Tile Compression

cc/resources/texture_compress/atc_dxt.cc

+// Copyright 2014 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+// This file is based on the public domain code "stb_dxt.h" originally written
+// by Fabian Giesen and Sean Barrett.
+//
+// The following changes were made:
+// - Added support for ATC format.
+// - Replaced the Principal Component Analysis based calculation to find the
+//   initial base colors with a much simpler bounding box implementation for low
+//   quality only.
+// - Removed dithering support.
+// - Some minor optimizations.
+// - Reformatted the code (mainly with clang-format).
+// - Swapped red and blue channels in the output to match Skia.
+
+#include "cc/resources/texture_compress/atc_dxt.h"
+#include <cmath>
+#include <cstdlib>
+#include "base/logging.h"
+
+namespace cc {
+namespace texture_compress {
+
+struct TYPE_ATC_GENERIC : public TYPE_ATC {
+  typedef TYPE_ATC BASE_TYPE;
+  static const int kRemap[8];
+  static const int kW1Table[4];
+  static const int kProds[4];
+};
+
+struct TYPE_DXT_GENERIC : public TYPE_DXT {
+  typedef TYPE_DXT BASE_TYPE;
+  static const int kRemap[8];
+  static const int kW1Table[4];
+  static const int kProds[4];
+};
+
+const int TYPE_ATC_GENERIC::kRemap[8] =
+    {0 << 30, 1 << 30, 0 << 30, 1 << 30, 2 << 30, 2 << 30, 3 << 30, 3 << 30};
+const int TYPE_ATC_GENERIC::kW1Table[4] = {3, 2, 1, 0};
+const int TYPE_ATC_GENERIC::kProds[4] = {0x090000,
+                                         0x040102,
+                                         0x010402,
+                                         0x000900};
+
+const int TYPE_DXT_GENERIC::kRemap[8] =
+    {0 << 30, 2 << 30, 0 << 30, 2 << 30, 3 << 30, 3 << 30, 1 << 30, 1 << 30};
+const int TYPE_DXT_GENERIC::kW1Table[4] = {3, 0, 2, 1};
+const int TYPE_DXT_GENERIC::kProds[4] = {0x090000,
+                                         0x000900,
+                                         0x040102,
+                                         0x010402};
+
+// Number of passes over the block that's done to refine the base colors.
+const int kNumRefinements = 2;
+
+uint8_t g_o_match55[256][2];
+uint8_t g_o_match66[256][2];
+uint8_t g_o_match56[256][2];
+
+namespace {
+
+inline int Mul8Bit(int a, int b) {
+  int t = a * b + 128;
+  return (t + (t >> 8)) >> 8;
+}
+
+// Linear interpolation at 1/3 point between a and b, using desired rounding
+// type.
+inline int Lerp13(int a, int b) {
+  // Without rounding bias.
+  // (2 * a + b) / 3;
+  return ((2 * a + b) * 0xaaab) >> 17;
+}
+
+inline void Lerp13RGB(uint8_t* out, const uint8_t* p1, const uint8_t* p2) {
+  out[0] = Lerp13(p1[0], p2[0]);
+  out[1] = Lerp13(p1[1], p2[1]);
+  out[2] = Lerp13(p1[2], p2[2]);
+}
+
+// Compute table to reproduce constant colors as accurately as possible.
+void PrepareOptTable(uint8_t* table,
+                     const uint8_t* expand_max,
+                     const uint8_t* expand_min,
+                     int size_max,
+                     int size_min) {
+  for (int i = 0; i < 256; ++i) {
+    int best_err = 256;
+    for (int mn = 0; mn < size_min; ++mn) {
+      for (int mx = 0; mx < size_max; ++mx) {
+        int mine = expand_min[mn];
+        int maxe = expand_max[mx];
+        int err = std::abs(Lerp13(maxe, mine) - i);
+
+        // DX10 spec says that interpolation must be within 3% of "correct"
+        // result, add this as error term. (normally we'd expect a random
+        // distribution of +-1.5% error, but nowhere in the spec does it say
+        // that the error has to be unbiased - better safe than sorry).
