simd/jdcolext-altivec.c

/*
 * AltiVec optimizations for libjpeg-turbo
 *
 * Copyright (C) 2015, D. R. Commander.
 * All rights reserved.
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 */

/* This file is included by jdcolor-altivec.c */


void jsimd_ycc_rgb_convert_altivec (JDIMENSION out_width, JSAMPIMAGE input_buf,
                                    JDIMENSION input_row,
                                    JSAMPARRAY output_buf, int num_rows)
{
  JSAMPROW outptr, inptr0, inptr1, inptr2;
  int pitch = out_width * RGB_PIXELSIZE, offset, num_cols;
  unsigned char __attribute__((aligned(16))) tmpbuf[RGB_PIXELSIZE * 16];

  __vector unsigned char rgb0, rgb1, rgb2, rgbx0, rgbx1, rgbx2, rgbx3,
    y, cb, cr, edgel, edgeh, edges, out0, out1, out2, out3;
#if RGB_PIXELSIZE == 4
  __vector unsigned char rgb3, out4;
#endif
  __vector short rg0, rg1, rg2, rg3, bx0, bx1, bx2, bx3, yl, yh, cbl, cbh,
    crl, crh, rl, rh, gl, gh, bl, bh, g0w, g1w, g2w, g3w;
  __vector int g0, g1, g2, g3;

  /* Constants
   * NOTE: The >> 1 is to compensate for the fact that vec_madds() returns 17
   * high-order bits, not 16.
   */
  __vector short pw_f0402 = { __8X(F_0_402 >> 1) },
    pw_mf0228 = { __8X(-F_0_228 >> 1) },
    pw_mf0344_f0285 = { __4X2(-F_0_344, F_0_285) },
    pw_one = { __8X(1) }, pw_255 = { __8X(255) },
    pw_cj = { __8X(CENTERJSAMPLE) };
  __vector int pd_onehalf = { __4X(ONE_HALF) };
  __vector unsigned char pb_zero = { __16X(0) },
    shift_pack_index =
      { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29};

  while (--num_rows >= 0) {
    inptr0 = input_buf[0][input_row];
    inptr1 = input_buf[1][input_row];
    inptr2 = input_buf[2][input_row];
    input_row++;
    outptr = *output_buf++;

    for (num_cols = pitch; num_cols > 0;
         num_cols -= RGB_PIXELSIZE * 16, outptr += RGB_PIXELSIZE * 16,
         inptr0 += 16, inptr1 += 16, inptr2 += 16) {

      y = vec_ld(0, inptr0);
      /* NOTE: We have to use vec_merge*() here because vec_unpack*() doesn't
       * support unsigned vectors.
       */
      yl = (__vector signed short)vec_mergeh(pb_zero, y);
      yh = (__vector signed short)vec_mergel(pb_zero, y);

      cb = vec_ld(0, inptr1);
      cbl = (__vector signed short)vec_mergeh(pb_zero, cb);
      cbh = (__vector signed short)vec_mergel(pb_zero, cb);
      cbl = vec_sub(cbl, pw_cj);
      cbh = vec_sub(cbh, pw_cj);

      cr = vec_ld(0, inptr2);
      crl = (__vector signed short)vec_mergeh(pb_zero, cr);
      crh = (__vector signed short)vec_mergel(pb_zero, cr);
      crl = vec_sub(crl, pw_cj);
      crh = vec_sub(crh, pw_cj);

      /* (Original)
       * R = Y                + 1.40200 * Cr
       * G = Y - 0.34414 * Cb - 0.71414 * Cr
       * B = Y + 1.77200 * Cb
       *
       * (This implementation)
       * R = Y                + 0.40200 * Cr + Cr
       * G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
       * B = Y - 0.22800 * Cb + Cb + Cb
       */
      bl = vec_add(cbl, cbl);
      bh = vec_add(cbh, cbh);
      bl = vec_madds(bl, pw_mf0228, pw_one);
      bh = vec_madds(bh, pw_mf0228, pw_one);
      bl = vec_sra(bl, (__vector unsigned short)pw_one);
      bh = vec_sra(bh, (__vector unsigned short)pw_one);
      bl = vec_add(bl, cbl);
      bh = vec_add(bh, cbh);
      bl = vec_add(bl, cbl);
      bh = vec_add(bh, cbh);
      bl = vec_add(bl, yl);
      bh = vec_add(bh, yh);

