void jsimd_h2v2_fancy_upsample_mmi()

in simd/mips64/jdsample-mmi.c [107:239]


void jsimd_h2v2_fancy_upsample_mmi(int max_v_samp_factor,
                                   JDIMENSION downsampled_width,
                                   JSAMPARRAY input_data,
                                   JSAMPARRAY *output_data_ptr)
{
  JSAMPARRAY output_data = *output_data_ptr;
  JSAMPROW inptr_1, inptr0, inptr1, outptr0, outptr1;
  int inrow, outrow, incol, tmp, tmp1;
  __m64 this_1l, this_1h, this_1, thiscolsum_1l, thiscolsum_1h;
  __m64 this0l, this0h, this0;
  __m64 this1l, this1h, this1, thiscolsum1l, thiscolsum1h;
  __m64 next_1l, next_1h, next_1, nextcolsum_1l, nextcolsum_1h;
  __m64 next0l, next0h, next0;
  __m64 next1l, next1h, next1, nextcolsum1l, nextcolsum1h;
  __m64 mask0 = 0.0, masklast, samp0123, samp4567, wk[4], zero = 0.0;

  mask0 = _mm_cmpeq_pi8(mask0, mask0);
  masklast = _mm_slli_si64(mask0, (SIZEOF_MMWORD - 2) * BYTE_BIT);
  mask0 = _mm_srli_si64(mask0, (SIZEOF_MMWORD - 2) * BYTE_BIT);

  for (inrow = 0, outrow = 0; outrow < max_v_samp_factor; inrow++) {

    inptr_1 = input_data[inrow - 1];
    inptr0 = input_data[inrow];
    inptr1 = input_data[inrow + 1];
    outptr0 = output_data[outrow++];
    outptr1 = output_data[outrow++];

    if (downsampled_width & 7) {
      tmp = (downsampled_width - 1) * sizeof(JSAMPLE);
      tmp1 = downsampled_width * sizeof(JSAMPLE);
      asm(PTR_ADDU  "$8, %3, %6\r\n"
          "lb       $9, ($8)\r\n"
          PTR_ADDU  "$8, %3, %7\r\n"
          "sb       $9, ($8)\r\n"
          PTR_ADDU  "$8, %4, %6\r\n"
          "lb       $9, ($8)\r\n"
          PTR_ADDU  "$8, %4, %7\r\n"
          "sb       $9, ($8)\r\n"
          PTR_ADDU  "$8, %5, %6\r\n"
          "lb       $9, ($8)\r\n"
          PTR_ADDU  "$8, %5, %7\r\n"
          "sb       $9, ($8)\r\n"
          : "=m" (*inptr_1), "=m" (*inptr0), "=m" (*inptr1)
          : "r" (inptr_1), "r" (inptr0), "r" (inptr1), "r" (tmp), "r" (tmp1)
          : "$8", "$9"
         );
    }

    /* process the first column block */
    this0 = _mm_load_si64((__m64 *)inptr0);    /* row[ 0][0] */
    this_1 = _mm_load_si64((__m64 *)inptr_1);  /* row[-1][0] */
    this1 = _mm_load_si64((__m64 *)inptr1);    /* row[ 1][0] */

    this0l = _mm_unpacklo_pi8(this0, zero);    /* row[ 0][0]( 0 1 2 3) */
    this0h = _mm_unpackhi_pi8(this0, zero);    /* row[ 0][0]( 4 5 6 7) */
    this_1l = _mm_unpacklo_pi8(this_1, zero);  /* row[-1][0]( 0 1 2 3) */
    this_1h = _mm_unpackhi_pi8(this_1, zero);  /* row[-1][0]( 4 5 6 7) */
    this1l = _mm_unpacklo_pi8(this1, zero);    /* row[+1][0]( 0 1 2 3) */
    this1h = _mm_unpackhi_pi8(this1, zero);    /* row[+1][0]( 4 5 6 7) */

    this0l = _mm_mullo_pi16(this0l, PW_THREE);
    this0h = _mm_mullo_pi16(this0h, PW_THREE);

    thiscolsum_1l = _mm_add_pi16(this_1l, this0l);  /* ( 0 1 2 3) */
    thiscolsum_1h = _mm_add_pi16(this_1h, this0h);  /* ( 4 5 6 7) */
    thiscolsum1l = _mm_add_pi16(this0l, this1l);    /* ( 0 1 2 3) */
    thiscolsum1h = _mm_add_pi16(this0h, this1h);    /* ( 4 5 6 7) */

