gecko-dev/gfx/2d/Swizzle.cpp

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

#include "Swizzle.h"
#include "Logging.h"
#include "Tools.h"
#include "mozilla/CheckedInt.h"
#include "mozilla/EndianUtils.h"

#ifdef BUILD_ARM_NEON
#include "mozilla/arm.h"
#endif

namespace mozilla {
namespace gfx {

/**
 * Convenience macros for dispatching to various format combinations.
 */

// Hash the formats to a relatively dense value to optimize jump table generation.
// The first 6 formats in SurfaceFormat are the 32-bit BGRA variants and are the most
// common formats dispatched here. Room is reserved in the lowish bits for up to
// these 6 destination formats. If a destination format is >= 6, the 6th bit is set
// to avoid collisions.
#define FORMAT_KEY(aSrcFormat, aDstFormat) \
  (int(aSrcFormat) * 6 + int(aDstFormat) + (int(int(aDstFormat) >= 6) << 6))

#define FORMAT_CASE_EXPR(aSrcFormat, aDstFormat, ...) \
  case FORMAT_KEY(aSrcFormat, aDstFormat): \
    __VA_ARGS__; \
    return true;

#define FORMAT_CASE(aSrcFormat, aDstFormat, ...) \
  FORMAT_CASE_EXPR(aSrcFormat, aDstFormat, FORMAT_CASE_CALL(__VA_ARGS__))

/**
 * Constexpr functions for analyzing format attributes in templates.
 */

// Whether B comes before R in pixel memory layout.
static constexpr bool
IsBGRFormat(SurfaceFormat aFormat)
{
  return aFormat == SurfaceFormat::B8G8R8A8 ||
#if MOZ_LITTLE_ENDIAN
         aFormat == SurfaceFormat::R5G6B5_UINT16 ||
#endif
         aFormat == SurfaceFormat::B8G8R8X8 ||
         aFormat == SurfaceFormat::B8G8R8;
}

// Whether the order of B and R need to be swapped to map from src to dst.
static constexpr bool
ShouldSwapRB(SurfaceFormat aSrcFormat, SurfaceFormat aDstFormat)
{
  return IsBGRFormat(aSrcFormat) != IsBGRFormat(aDstFormat);
}

// The starting byte of the RGB components in pixel memory.
static constexpr uint32_t
RGBByteIndex(SurfaceFormat aFormat)
{
  return aFormat == SurfaceFormat::A8R8G8B8 ||
         aFormat == SurfaceFormat::X8R8G8B8
           ? 1 : 0;
}

// The byte of the alpha component, which just comes after RGB.
static constexpr uint32_t
AlphaByteIndex(SurfaceFormat aFormat)
{
  return (RGBByteIndex(aFormat) + 3) % 4;
}

// The endian-dependent bit shift to access RGB of a UINT32 pixel.
static constexpr uint32_t
RGBBitShift(SurfaceFormat aFormat)
{
#if MOZ_LITTLE_ENDIAN
  return 8 * RGBByteIndex(aFormat);
#else
  return 24 - 8 * RGBByteIndex(aFormat);
#endif
}

// The endian-dependent bit shift to access alpha of a UINT32 pixel.
static constexpr uint32_t
AlphaBitShift(SurfaceFormat aFormat)
{
  return (RGBBitShift(aFormat) + 24) % 32;
}

// Whether the pixel format should ignore the value of the alpha channel and treat it as opaque.
static constexpr bool
IgnoreAlpha(SurfaceFormat aFormat)
{
  return aFormat == SurfaceFormat::B8G8R8X8 ||
         aFormat == SurfaceFormat::R8G8B8X8 ||
         aFormat == SurfaceFormat::X8R8G8B8;
}

// Whether to force alpha to opaque to map from src to dst.
static constexpr bool
ShouldForceOpaque(SurfaceFormat aSrcFormat, SurfaceFormat aDstFormat)
{
  return IgnoreAlpha(aSrcFormat) != IgnoreAlpha(aDstFormat);
}