+        err += std::abs(maxe - mine) * 3 / 100;
+
+        if (err < best_err) {
+          table[i * 2 + 0] = mx;
+          table[i * 2 + 1] = mn;
+          best_err = err;
+        }
+      }
+    }
+  }
+}
+
+inline uint8_t Expand5(int i) {
+  return (i << 3) | (i >> 2);
+}
+
+inline uint8_t Expand6(int i) {
+  return (i << 2) | (i >> 4);
+}
+
+inline uint16_t As16Bit(int r, int g, int b) {
+  return (Mul8Bit(r, 31) << 11) + (Mul8Bit(g, 63) << 5) + Mul8Bit(b, 31);
+}
+
+inline int sclamp(float y, int p0, int p1) {
+  int x = static_cast<int>(y);
+  if (x < p0) {
+    return p0;
+  }
+  if (x > p1) {
+    return p1;
+  }
+  return x;
+}
+
+#if defined(OS_ANDROID)
+inline uint16_t RGB2BGR(uint16_t rgb) {
+  uint16_t r = rgb & 0xf800;
+  uint16_t g = rgb & 0x07e0;
+  uint16_t b = rgb & 0x001f;
+  return (r >> 11) | g | (b << 11);
+}
+#endif
+
+// Take two 16-bit base colors and generate 4 32-bit RGBX colors where:
+//  0 = c0
+//  1 = c1
+//  2 = (2 * c0 + c1) / 3
+//  3 = (2 * c1 + c0) / 3
+//
+// The base colors are expanded by reusing the top bits at the end. That makes
+// sure that white is still white after being quantized and converted back.
+//
+// params:
+//  color   (output) 4 RGBA pixels.
+//  c0      base color 0 (16 bit RGB)
+//  c1      base color 1 (16 bit RGB)
+inline void EvalColors(uint8_t* color, uint16_t c0, uint16_t c1) {
+  // Expand the two base colors to 32-bit
+  // From: [00000000][00000000][rrrrrggg][gggbbbbb]
+  // To:   [00000000][bbbbbxxx][ggggggxx][rrrrrxxx]
+  // Where x means repeat the upper bits for that color component.
+
+  // Take shortcut if either color is zero. Both will never be zero.
+  DCHECK(c0 | c1);
+  if (c0 && c1) {
+    // Combine the two base colors into one register to allow operating on both
+    // pixels at the same time.
+    // [rrrrrggg][gggbbbbb][RRRRRGGG][GGGBBBBB]
+    uint32_t c01 = c1 | (c0 << 16);
+
+    // Mask out the red components and shift it down one channel to avoid some
+    // shifts when combining the channels.
+    // [00000000][rrrrr000][00000000][RRRRR000]
+    uint32_t c01_r = (c01 & 0xf800f800) >> 8;
+    // Extend to be 8-bit by reusing top bits at the end.
+    // Note that we leave some extra garbage bits in the other channels, but
+    // that's ok since we mask that off when we combine the different
+    // components.
+    // [00000000][rrrrrrrr][xx000000][RRRRRRRR]
+    c01_r |= (c01_r >> 5);
+
+    // Mask out the green components.
+    // [00000ggg][ggg00000][00000GGG][GGG00000]
+    uint32_t c01_g = c01 & 0x07e007e0;
+    // Shift it into place and extend.
+    // [gggggggg][xxxx0000][GGGGGGGG][xxxx0000]
+    c01_g = ((c01_g << 5) | (c01_g >> 1));
+
+    // Mask out the blue components.
+    // [00000000][000bbbbb][00000000][000BBBBB]
+    uint32_t c01_b = c01 & 0x001f001f;
+    // Shift it into place and extend.
