gfx/2d/SwizzleAVX2.cpp

   1 /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
   2 /* vim: set ts=8 sts=2 et sw=2 tw=80: */
   3 /* This Source Code Form is subject to the terms of the Mozilla Public
   4  * License, v. 2.0. If a copy of the MPL was not distributed with this
   5  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
   6
   7 #include "Swizzle.h"
   8
   9 #include <immintrin.h>
  10 #include <tmmintrin.h>
  11
  12 namespace mozilla::gfx {
  13
  14 template <bool aSwapRB>
  15 void UnpackRowRGB24_SSSE3(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength);
  16
  17 template <bool aSwapRB>
  18 void UnpackRowRGB24_AVX2(const uint8_t* aSrc, uint8_t* aDst, int32_t aLength) {
  19   // Because this implementation will read an additional 8 bytes of data that
  20   // is ignored and masked over, we cannot use the accelerated version for the
  21   // last 1-10 pixels (3-30 bytes remaining) to guarantee we don't access memory
  22   // outside the buffer (we read in 32 byte chunks).
  23   if (aLength < 11) {
  24     UnpackRowRGB24_SSSE3<aSwapRB>(aSrc, aDst, aLength);
  25     return;
  26   }
  27
  28   // Because we are expanding, we can only process the data back to front in
  29   // case we are performing this in place.
  30   int32_t alignedRow = (aLength - 4) & ~7;
  31   int32_t remainder = aLength - alignedRow;
  32
  33   const uint8_t* src = aSrc + alignedRow * 3;
  34   uint8_t* dst = aDst + alignedRow * 4;
  35
  36   // Handle any 3-10 remaining pixels.
  37   UnpackRowRGB24_SSSE3<aSwapRB>(src, dst, remainder);
  38
  39   // Used to shuffle the two final 32-bit words which we ignore into the last
  40   // 32-bit word of each 128-bit lane, such that
  41   //   RGBR GBRG BRGB RGBR GBRG BRGB RGBR GBRG
  42   //   BRGB RGBR GBRG BRGB ZZZZ ZZZZ ZZZZ ZZZZ
  43   // becomes
  44   //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
  45   //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
  46   const __m256i discardMask = _mm256_set_epi32(7, 5, 4, 3, 6, 2, 1, 0);
  47
  48   // Used to shuffle 8-bit words within a 128-bit lane, such that we transform
  49   //   RGBR GBRG BRGB RGBR GBRG BRGB ZZZZ ZZZZ
  50   // into
  51   //   RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ RGBZ
  52   // or
  53   //   BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ BGRZ
  54   const __m256i colorMask =
  55       aSwapRB ? _mm256_set_epi8(15, 9, 10, 11, 14, 6, 7, 8, 13, 3, 4, 5, 12, 0,
  56                                 1, 2, 15, 9, 10, 11, 14, 6, 7, 8, 13, 3, 4, 5,
  57                                 12, 0, 1, 2)
  58               : _mm256_set_epi8(15, 11, 10, 9, 14, 8, 7, 6, 13, 5, 4, 3, 12, 2,
  59                                 1, 0, 15, 11, 10, 9, 14, 8, 7, 6, 13, 5, 4, 3,
  60                                 12, 2, 1, 0);
  61
  62   // Used to transform RGBZ/BGRZ to RGBX/BGRX, or force the alpha opaque.
  63   const __m256i alphaMask = _mm256_set1_epi32(0xFF000000);
  64
  65   // Process all 8-pixel chunks as one vector.
  66   src -= 8 * 3;
  67   dst -= 8 * 4;
  68   while (src >= aSrc) {
  69     __m256i px = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(src));
  70     px = _mm256_permutevar8x32_epi32(px, discardMask);
  71     px = _mm256_shuffle_epi8(px, colorMask);
  72     px = _mm256_or_si256(px, alphaMask);
  73     _mm256_storeu_si256(reinterpret_cast<__m256i*>(dst), px);
  74     src -= 8 * 3;
  75     dst -= 8 * 4;
  76   }
  77 }
  78
  79 // Force instantiation of swizzle variants here.
  80 template void UnpackRowRGB24_AVX2<false>(const uint8_t*, uint8_t*, int32_t);
  81 template void UnpackRowRGB24_AVX2<true>(const uint8_t*, uint8_t*, int32_t);
  82
  83 }  // namespace mozilla::gfx