#ifdef USE_SSE2
/**
 * SSE2 optimizations
 */

template<bool aSwapRB, bool aOpaqueAlpha>
void Premultiply_SSE2(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define PREMULTIPLY_SSE2(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Premultiply_SSE2 \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat)>)

template<bool aSwapRB>
void Unpremultiply_SSE2(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define UNPREMULTIPLY_SSE2(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Unpremultiply_SSE2<ShouldSwapRB(aSrcFormat, aDstFormat)>)

template<bool aSwapRB, bool aOpaqueAlpha>
void Swizzle_SSE2(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define SWIZZLE_SSE2(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Swizzle_SSE2 \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat)>)

#endif

#ifdef BUILD_ARM_NEON
/**
 * ARM NEON optimizations
 */

template<bool aSwapRB, bool aOpaqueAlpha>
void Premultiply_NEON(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define PREMULTIPLY_NEON(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Premultiply_NEON \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat)>)

template<bool aSwapRB>
void Unpremultiply_NEON(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define UNPREMULTIPLY_NEON(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Unpremultiply_NEON<ShouldSwapRB(aSrcFormat, aDstFormat)>)

template<bool aSwapRB, bool aOpaqueAlpha>
void Swizzle_NEON(const uint8_t*, int32_t, uint8_t*, int32_t, IntSize);

#define SWIZZLE_NEON(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    Swizzle_NEON \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat)>)

#endif

/**
 * Premultiplying
 */

// Fallback premultiply implementation that uses splayed pixel math to reduce the
// multiplications used. That is, the R and B components are isolated from the G and A
// components, which then can be multiplied as if they were two 2-component vectors.
// Otherwise, an approximation if divide-by-255 is used which is faster than an actual
// division. These optimizations are also used for the SSE2 and NEON implementations.
template<bool aSwapRB, bool aOpaqueAlpha,
         uint32_t aSrcRGBShift, uint32_t aSrcAShift,
         uint32_t aDstRGBShift, uint32_t aDstAShift>
static void
PremultiplyFallback(const uint8_t* aSrc, int32_t aSrcGap,
                    uint8_t* aDst, int32_t aDstGap,
                    IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      // Load and process 1 entire pixel at a time.
      uint32_t color = *reinterpret_cast<const uint32_t*>(aSrc);

      uint32_t a = aSrcAShift ? color >> aSrcAShift : color & 0xFF;

      // Isolate the R and B components.
      uint32_t rb = (color >> aSrcRGBShift) & 0x00FF00FF;
      // Swap the order of R and B if necessary.
      if (aSwapRB) {
        rb = (rb >> 16) | (rb << 16);
      }
      // Approximate the multiply by alpha and divide by 255 which is essentially:
      // c = c*a + 255; c = (c + (c >> 8)) >> 8;
      // However, we omit the final >> 8 to fold it with the final shift into place
      // depending on desired output format.
      rb = rb*a + 0x00FF00FF;
      rb = (rb + ((rb >> 8) & 0x00FF00FF)) & 0xFF00FF00;

      // Use same approximation as above, but G is shifted 8 bits left.
      // Alpha is left out and handled separately.
      uint32_t g = color & (0xFF00 << aSrcRGBShift);
      g = g*a + (0xFF00 << aSrcRGBShift);
      g = (g + (g >> 8)) & (0xFF0000 << aSrcRGBShift);

      // The above math leaves RGB shifted left by 8 bits.
      // Shift them right if required for the output format.
      // then combine them back together to produce output pixel.
      // Add the alpha back on if the output format is not opaque.
      *reinterpret_cast<uint32_t*>(aDst) =
         (rb >> (8 - aDstRGBShift)) |
         (g >> (8 + aSrcRGBShift - aDstRGBShift)) |
         (aOpaqueAlpha ? 0xFF << aDstAShift : a << aDstAShift);

      aSrc += 4;
      aDst += 4;
    } while (aSrc < end);

    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define PREMULTIPLY_FALLBACK_CASE(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    PremultiplyFallback \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat), \
       RGBBitShift(aSrcFormat), AlphaBitShift(aSrcFormat), \
       RGBBitShift(aDstFormat), AlphaBitShift(aDstFormat)>)

