we can square them */
      "movdqa %%xmm2,%%xmm1\n"
      "movdqa %%xmm4,%%xmm3\n"

      "punpckhbw %%xmm0,%%xmm2\n"
      "punpckhbw %%xmm0,%%xmm4\n"
      "punpcklbw %%xmm0,%%xmm1\n"	/* mm1 now spread over (mm1,mm2) */
      "punpcklbw %%xmm0,%%xmm3\n"	/* mm4 now spread over (mm3,mm4) */

      "pmaddwd %%xmm2,%%xmm2\n"
      "pmaddwd %%xmm4,%%xmm4\n"
      "pmaddwd %%xmm1,%%xmm1\n"
      "pmaddwd %%xmm3,%%xmm3\n"

      "lea (%0,%4,2), %0\n"	/* pix1 += 2*line_size */
      "lea (%1,%4,2), %1\n"	/* pix2 += 2*line_size */

      "paddd %%xmm2,%%xmm1\n"
      "paddd %%xmm4,%%xmm3\n"
      "paddd %%xmm1,%%xmm7\n"
      "paddd %%xmm3,%%xmm7\n"

      "decl %2\n"
      "jnz 1b\n"

      "movdqa %%xmm7,%%xmm1\n"
      "psrldq $8, %%xmm7\n"	/* shift hi qword to lo */
      "paddd %%xmm1,%%xmm7\n"
      "movdqa %%xmm7,%%xmm1\n"
      "psrldq $4, %%xmm7\n"	/* shift hi dword to lo */
      "paddd %%xmm1,%%xmm7\n"
      "movd %%xmm7,%3\n"
      : "+r" (pix1), "+r" (pix2), "+r"(h), "=r"(tmp) 
      : "r" ((long)line_size));
    return tmp;
}

static int hf_noise8_mmx(uint8_t * pix1, int line_size, int h) {
    int tmp;
  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm7,%%mm7\n"
      "pxor %%mm6,%%mm6\n"
      
      "movq (%0),%%mm0\n"
      "movq %%mm0, %%mm1\n"
      "psllq $8, %%mm0\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm0\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq %%mm4, %%mm1\n"
      "psllq $8, %%mm4\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm4\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "1:\n"
  
      "movq (%0),%%mm0\n"
      "movq %%mm0, %%mm1\n"
      "psllq $8, %%mm0\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm0\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      "psubw %%mm0, %%mm4\n"
      "psubw %%mm2, %%mm5\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm4, %%mm3\n\t"
      "pcmpgtw %%mm5, %%mm1\n\t"
      "pxor %%mm3, %%mm4\n"
      "pxor %%mm1, %%mm5\n"
      "psubw %%mm3, %%mm4\n" 
      "psubw %%mm1, %%mm5\n"
      "paddw %%mm4, %%mm5\n"
      "paddw %%mm5, %%mm6\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq %%mm4, %%mm1\n"
      "psllq $8, %%mm4\n"
      "psrlq $8, %%mm1\n"
      "psrlq $8, %%mm4\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "subl $2, %%ecx\n"
      " jnz 1b\n"

      "movq %%mm6, %%mm0\n"
      "punpcklwd %%mm7,%%mm0\n"
      "punpckhwd %%mm7,%%mm6\n"
      "paddd %%mm0, %%mm6\n"
      
      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddd %%mm6,%%mm0\n"
      "movd %%mm0,%1\n"
      : "+r" (pix1), "=r"(tmp) 
      : "r" ((long)line_size) , "g" (h-2)
      : "%ecx");
      return tmp;
}

static int hf_noise16_mmx(uint8_t * pix1, int line_size, int h) {
    int tmp;
    uint8_t * pix= pix1;
  asm volatile (
      "movl %3,%%ecx\n"
      "pxor %%mm7,%%mm7\n"
      "pxor %%mm6,%%mm6\n"
      
      "movq (%0),%%mm0\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "1:\n"
  
      "movq (%0),%%mm0\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm0, %%mm2\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm0\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm2\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm0\n"
      "psubw %%mm3, %%mm2\n"
      "psubw %%mm0, %%mm4\n"
      "psubw %%mm2, %%mm5\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm4, %%mm3\n\t"
      "pcmpgtw %%mm5, %%mm1\n\t"
      "pxor %%mm3, %%mm4\n"
      "pxor %%mm1, %%mm5\n"
      "psubw %%mm3, %%mm4\n"
      "psubw %%mm1, %%mm5\n"
      "paddw %%mm4, %%mm5\n"
      "paddw %%mm5, %%mm6\n"
      