+    // [bbbbbbbb][xx000000][BBBBBBBB][xx000000]
+    c01_b = ((c01_b << 11) | (c01_b << 6));
+
+    // Combine the components into base color 0.
+    // Shift the components into place and mask of each channel.
+    // [00000000][bbbbbbbb][gggggggg][rrrrrrrr]
+    *reinterpret_cast<uint32_t*>(color + 0) = ((c01_r >> 16) & 0x000000ff) |
+                                              ((c01_g >> 16) & 0x0000ff00) |
+                                              ((c01_b >> 8) & 0x00ff0000);
+
+    // Combine the components into base color 1.
+    // [00000000][BBBBBBBB][GGGGGGGG][RRRRRRRR]
+    *reinterpret_cast<uint32_t*>(color + 4) = (c01_r & 0x000000ff) |
+                                              (c01_g & 0x0000ff00) |
+                                              ((c01_b << 8) & 0x00ff0000);
+
+    Lerp13RGB(color + 8, color, color + 4);
+    Lerp13RGB(color + 12, color + 4, color);
+  } else {
+    // Combine the two base colors into one register, one of them will be zero.
+    // [00000000][00000000][rrrrrggg][gggbbbbb]
+    uint32_t c = c0 | c1;
+
+    // Mask out the red components and shift it down one channel to avoid some
+    // shifts when combining the channels.
+    // [00000000][00000000][00000000][rrrrr000]
+    uint32_t c_r = (c & 0xf800) >> 8;
+    // Extend to be 8-bit by reusing top bits at the end.
+    // [00000000][00000000][00000000][rrrrrrrr]
+    c_r |= c_r >> 5;
+
+    // Mask out the green components.
+    // [00000000][00000000][00000ggg][ggg00000]
+    uint32_t c_g = c & 0x07e0;
+    // Shift it into place and extend. Then mask off garbage bits.
+    // [00000000][00000000][gggggggg][xxxx0000]
+    c_g = ((c_g << 5) | (c_g >> 1)) & 0x0000ff00;
+
+    // Mask out the blue components.
+    // [00000000][00000000][00000000][000bbbbb]
+    uint32_t c_b = c & 0x001f;
+    // Shift it into place and extend. Then mask off garbage bits.
+    // [00000000][bbbbbbbb][xx000000][00000000]
+    c_b = ((c_b << 19) | (c_b << 14)) & 0x00ff0000;
+
+    size_t zero_offset = !!c0 * 4;
+    size_t nonzero_offset = !!c1 * 4;
+
+    // Combine the components into non zero base color.
+    // [00000000][bbbbbbbb][gggggggg][rrrrrrrr]
+    *reinterpret_cast<uint32_t*>(color + nonzero_offset) = c_r | c_g | c_b;
+
+    // We already know that the other base color is zero.
+    *reinterpret_cast<uint32_t*>(color + zero_offset) = 0;
+
+    color[8 + nonzero_offset + 0] =
+        (color[nonzero_offset + 0] * (2 * 0xaaab)) >> 17;
+    color[8 + nonzero_offset + 1] =
+        (color[nonzero_offset + 1] * (2 * 0xaaab)) >> 17;
+    color[8 + nonzero_offset + 2] =
+        (color[nonzero_offset + 2] * (2 * 0xaaab)) >> 17;
+
+    color[8 + zero_offset + 0] = (color[nonzero_offset + 0] * 0xaaab) >> 17;
+    color[8 + zero_offset + 1] = (color[nonzero_offset + 1] * 0xaaab) >> 17;
+    color[8 + zero_offset + 2] = (color[nonzero_offset + 2] * 0xaaab) >> 17;
+  }
+}
+
+// The color matching function.