#define PREMULTIPLY_FALLBACK(aSrcFormat) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::B8G8R8A8) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::B8G8R8X8) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::R8G8B8A8) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::R8G8B8X8) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::A8R8G8B8) \
  PREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::X8R8G8B8)

// If rows are tightly packed, and the size of the total area will fit within
// the precision range of a single row, then process all the data as if it was
// a single row.
static inline IntSize
CollapseSize(const IntSize& aSize, int32_t aSrcStride, int32_t aDstStride)
{
  if (aSrcStride == aDstStride &&
      (aSrcStride & 3) == 0 &&
      aSrcStride / 4 == aSize.width) {
    CheckedInt32 area = CheckedInt32(aSize.width) * CheckedInt32(aSize.height);
    if (area.isValid()) {
      return IntSize(area.value(), 1);
    }
  }
  return aSize;
}

static inline int32_t
GetStrideGap(int32_t aWidth, SurfaceFormat aFormat, int32_t aStride)
{
  CheckedInt32 used = CheckedInt32(aWidth) * BytesPerPixel(aFormat);
  if (!used.isValid() || used.value() < 0) {
    return -1;
  }
  return aStride - used.value();
}

bool
PremultiplyData(const uint8_t* aSrc, int32_t aSrcStride, SurfaceFormat aSrcFormat,
                uint8_t* aDst, int32_t aDstStride, SurfaceFormat aDstFormat,
                const IntSize& aSize)
{
  if (aSize.IsEmpty()) {
    return true;
  }
  IntSize size = CollapseSize(aSize, aSrcStride, aDstStride);
  // Find gap from end of row to the start of the next row.
  int32_t srcGap = GetStrideGap(aSize.width, aSrcFormat, aSrcStride);
  int32_t dstGap = GetStrideGap(aSize.width, aDstFormat, aDstStride);
  MOZ_ASSERT(srcGap >= 0 && dstGap >= 0);
  if (srcGap < 0 || dstGap < 0) {
    return false;
  }

#define FORMAT_CASE_CALL(...) __VA_ARGS__(aSrc, srcGap, aDst, dstGap, size)

#ifdef USE_SSE2
  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  PREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8A8)
  PREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8X8)
  PREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  PREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8X8)
  PREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8A8)
  PREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8X8)
  PREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  PREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8X8)
  default: break;
  }
#endif

#ifdef BUILD_ARM_NEON
  if (mozilla::supports_neon()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  PREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8A8)
  PREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8X8)
  PREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  PREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8X8)
  PREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8A8)
  PREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8X8)
  PREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  PREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8X8)
  default: break;
  }
#endif

  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  PREMULTIPLY_FALLBACK(SurfaceFormat::B8G8R8A8)
  PREMULTIPLY_FALLBACK(SurfaceFormat::R8G8B8A8)
  PREMULTIPLY_FALLBACK(SurfaceFormat::A8R8G8B8)
  default: break;
  }

#undef FORMAT_CASE_CALL

  MOZ_ASSERT(false, "Unsupported premultiply formats");
  return false;
}

/**
 * Unpremultiplying
 */

// Generate a table of 8.16 fixed-point reciprocals representing 1/alpha.
#define UNPREMULQ(x) (0xFF00FFU / (x))
#define UNPREMULQ_2(x) UNPREMULQ(x), UNPREMULQ((x) + 1)
#define UNPREMULQ_4(x) UNPREMULQ_2(x), UNPREMULQ_2((x) + 2)
#define UNPREMULQ_8(x) UNPREMULQ_4(x), UNPREMULQ_4((x) + 4)
#define UNPREMULQ_16(x) UNPREMULQ_8(x), UNPREMULQ_8((x) + 8)
#define UNPREMULQ_32(x) UNPREMULQ_16(x), UNPREMULQ_16((x) + 16)
static const uint32_t sUnpremultiplyTable[256] =
{
  0, UNPREMULQ(1), UNPREMULQ_2(2), UNPREMULQ_4(4),
  UNPREMULQ_8(8), UNPREMULQ_16(16), UNPREMULQ_32(32),
  UNPREMULQ_32(64), UNPREMULQ_32(96), UNPREMULQ_32(128),
  UNPREMULQ_32(160), UNPREMULQ_32(192), UNPREMULQ_32(224)
};