      rl = vec_add(crl, crl);
      rh = vec_add(crh, crh);
      rl = vec_madds(rl, pw_f0402, pw_one);
      rh = vec_madds(rh, pw_f0402, pw_one);
      rl = vec_sra(rl, (__vector unsigned short)pw_one);
      rh = vec_sra(rh, (__vector unsigned short)pw_one);
      rl = vec_add(rl, crl);
      rh = vec_add(rh, crh);
      rl = vec_add(rl, yl);
      rh = vec_add(rh, yh);

      g0w = vec_mergeh(cbl, crl);
      g1w = vec_mergel(cbl, crl);
      g0 = vec_msums(g0w, pw_mf0344_f0285, pd_onehalf);
      g1 = vec_msums(g1w, pw_mf0344_f0285, pd_onehalf);
      g2w = vec_mergeh(cbh, crh);
      g3w = vec_mergel(cbh, crh);
      g2 = vec_msums(g2w, pw_mf0344_f0285, pd_onehalf);
      g3 = vec_msums(g3w, pw_mf0344_f0285, pd_onehalf);
      /* Clever way to avoid 4 shifts + 2 packs.  This packs the high word from
       * each dword into a new 16-bit vector, which is the equivalent of
       * descaling the 32-bit results (right-shifting by 16 bits) and then
       * packing them.
       */
      gl = vec_perm((__vector short)g0, (__vector short)g1, shift_pack_index);
      gh = vec_perm((__vector short)g2, (__vector short)g3, shift_pack_index);
      gl = vec_sub(gl, crl);
      gh = vec_sub(gh, crh);
      gl = vec_add(gl, yl);
      gh = vec_add(gh, yh);

      rg0 = vec_mergeh(rl, gl);
      bx0 = vec_mergeh(bl, pw_255);
      rg1 = vec_mergel(rl, gl);
      bx1 = vec_mergel(bl, pw_255);
      rg2 = vec_mergeh(rh, gh);
      bx2 = vec_mergeh(bh, pw_255);
      rg3 = vec_mergel(rh, gh);
      bx3 = vec_mergel(bh, pw_255);

      rgbx0 = vec_packsu(rg0, bx0);
      rgbx1 = vec_packsu(rg1, bx1);
      rgbx2 = vec_packsu(rg2, bx2);
      rgbx3 = vec_packsu(rg3, bx3);

#if RGB_PIXELSIZE == 3
      /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
       * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
       * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
       * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
       *
       * rgb0 = R0 G0 B0 R1 G1 B1 R2 G2 B2 R3 G3 B3 R4 G4 B4 R5
       * rgb1 = G5 B5 R6 G6 B6 R7 G7 B7 R8 G8 B8 R9 G9 B9 Ra Ga
       * rgb2 = Ba Rb Gb Bb Rc Gc Bc Rd Gd Bd Re Ge Be Rf Gf Bf
       */
      rgb0 = vec_perm(rgbx0, rgbx1, (__vector unsigned char)RGB_INDEX0);
      rgb1 = vec_perm(rgbx1, rgbx2, (__vector unsigned char)RGB_INDEX1);
      rgb2 = vec_perm(rgbx2, rgbx3, (__vector unsigned char)RGB_INDEX2);
#else
      /* rgbx0 = R0 G0 R1 G1 R2 G2 R3 G3 B0 X0 B1 X1 B2 X2 B3 X3
       * rgbx1 = R4 G4 R5 G5 R6 G6 R7 G7 B4 X4 B5 X5 B6 X6 B7 X7
       * rgbx2 = R8 G8 R9 G9 Ra Ga Rb Gb B8 X8 B9 X9 Ba Xa Bb Xb
       * rgbx3 = Rc Gc Rd Gd Re Ge Rf Gf Bc Xc Bd Xd Be Xe Bf Xf
       *
       * rgb0 = R0 G0 B0 X0 R1 G1 B1 X1 R2 G2 B2 X2 R3 G3 B3 X3
       * rgb1 = R4 G4 B4 X4 R5 G5 B5 X5 R6 G6 B6 X6 R7 G7 B7 X7
       * rgb2 = R8 G8 B8 X8 R9 G9 B9 X9 Ra Ga Ba Xa Rb Gb Bb Xb
       * rgb3 = Rc Gc Bc Xc Rd Gd Bd Xd Re Ge Be Xe Rf Gf Bf Xf
       */
      rgb0 = vec_perm(rgbx0, rgbx0, (__vector unsigned char)RGB_INDEX);
      rgb1 = vec_perm(rgbx1, rgbx1, (__vector unsigned char)RGB_INDEX);
      rgb2 = vec_perm(rgbx2, rgbx2, (__vector unsigned char)RGB_INDEX);
      rgb3 = vec_perm(rgbx3, rgbx3, (__vector unsigned char)RGB_INDEX);
#endif