+template <typename T>
+uint32_t MatchColorsBlock(const uint8_t* block, uint8_t* color) {
+  int dirr = color[0 * 4 + 0] - color[1 * 4 + 0];
+  int dirg = color[0 * 4 + 1] - color[1 * 4 + 1];
+  int dirb = color[0 * 4 + 2] - color[1 * 4 + 2];
+
+  int stops[4];
+  for (int i = 0; i < 4; ++i) {
+    stops[i] = color[i * 4 + 0] * dirr + color[i * 4 + 1] * dirg +
+               color[i * 4 + 2] * dirb;
+  }
+
+  // Think of the colors as arranged on a line; project point onto that line,
+  // then choose next color out of available ones. we compute the crossover
+  // points for "best color in top half"/"best in bottom half" and then the same
+  // inside that subinterval.
+  //
+  // Relying on this 1d approximation isn't always optimal in terms of euclidean
+  // distance, but it's very close and a lot faster.
+  // http://cbloomrants.blogspot.com/2008/12/12-08-08-dxtc-summary.html
+
+  int c0_point = (stops[1] + stops[3]) >> 1;
+  int half_point = (stops[3] + stops[2]) >> 1;
+  int c3_point = (stops[2] + stops[0]) >> 1;
+
+  uint32_t mask = 0;
+  for (int i = 0; i < 16; i++) {
+    int dot = block[i * 4 + 0] * dirr + block[i * 4 + 1] * dirg +
+              block[i * 4 + 2] * dirb;
+
+    int bits = ((dot < half_point) ? 4 : 0) | ((dot < c0_point) ? 2 : 0) |
+               ((dot < c3_point) ? 1 : 0);
+
+    mask >>= 2;
+    mask |= T::kRemap[bits];
+  }
+
+  return mask;
+}
+
+void GetBaseColors(const uint8_t* block,
+                   int v_r,
+                   int v_g,
+                   int v_b,
+                   uint16_t* pmax16,
+                   uint16_t* pmin16) {
+// Pick colors at extreme points.
+#ifdef VERIFY_RESULTS
+  // Rewritten to match the SIMD implementation, not as efficient.
+  int dots[16];
+  for (int i = 0; i < 16; ++i) {
+    dots[i] = block[i * 4 + 0] * v_r + block[i * 4 + 1] * v_g +
+              block[i * 4 + 2] * v_b;
+  }
+  int max_dot = dots[0];
+  int min_dot = dots[0];
+  for (int i = 1; i < 16; ++i) {
+    if (dots[i] > max_dot)
+      max_dot = dots[i];
+    if (dots[i] < min_dot)
+      min_dot = dots[i];
+  }
+  uint32_t max_pixels[16];
+  uint32_t min_pixels[16];
+  for (int i = 0; i < 16; ++i) {
+    const uint32_t* p = reinterpret_cast<const uint32_t*>(block) + i;
+    max_pixels[i] = (dots[i] == max_dot) ? *p : 0;
+    min_pixels[i] = (dots[i] == min_dot) ? *p : 0;
+  }
+  uint32_t max_pixel = max_pixels[0];
+  uint32_t min_pixel = min_pixels[0];
+  for (int i = 1; i < 16; ++i) {
+    if (max_pixels[i] > max_pixel) {
+      max_pixel = max_pixels[i];
+    }
+    if (min_pixels[i] > min_pixel) {
+      min_pixel = min_pixels[i];
+    }
+  }
+  uint8_t* maxp = reinterpret_cast<uint8_t*>(&max_pixel);
+  uint8_t* minp = reinterpret_cast<uint8_t*>(&min_pixel);
+#else
+  int mind = 0x7fffffff;
+  int maxd = -0x7fffffff;
+  const uint8_t* minp = block;
+  const uint8_t* maxp = block;
+  for (int i = 0; i < 16; ++i) {
+    int dot = block[i * 4 + 0] * v_r + block[i * 4 + 1] * v_g +
+              block[i * 4 + 2] * v_b;
+
+    if (dot < mind) {
+      mind = dot;
+      minp = block + i * 4;
+    }
+
+    if (dot > maxd) {
+      maxd = dot;
+      maxp = block + i * 4;
+    }
+  }
+#endif
+
+  *pmax16 = As16Bit(maxp[0], maxp[1], maxp[2]);
+  *pmin16 = As16Bit(minp[0], minp[1], minp[2]);
+}
+
+// Figure out the two base colors to use from a block of 16 pixels
+// by Primary Component Analysis and map along principal axis.