// Fallback unpremultiply implementation that uses 8.16 fixed-point reciprocal math
// to eliminate any division by the alpha component. This optimization is used for the
// SSE2 and NEON implementations, with some adaptations. This implementation also accesses
// color components using individual byte accesses as this profiles faster than accessing
// the pixel as a uint32_t and shifting/masking to access components.
template<bool aSwapRB,
         uint32_t aSrcRGBIndex, uint32_t aSrcAIndex,
         uint32_t aDstRGBIndex, uint32_t aDstAIndex>
static void
UnpremultiplyFallback(const uint8_t* aSrc, int32_t aSrcGap,
                      uint8_t* aDst, int32_t aDstGap,
                      IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      uint8_t r = aSrc[aSrcRGBIndex + (aSwapRB ? 2 : 0)];
      uint8_t g = aSrc[aSrcRGBIndex + 1];
      uint8_t b = aSrc[aSrcRGBIndex + (aSwapRB ? 0 : 2)];
      uint8_t a = aSrc[aSrcAIndex];

      // Access the 8.16 reciprocal from the table based on alpha. Multiply by the
      // reciprocal and shift off the fraction bits to approximate the division by alpha.
      uint32_t q = sUnpremultiplyTable[a];
      aDst[aDstRGBIndex + 0] = (r * q) >> 16;
      aDst[aDstRGBIndex + 1] = (g * q) >> 16;
      aDst[aDstRGBIndex + 2] = (b * q) >> 16;
      aDst[aDstAIndex] = a;

      aSrc += 4;
      aDst += 4;
    } while (aSrc < end);

    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define UNPREMULTIPLY_FALLBACK_CASE(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    UnpremultiplyFallback \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       RGBByteIndex(aSrcFormat), AlphaByteIndex(aSrcFormat), \
       RGBByteIndex(aDstFormat), AlphaByteIndex(aDstFormat)>)

#define UNPREMULTIPLY_FALLBACK(aSrcFormat) \
  UNPREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::B8G8R8A8) \
  UNPREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::R8G8B8A8) \
  UNPREMULTIPLY_FALLBACK_CASE(aSrcFormat, SurfaceFormat::A8R8G8B8)

bool
UnpremultiplyData(const uint8_t* aSrc, int32_t aSrcStride, SurfaceFormat aSrcFormat,
                  uint8_t* aDst, int32_t aDstStride, SurfaceFormat aDstFormat,
                  const IntSize& aSize)
{
  if (aSize.IsEmpty()) {
    return true;
  }
  IntSize size = CollapseSize(aSize, aSrcStride, aDstStride);
  // Find gap from end of row to the start of the next row.
  int32_t srcGap = GetStrideGap(aSize.width, aSrcFormat, aSrcStride);
  int32_t dstGap = GetStrideGap(aSize.width, aDstFormat, aDstStride);
  MOZ_ASSERT(srcGap >= 0 && dstGap >= 0);
  if (srcGap < 0 || dstGap < 0) {
    return false;
  }

#define FORMAT_CASE_CALL(...) __VA_ARGS__(aSrc, srcGap, aDst, dstGap, size)

#ifdef USE_SSE2
  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  UNPREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8A8)
  UNPREMULTIPLY_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  UNPREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8A8)
  UNPREMULTIPLY_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  default: break;
  }
#endif

#ifdef BUILD_ARM_NEON
  if (mozilla::supports_neon()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  UNPREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8A8)
  UNPREMULTIPLY_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  UNPREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8A8)
  UNPREMULTIPLY_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  default: break;
  }
#endif