      "add %2,%0\n"
      
      "movq (%0),%%mm4\n"
      "movq 1(%0),%%mm1\n"
      "movq %%mm4, %%mm5\n"
      "movq %%mm1, %%mm3\n"
      "punpcklbw %%mm7,%%mm4\n"
      "punpcklbw %%mm7,%%mm1\n"
      "punpckhbw %%mm7,%%mm5\n"
      "punpckhbw %%mm7,%%mm3\n"
      "psubw %%mm1, %%mm4\n"
      "psubw %%mm3, %%mm5\n"
      "psubw %%mm4, %%mm0\n"
      "psubw %%mm5, %%mm2\n"
      "pxor %%mm3, %%mm3\n"
      "pxor %%mm1, %%mm1\n"
      "pcmpgtw %%mm0, %%mm3\n\t"
      "pcmpgtw %%mm2, %%mm1\n\t"
      "pxor %%mm3, %%mm0\n"
      "pxor %%mm1, %%mm2\n"
      "psubw %%mm3, %%mm0\n" 
      "psubw %%mm1, %%mm2\n"
      "paddw %%mm0, %%mm2\n"
      "paddw %%mm2, %%mm6\n"

      "add %2,%0\n"
      "subl $2, %%ecx\n"
      " jnz 1b\n"

      "movq %%mm6, %%mm0\n"
      "punpcklwd %%mm7,%%mm0\n"
      "punpckhwd %%mm7,%%mm6\n"
      "paddd %%mm0, %%mm6\n"
      
      "movq %%mm6,%%mm0\n"
      "psrlq $32, %%mm6\n"
      "paddd %%mm6,%%mm0\n"
      "movd %%mm0,%1\n"
      : "+r" (pix1), "=r"(tmp) 
      : "r" ((long)line_size) , "g" (h-2)
      : "%ecx");
      return tmp + hf_noise8_mmx(pix+8, line_size, h);
}

static int nsse16_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    MpegEncContext *c = p;
    int score1, score2;

    if(c) score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h);
    else  score1 = sse16_mmx(c, pix1, pix2, line_size, h);
    score2= hf_noise16_mmx(pix1, line_size, h) - hf_noise16_mmx(pix2, line_size, h);

    if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
    else  return score1 + ABS(score2)*8;
}

static int nsse8_mmx(void *p, uint8_t * pix1, uint8_t * pix2, int line_size, int h) {
    MpegEncContext *c = p;
    int score1= sse8_mmx(c, pix1, pix2, line_size, h);
    int score2= hf_noise8_mmx(pix1, line_size, h) - hf_noise8_mmx(pix2, line_size, h);

    if(c) return score1 + ABS(score2)*c->avctx->nsse_weight;
    else  return score1 + ABS(score2)*8;
}

static int vsad_intra16_mmx(void *v, uint8_t * pix, uint8_t * dummy, int line_size, int h) {
    int tmp;
    
    assert( (((int)pix) & 7) == 0);
    assert((line_size &7) ==0);
    
#define SUM(in0, in1, out0, out1) \
      "movq (%0), %%mm2\n"\
      "movq 8(%0), %%mm3\n"\
      "add %2,%0\n"\
      "movq %%mm2, " #out0 "\n"\
      "movq %%mm3, " #out1 "\n"\
      "psubusb " #in0 ", %%mm2\n"\
      "psubusb " #in1 ", %%mm3\n"\
      "psubusb " #out0 ", " #in0 "\n"\
      "psubusb " #out1 ", " #in1 "\n"\
      "por %%mm2, " #in0 "\n"\
      "por %%mm3, " #in1 "\n"\
      "movq " #in0 ", %%mm2\n"\
      "movq " #in1 ", %%mm3\n"\
      "punpcklbw %%mm7, " #in0 "\n"\
      "punpcklbw %%mm7, " #in1 "\n"\
      "punpckhbw %%mm7, %%mm2\n"\
      "punpckhbw %%mm7, %%mm3\n"\
      "paddw " #in1 ", " #in0 "\n"\