+//
+// params:
+//  block   16 32-bit RGBX colors.
+//  pmax16  (output) base color 0 (minimum value), 16-bit RGB
+//  pmin16  (output) base color 1 (maximum value), 16-bit RGB
+void OptimizeColorsBlock(const uint8_t* block,
+                         uint16_t* pmax16,
+                         uint16_t* pmin16) {
+  // Determine color distribution.
+  int mu[3];
+  int min[3];
+  int max[3];
+  for (int ch = 0; ch < 3; ++ch) {
+    const uint8_t* bp = block + ch;
+    int muv = bp[0];
+    int minv = muv;
+    int maxv = muv;
+    for (int i = 4; i < 64; i += 4) {
+      int pixel = bp[i];
+      muv += pixel;
+      if (pixel < minv) {
+        minv = pixel;
+      } else if (pixel > maxv) {
+        maxv = pixel;
+      }
+    }
+
+    mu[ch] = (muv + 8) >> 4;
+    min[ch] = minv;
+    max[ch] = maxv;
+  }
+
+  // Determine covariance matrix.
+  int cov[6] = {0, 0, 0, 0, 0, 0};
+  for (int i = 0; i < 16; ++i) {
+    int r = block[i * 4 + 0] - mu[0];
+    int g = block[i * 4 + 1] - mu[1];
+    int b = block[i * 4 + 2] - mu[2];
+
+    cov[0] += r * r;
+    cov[1] += r * g;
+    cov[2] += r * b;
+    cov[3] += g * g;
+    cov[4] += g * b;
+    cov[5] += b * b;
+  }
+
+  // Convert covariance matrix to float, find principal axis via power iter.
+  float covf[6];
+  for (int i = 0; i < 6; ++i) {
+    covf[i] = cov[i] / 255.0f;
+  }
+
+  float vfr = static_cast<float>(max[0] - min[0]);
+  float vfg = static_cast<float>(max[1] - min[1]);
+  float vfb = static_cast<float>(max[2] - min[2]);
+
+  // Iterate to the power of 4.
+  for (int iter = 0; iter < 4; ++iter) {
+    float r = vfr * covf[0] + vfg * covf[1] + vfb * covf[2];
+    float g = vfr * covf[1] + vfg * covf[3] + vfb * covf[4];
+    float b = vfr * covf[2] + vfg * covf[4] + vfb * covf[5];
+
+    vfr = r;
+    vfg = g;
+    vfb = b;
+  }
+
+  double magn = std::abs(vfr);
+  double mag_g = std::abs(vfg);
+  double mag_b = std::abs(vfb);
+  if (mag_g > magn) {
+    magn = mag_g;
+  }
+  if (mag_b > magn) {
+    magn = mag_b;
+  }
+
+  int v_r = 299;  // JPEG YCbCr luma coefs, scaled by 1000.
+  int v_g = 587;
+  int v_b = 114;
+  if (magn >= 4.0f) {  // Too small, default to luminance.
+    magn = 512.0 / magn;
+    v_r = static_cast<int>(vfr * magn);
+    v_g = static_cast<int>(vfg * magn);
+    v_b = static_cast<int>(vfb * magn);
+  }
+
+  GetBaseColors(block, v_r, v_g, v_b, pmax16, pmin16);
+}
+
+// Figure out the two base colors simply using a direction vector between min
+// and max colors.
+//
+// params:
+//  block   16 32-bit RGBX colors.