  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  UNPREMULTIPLY_FALLBACK(SurfaceFormat::B8G8R8A8)
  UNPREMULTIPLY_FALLBACK(SurfaceFormat::R8G8B8A8)
  UNPREMULTIPLY_FALLBACK(SurfaceFormat::A8R8G8B8)
  default: break;
  }

#undef FORMAT_CASE_CALL

  MOZ_ASSERT(false, "Unsupported unpremultiply formats");
  return false;
}

/**
 * Swizzling
 */

// Fallback swizzle implementation that uses shifting and masking to reorder pixels.
template<bool aSwapRB, bool aOpaqueAlpha,
         uint32_t aSrcRGBShift, uint32_t aSrcAShift,
         uint32_t aDstRGBShift, uint32_t aDstAShift>
static void
SwizzleFallback(const uint8_t* aSrc, int32_t aSrcGap,
                uint8_t* aDst, int32_t aDstGap,
                IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      uint32_t rgba = *reinterpret_cast<const uint32_t*>(aSrc);

      if (aSwapRB) {
        // Handle R and B swaps by exchanging words and masking.
        uint32_t rb = ((rgba << 16) | (rgba >> 16)) & (0x00FF00FF << aSrcRGBShift);
        uint32_t ga = rgba & ((0xFF << aSrcAShift) | (0xFF00 << aSrcRGBShift));
        rgba = rb | ga;
      }

      // If src and dst shifts differ, rotate left or right to move RGB into place,
      // i.e. ARGB -> RGBA or ARGB -> RGBA.
      if (aDstRGBShift > aSrcRGBShift) {
        rgba = (rgba << 8) | (aOpaqueAlpha ? 0x000000FF : rgba >> 24);
      } else if (aSrcRGBShift > aDstRGBShift) {
        rgba = (rgba >> 8) | (aOpaqueAlpha ? 0xFF000000 : rgba << 24);
      } else if (aOpaqueAlpha) {
        rgba |= 0xFF << aDstAShift;
      }

      *reinterpret_cast<uint32_t*>(aDst) = rgba;

      aSrc += 4;
      aDst += 4;
    } while (aSrc < end);

    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define SWIZZLE_FALLBACK(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    SwizzleFallback \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       ShouldForceOpaque(aSrcFormat, aDstFormat), \
       RGBBitShift(aSrcFormat), AlphaBitShift(aSrcFormat), \
       RGBBitShift(aDstFormat), AlphaBitShift(aDstFormat)>)

// Fast-path for matching formats.
static void
SwizzleCopy(const uint8_t* aSrc, int32_t aSrcGap,
            uint8_t* aDst, int32_t aDstGap,
            IntSize aSize, int32_t aBPP)
{
  if (aSrc != aDst) {
    int32_t rowLength = aBPP * aSize.width;
    for (int32_t height = aSize.height; height > 0; height--) {
      memcpy(aDst, aSrc, rowLength);
      aSrc += rowLength + aSrcGap;
      aDst += rowLength + aDstGap;
    }
  }
}

// Fast-path for conversions that swap all bytes.
template<bool aOpaqueAlpha, uint32_t aSrcAShift, uint32_t aDstAShift>
static void
SwizzleSwap(const uint8_t* aSrc, int32_t aSrcGap,
            uint8_t* aDst, int32_t aDstGap,
            IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      // Use an endian swap to move the bytes, i.e. BGRA -> ARGB.
      uint32_t rgba = *reinterpret_cast<const uint32_t*>(aSrc);
#if MOZ_LITTLE_ENDIAN
      rgba = NativeEndian::swapToBigEndian(rgba);
#else
      rgba = NativeEndian::swapToLittleEndian(rgba);
#endif
      if (aOpaqueAlpha) {
        rgba |= 0xFF << aDstAShift;
      }
      *reinterpret_cast<uint32_t*>(aDst) = rgba;
      aSrc += 4;
      aDst += 4;
    } while (aSrc < end);
    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define SWIZZLE_SWAP(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    SwizzleSwap \
      <ShouldForceOpaque(aSrcFormat, aDstFormat), \
       AlphaBitShift(aSrcFormat), AlphaBitShift(aDstFormat)>)