    /* temporarily save the intermediate data */
    _mm_store_si64((__m64 *)outptr0, thiscolsum_1l);
    _mm_store_si64((__m64 *)outptr0 + 1, thiscolsum_1h);
    _mm_store_si64((__m64 *)outptr1, thiscolsum1l);
    _mm_store_si64((__m64 *)outptr1 + 1, thiscolsum1h);

    wk[0] = _mm_and_si64(thiscolsum_1l, mask0);  /* ( 0 - - -) */
    wk[1] = _mm_and_si64(thiscolsum1l, mask0);   /* ( 0 - - -) */

    for (incol = downsampled_width; incol > 0;
         incol -= 8, inptr_1 += 8, inptr0 += 8, inptr1 += 8,
         outptr0 += 16, outptr1 += 16) {

      if (incol > 8) {
        /* process the next column block */
        next0 = _mm_load_si64((__m64 *)inptr0 + 1);    /* row[ 0][1] */
        next_1 = _mm_load_si64((__m64 *)inptr_1 + 1);  /* row[-1][1] */
        next1 = _mm_load_si64((__m64 *)inptr1 + 1);    /* row[+1][1] */

        next0l = _mm_unpacklo_pi8(next0, zero);    /* row[ 0][1]( 0 1 2 3) */
        next0h = _mm_unpackhi_pi8(next0, zero);    /* row[ 0][1]( 4 5 6 7) */
        next_1l = _mm_unpacklo_pi8(next_1, zero);  /* row[-1][1]( 0 1 2 3) */
        next_1h = _mm_unpackhi_pi8(next_1, zero);  /* row[-1][1]( 4 5 6 7) */
        next1l = _mm_unpacklo_pi8(next1, zero);    /* row[+1][1]( 0 1 2 3) */
        next1h = _mm_unpackhi_pi8(next1, zero);    /* row[+1][1]( 4 5 6 7) */

        next0l = _mm_mullo_pi16(next0l, PW_THREE);
        next0h = _mm_mullo_pi16(next0h, PW_THREE);

        nextcolsum_1l = _mm_add_pi16(next_1l, next0l);  /* ( 0 1 2 3) */
        nextcolsum_1h = _mm_add_pi16(next_1h, next0h);  /* ( 4 5 6 7) */
        nextcolsum1l = _mm_add_pi16(next0l, next1l);    /* ( 0 1 2 3) */
        nextcolsum1h = _mm_add_pi16(next0h, next1h);    /* ( 4 5 6 7) */

        /* temporarily save the intermediate data */
        _mm_store_si64((__m64 *)outptr0 + 2, nextcolsum_1l);
        _mm_store_si64((__m64 *)outptr0 + 3, nextcolsum_1h);
        _mm_store_si64((__m64 *)outptr1 + 2, nextcolsum1l);
        _mm_store_si64((__m64 *)outptr1 + 3, nextcolsum1h);

        wk[2] = _mm_slli_si64(nextcolsum_1l, (SIZEOF_MMWORD - 2) * BYTE_BIT);  /* ( - - - 0) */
        wk[3] = _mm_slli_si64(nextcolsum1l, (SIZEOF_MMWORD - 2) * BYTE_BIT);   /* ( - - - 0) */
      } else {
        __m64 tmp;

        /* process the last column block */
        tmp = _mm_load_si64((__m64 *)outptr0 + 1);
        wk[2] = _mm_and_si64(masklast, tmp);        /* ( - - - 7) */
        tmp = _mm_load_si64((__m64 *)outptr1 + 1);
        wk[3] = _mm_and_si64(masklast, tmp);        /* ( - - - 7) */
      }

      /* process the upper row */
      samp0123 = _mm_load_si64((__m64 *)outptr0);      /* ( 0 1 2 3) */ \
      samp4567 = _mm_load_si64((__m64 *)outptr0 + 1);  /* ( 4 5 6 7) */ \
      PROCESS_ROW(0, 2, PW_EIGHT, PW_SEVEN, 4)

      /* process the lower row */
      samp0123 = _mm_load_si64((__m64 *)outptr1);      /* ( 0 1 2 3) */ \
      samp4567 = _mm_load_si64((__m64 *)outptr1 + 1);  /* ( 4 5 6 7) */ \
      PROCESS_ROW(1, 2, PW_EIGHT, PW_SEVEN, 4)
    }
  }
}