+//  pmax16  (output) base color 0 (minimum value), 16-bit RGB
+//  pmin16  (output) base color 1 (maximum value), 16-bit RGB
+void GetApproximateBaseColors(const uint8_t* block,
+                              uint16_t* pmax16,
+                              uint16_t* pmin16) {
+  uint8_t dir[3];
+  for (int ch = 0; ch < 3; ++ch) {
+    const uint8_t* bp = block + ch;
+    uint8_t minv = bp[0];
+    uint8_t maxv = bp[0];
+    for (int i = 4; i < 64; i += 4) {
+      uint8_t pixel = bp[i];
+      if (pixel < minv) {
+        minv = pixel;
+      } else if (pixel > maxv) {
+        maxv = pixel;
+      }
+    }
+
+    dir[ch] = maxv - minv;
+  }
+
+  GetBaseColors(block, dir[0], dir[1], dir[2], pmax16, pmin16);
+}
+
+// The refinement function.
+// Tries to optimize colors to suit block contents better.
+// (By solving a least squares system via normal equations+Cramer's rule)
+//
+// params:
+//  block   16 32-bit RGBX colors.
+//  pmax16  (output) base color 0 (minimum value), 16-bit RGB
+//  pmin16  (output) base color 1 (maximum value), 16-bit RGB
+//  mask    16 2-bit color indices.
+template <typename T>
+int RefineBlock(const uint8_t* block,
+                uint16_t* pmax16,
+                uint16_t* pmin16,
+                uint32_t mask) {
+  uint16_t min16 = 0;
+  uint16_t max16 = 0;
+  if ((mask ^ (mask << 2)) < 4) {  // All pixels have the same index?
+    // Yes, linear system would be singular; solve using optimal
+    // single-color match on average color.
+    int r = 8;
+    int g = 8;
+    int b = 8;
+    for (int i = 0; i < 16; ++i) {
+      r += block[i * 4 + 0];
+      g += block[i * 4 + 1];
+      b += block[i * 4 + 2];
+    }
+
+    r >>= 4;
+    g >>= 4;
+    b >>= 4;
+
+    max16 = MatchSingleColorMax<typename T::BASE_TYPE>(r, g, b);
+    min16 = MatchSingleColorMin<typename T::BASE_TYPE>(r, g, b);
+  } else {
+    int at1_r = 0;
+    int at1_g = 0;
+    int at1_b = 0;
+    int at2_r = 0;
+    int at2_g = 0;
+    int at2_b = 0;
+    int akku = 0;
+    uint32_t cm = mask;
+    for (int i = 0; i < 16; ++i, cm >>= 2) {
+      int step = cm & 3;
+
+      int w1 = T::kW1Table[step];
+      int r = block[i * 4 + 0];
+      int g = block[i * 4 + 1];
+      int b = block[i * 4 + 2];
+
+      // Some magic to save a lot of multiplies in the accumulating loop...
+      // (Precomputed products of weights for least squares system, accumulated
+      // inside one 32-bit register.)
+      akku += T::kProds[step];
+      at1_r += w1 * r;
+      at1_g += w1 * g;
+      at1_b += w1 * b;
+      at2_r += r;
+      at2_g += g;
+      at2_b += b;
+    }
+
+    at2_r = 3 * at2_r - at1_r;
+    at2_g = 3 * at2_g - at1_g;
+    at2_b = 3 * at2_b - at1_b;
+
+    // Extract solutions and decide solvability.
+    int xx = akku >> 16;
+    int yy = (akku >> 8) & 0xff;
+    int xy = (akku >> 0) & 0xff;
+
+    float frb = 3.0f * 31.0f / 255.0f / (xx * yy - xy * xy);
+    float fg = frb * 63.0f / 31.0f;
+
+    // Solve.