// Fast-path for conversions that force alpha to opaque.
template<uint32_t aDstAShift>
static void
SwizzleOpaque(const uint8_t* aSrc, int32_t aSrcGap,
              uint8_t* aDst, int32_t aDstGap,
              IntSize aSize)
{
  if (aSrc == aDst) {
    // Modifying in-place, so just write out the alpha.
    for (int32_t height = aSize.height; height > 0; height--) {
      const uint8_t* end = aDst + 4 * aSize.width;
      do {
        // ORing directly onto destination memory profiles faster than writing
        // individually to the alpha byte and also profiles equivalently to a
        // SSE2 implementation.
        *reinterpret_cast<uint32_t*>(aDst) |= 0xFF << aDstAShift;
        aDst += 4;
      } while (aDst < end);
      aDst += aDstGap;
    }
  } else {
    for (int32_t height = aSize.height; height > 0; height--) {
      const uint8_t* end = aSrc + 4 * aSize.width;
      do {
        uint32_t rgba = *reinterpret_cast<const uint32_t*>(aSrc);
        // Just add on the alpha bits to the source.
        rgba |= 0xFF << aDstAShift;
        *reinterpret_cast<uint32_t*>(aDst) = rgba;
        aSrc += 4;
        aDst += 4;
      } while (aSrc < end);
      aSrc += aSrcGap;
      aDst += aDstGap;
    }
  }
}

#define SWIZZLE_OPAQUE(aSrcFormat, aDstFormat) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    SwizzleOpaque<AlphaBitShift(aDstFormat)>)

// Packing of 32-bit formats to RGB565.
template<bool aSwapRB, uint32_t aSrcRGBShift, uint32_t aSrcRGBIndex>
static void
PackToRGB565(const uint8_t* aSrc, int32_t aSrcGap,
             uint8_t* aDst, int32_t aDstGap,
             IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      uint32_t rgba = *reinterpret_cast<const uint32_t*>(aSrc);

      // Isolate the R, G, and B components and shift to final endian-dependent locations.
      uint16_t rgb565;
      if (aSwapRB) {
        rgb565 = ((rgba & (0xF8 << aSrcRGBShift)) << (8 - aSrcRGBShift)) |
                 ((rgba & (0xFC00 << aSrcRGBShift)) >> (5 + aSrcRGBShift)) |
                 ((rgba & (0xF80000 << aSrcRGBShift)) >> (19 + aSrcRGBShift));
      } else {
        rgb565 = ((rgba & (0xF8 << aSrcRGBShift)) >> (3 + aSrcRGBShift)) |
                 ((rgba & (0xFC00 << aSrcRGBShift)) >> (5 + aSrcRGBShift)) |
                 ((rgba & (0xF80000 << aSrcRGBShift)) >> (8 + aSrcRGBShift));
      }

      *reinterpret_cast<uint16_t*>(aDst) = rgb565;

      aSrc += 4;
      aDst += 2;
    } while (aSrc < end);

    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

// Packing of 32-bit formats to 24-bit formats.
template<bool aSwapRB, uint32_t aSrcRGBShift, uint32_t aSrcRGBIndex>
static void
PackToRGB24(const uint8_t* aSrc, int32_t aSrcGap,
            uint8_t* aDst, int32_t aDstGap,
            IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      uint8_t r = aSrc[aSrcRGBIndex + (aSwapRB ? 2 : 0)];
      uint8_t g = aSrc[aSrcRGBIndex + 1];
      uint8_t b = aSrc[aSrcRGBIndex + (aSwapRB ? 0 : 2)];

      aDst[0] = r;
      aDst[1] = g;
      aDst[2] = b;

      aSrc += 4;
      aDst += 3;
    } while (aSrc < end);

    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define PACK_RGB_CASE(aSrcFormat, aDstFormat, aPackFunc) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    aPackFunc \
      <ShouldSwapRB(aSrcFormat, aDstFormat), \
       RGBBitShift(aSrcFormat), RGBByteIndex(aSrcFormat)>)