      offset = (size_t)outptr & 15;
      if (offset) {
        __vector unsigned char unaligned_shift_index;
        int bytes = num_cols + offset;

        if (bytes < (RGB_PIXELSIZE + 1) * 16 && (bytes & 15)) {
          /* Slow path to prevent buffer overwrite.  Since there is no way to
           * write a partial AltiVec register, overwrite would occur on the
           * last chunk of the last image row if the right edge is not on a
           * 16-byte boundary.  It could also occur on other rows if the bytes
           * per row is low enough.  Since we can't determine whether we're on
           * the last image row, we have to assume every row is the last.
           */
          vec_st(rgb0, 0, tmpbuf);
          vec_st(rgb1, 16, tmpbuf);
          vec_st(rgb2, 32, tmpbuf);
#if RGB_PIXELSIZE == 4
          vec_st(rgb3, 48, tmpbuf);
#endif
          memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
        } else {
          /* Fast path */
          unaligned_shift_index = vec_lvsl(0, outptr);
          edgel = vec_ld(0, outptr);
          edgeh = vec_ld(min(num_cols, RGB_PIXELSIZE * 16) + offset, outptr);
          edges = vec_perm(edgeh, edgel, unaligned_shift_index);
          unaligned_shift_index = vec_lvsr(0, outptr);
          out0 = vec_perm(edges, rgb0, unaligned_shift_index);
          out1 = vec_perm(rgb0, rgb1, unaligned_shift_index);
          out2 = vec_perm(rgb1, rgb2, unaligned_shift_index);
#if RGB_PIXELSIZE == 4
          out3 = vec_perm(rgb2, rgb3, unaligned_shift_index);
          out4 = vec_perm(rgb3, edges, unaligned_shift_index);
#else
          out3 = vec_perm(rgb2, edges, unaligned_shift_index);
#endif
          vec_st(out0, 0, outptr);
          if (bytes > 16)
            vec_st(out1, 16, outptr);
          if (bytes > 32)
            vec_st(out2, 32, outptr);
          if (bytes > 48)
            vec_st(out3, 48, outptr);
#if RGB_PIXELSIZE == 4
          if (bytes > 64)
            vec_st(out4, 64, outptr);
#endif
        }
      } else {
        if (num_cols < RGB_PIXELSIZE * 16 && (num_cols & 15)) {
          /* Slow path */
          vec_st(rgb0, 0, tmpbuf);
          vec_st(rgb1, 16, tmpbuf);
          vec_st(rgb2, 32, tmpbuf);
#if RGB_PIXELSIZE == 4
          vec_st(rgb3, 48, tmpbuf);
#endif
          memcpy(outptr, tmpbuf, min(num_cols, RGB_PIXELSIZE * 16));
        } else {
          /* Fast path */
          vec_st(rgb0, 0, outptr);
          if (num_cols > 16)
            vec_st(rgb1, 16, outptr);
          if (num_cols > 32)
            vec_st(rgb2, 32, outptr);
#if RGB_PIXELSIZE == 4
          if (num_cols > 48)
            vec_st(rgb3, 48, outptr);
#endif
        }
      }
    }
  }
}