+    max16 = sclamp((at1_r * yy - at2_r * xy) * frb + 0.5f, 0, 31) << 11;
+    max16 |= sclamp((at1_g * yy - at2_g * xy) * fg + 0.5f, 0, 63) << 5;
+    max16 |= sclamp((at1_b * yy - at2_b * xy) * frb + 0.5f, 0, 31) << 0;
+
+    min16 = sclamp((at2_r * xx - at1_r * xy) * frb + 0.5f, 0, 31) << 11;
+    min16 |= sclamp((at2_g * xx - at1_g * xy) * fg + 0.5f, 0, 63) << 5;
+    min16 |= sclamp((at2_b * xx - at1_b * xy) * frb + 0.5f, 0, 31) << 0;
+  }
+
+  uint16_t oldMin = *pmin16;
+  uint16_t oldMax = *pmax16;
+  *pmin16 = min16;
+  *pmax16 = max16;
+  return oldMin != min16 || oldMax != max16;
+}
+
+// Color block compression.
+template <typename T, Quality QUALITY>
+void CompressColorBlock(uint8_t* dst, const uint8_t* block) {
+  // Check if block is constant.
+  int i = 1;
+  uint32_t first_pixel =
+      reinterpret_cast<const uint32_t*>(block)[0] & 0x00ffffff;
+  for (; i < 16; ++i) {
+    if ((reinterpret_cast<const uint32_t*>(block)[i] & 0x00ffffff) !=
+        first_pixel) {
+      break;
+    }
+  }
+
+  uint32_t mask = 0;
+  uint16_t max16 = 0;
+  uint16_t min16 = 0;
+  if (i == 16) {  // Constant color
+    int r = block[0];
+    int g = block[1];
+    int b = block[2];
+    max16 = MatchSingleColorMax<typename T::BASE_TYPE>(r, g, b);
+    min16 = MatchSingleColorMin<typename T::BASE_TYPE>(r, g, b);
+    mask = T::kConstantColorIndices;
+  } else {
+    if (QUALITY == kQualityLow) {
+      GetApproximateBaseColors(block, &max16, &min16);
+    } else {
+      // Do Primary Component Analysis and map along principal axis.
+      OptimizeColorsBlock(block, &max16, &min16);
+    }
+
+    if (max16 != min16) {
+      uint8_t color[4 * 4];
+      EvalColors(color, max16, min16);
+      mask = MatchColorsBlock<T>(block, color);
+    }
+
+    if (QUALITY == kQualityHigh) {
+      // Refine (multiple times if requested).
+      for (int i = 0; i < kNumRefinements; ++i) {
+        uint32_t lastmask = mask;
+
+        if (RefineBlock<T>(block, &max16, &min16, mask)) {
+          if (max16 != min16) {
+            uint8_t color[4 * 4];
+            EvalColors(color, max16, min16);
+            mask = MatchColorsBlock<T>(block, color);
+          } else {
+            mask = 0;
+            break;
+          }
+        }
+
+        if (mask == lastmask) {
+          break;
+        }
+      }
+    }
+  }
+
+#if defined(OS_ANDROID)
+  // Swapping r and b channels to match Skia.
+  // See skia/config/SkUserConfig.h
+  max16 = RGB2BGR(max16);
+  min16 = RGB2BGR(min16);
+#endif
+
+  FormatFixup_Generic<typename T::BASE_TYPE>(&max16, &min16, &mask);
+
+  uint32_t* dst32 = reinterpret_cast<uint32_t*>(dst);
+  dst32[0] = max16 | (min16 << 16);
+  dst32[1] = mask;
+}
+
+// Alpha block compression.
+void CompressAlphaBlock(uint8_t* dst, const uint8_t* src) {
+  // Find min/max alpha.
+  int mn = src[3];
+  int mx = mn;
+  for (int i = 1; i < 16; ++i) {
+    int alpha = src[i * 4 + 3];
+    if (alpha < mn) {
+      mn = alpha;
+    } else if (alpha > mx) {
+      mx = alpha;
+    }
+  }
+
+  // Encode them.
+  dst[0] = mx;
+  dst[1] = mn;
+  dst += 2;
+
+  if (mx == mn) {
+    memset(dst, 0, 6);
+  } else {
+    // Determine bias and emit color indices.
+    // Given the choice of mx/mn, these indices are optimal:
+    // http://fgiesen.wordpress.com/2009/12/15/dxt5-alpha-block-index-determination/
+    int dist = mx - mn;
+    int dist4 = dist * 4;
+    int dist2 = dist * 2;
+    int bias = (dist < 8) ? (dist - 1) : (dist / 2 + 2);
+    bias -= mn * 7;
+    int bits = 0;
+    int mask = 0;
+
+    for (int i = 0; i < 16; ++i) {
+      int a = src[i * 4 + 3] * 7 + bias;
+
+      // Select index. this is a "linear scale" lerp factor between 0 (val=min)
+      // and 7 (val=max).
+      int t = (a >= dist4) ? -1 : 0;
+      int ind = t & 4;
+      a -= dist4 & t;
+
+      t = (a >= dist2) ? -1 : 0;
+      ind += t & 2;
+      a -= dist2 & t;
+
+      ind += (a >= dist);
+
+      // Turn linear scale into DXT index (0/1 are extremal pts).
+      ind = -ind & 7;
+      ind ^= (2 > ind);
+
+      // Write index.
+      mask |= ind << bits;
+      if ((bits += 3) >= 8) {
+        *dst++ = mask;
+        mask >>= 8;
+        bits -= 8;
+      }
+    }
+  }
+}
+
+static void ExtractBlock(uint8_t* dst, const uint8_t* src, int width) {
+  for (int j = 0; j < 4; ++j) {
+    memcpy(&dst[j * 4 * 4], src, 4 * 4);
+    src += width * 4;
+  }
+}
+
+template <typename T, bool OPAQUE, Quality QUALITY>
+void CompressImage(const uint8_t* src, uint8_t* dst, int width, int height) {
+  for (int y = 0; y < height; y += 4, src += width * 4 * 4) {
+    for (int x = 0; x < width; x += 4) {
+      uint8_t block[64];
+      ExtractBlock(block, src + x * 4, width);
+
+      if (!OPAQUE) {
+        CompressAlphaBlock(dst, block);
+        dst += 8;
+      }
+
+      CompressColorBlock<T, QUALITY>(dst, block);
+      dst += 8;
+    }
+  }
+}
+
+}  // namespace
+
+void Init_ATC_DXT() {
+  uint8_t expand5[32];
+  for (int i = 0; i < 32; ++i) {
+    expand5[i] = Expand5(i);
+  }
+
+  uint8_t expand6[64];
+  for (int i = 0; i < 64; ++i) {
+    expand6[i] = Expand6(i);
+  }
+
+  PrepareOptTable(&g_o_match55[0][0], expand5, expand5, 32, 32);
+  PrepareOptTable(&g_o_match66[0][0], expand6, expand6, 64, 64);
+  PrepareOptTable(&g_o_match56[0][0], expand5, expand6, 32, 64);
+}
+
+void CompressATC_Generic(const uint8_t* src,
+                         uint8_t* dst,
+                         int width,
+                         int height) {
+  CompressImage<TYPE_ATC_GENERIC, true, kQualityHigh>(src, dst, width, height);
+}
+
+void CompressATCIA_Generic(const uint8_t* src,
+                           uint8_t* dst,
+                           int width,
+                           int height) {
+  CompressImage<TYPE_ATC_GENERIC, false, kQualityHigh>(src, dst, width, height);
+}
+
+void CompressDXT1_Generic(const uint8_t* src,
+                          uint8_t* dst,
+                          int width,
+                          int height) {
+  CompressImage<TYPE_DXT_GENERIC, true, kQualityHigh>(src, dst, width, height);
+}
+
+void CompressDXT5_Generic(const uint8_t* src,
+                          uint8_t* dst,
+                          int width,
+                          int height) {
+  CompressImage<TYPE_DXT_GENERIC, false, kQualityHigh>(src, dst, width, height);
+}
+
+}  // namespace texture_compress
+}  // namespace cc