#define PACK_RGB(aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::B8G8R8A8, aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::B8G8R8X8, aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::R8G8B8A8, aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::R8G8B8X8, aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::A8R8G8B8, aDstFormat, aPackFunc) \
  PACK_RGB_CASE(SurfaceFormat::X8R8G8B8, aDstFormat, aPackFunc)

// Packing of 32-bit formats to A8.
template<uint32_t aSrcAIndex>
static void
PackToA8(const uint8_t* aSrc, int32_t aSrcGap,
         uint8_t* aDst, int32_t aDstGap,
         IntSize aSize)
{
  for (int32_t height = aSize.height; height > 0; height--) {
    const uint8_t* end = aSrc + 4 * aSize.width;
    do {
      *aDst++ = aSrc[aSrcAIndex];
      aSrc += 4;
    } while (aSrc < end);
    aSrc += aSrcGap;
    aDst += aDstGap;
  }
}

#define PACK_ALPHA_CASE(aSrcFormat, aDstFormat, aPackFunc) \
  FORMAT_CASE(aSrcFormat, aDstFormat, \
    aPackFunc<AlphaByteIndex(aSrcFormat)>)

#define PACK_ALPHA(aDstFormat, aPackFunc) \
  PACK_ALPHA_CASE(SurfaceFormat::B8G8R8A8, aDstFormat, aPackFunc) \
  PACK_ALPHA_CASE(SurfaceFormat::R8G8B8A8, aDstFormat, aPackFunc) \
  PACK_ALPHA_CASE(SurfaceFormat::A8R8G8B8, aDstFormat, aPackFunc)

bool
SwizzleData(const uint8_t* aSrc, int32_t aSrcStride, SurfaceFormat aSrcFormat,
            uint8_t* aDst, int32_t aDstStride, SurfaceFormat aDstFormat,
            const IntSize& aSize)
{
  if (aSize.IsEmpty()) {
    return true;
  }
  IntSize size = CollapseSize(aSize, aSrcStride, aDstStride);
  // Find gap from end of row to the start of the next row.
  int32_t srcGap = GetStrideGap(aSize.width, aSrcFormat, aSrcStride);
  int32_t dstGap = GetStrideGap(aSize.width, aDstFormat, aDstStride);
  MOZ_ASSERT(srcGap >= 0 && dstGap >= 0);
  if (srcGap < 0 || dstGap < 0) {
    return false;
  }

#define FORMAT_CASE_CALL(...) __VA_ARGS__(aSrc, srcGap, aDst, dstGap, size)

#ifdef USE_SSE2
  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  SWIZZLE_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_SSE2(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_SSE2(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_SSE2(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_SSE2(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_SSE2(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_SSE2(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8A8)
  default: break;
  }
#endif

#ifdef BUILD_ARM_NEON
  if (mozilla::supports_neon()) switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {
  SWIZZLE_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_NEON(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_NEON(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_NEON(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_NEON(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_NEON(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_NEON(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8A8)
  default: break;
  }
#endif

  switch (FORMAT_KEY(aSrcFormat, aDstFormat)) {

  SWIZZLE_FALLBACK(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_FALLBACK(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::B8G8R8A8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::B8G8R8X8, SurfaceFormat::R8G8B8A8)

  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8A8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8X8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8A8, SurfaceFormat::A8R8G8B8)
  SWIZZLE_FALLBACK(SurfaceFormat::R8G8B8X8, SurfaceFormat::X8R8G8B8)

  SWIZZLE_FALLBACK(SurfaceFormat::A8R8G8B8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_FALLBACK(SurfaceFormat::X8R8G8B8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::A8R8G8B8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_FALLBACK(SurfaceFormat::X8R8G8B8, SurfaceFormat::R8G8B8A8)

  SWIZZLE_SWAP(SurfaceFormat::B8G8R8A8, SurfaceFormat::A8R8G8B8)
  SWIZZLE_SWAP(SurfaceFormat::B8G8R8A8, SurfaceFormat::X8R8G8B8)
  SWIZZLE_SWAP(SurfaceFormat::B8G8R8X8, SurfaceFormat::X8R8G8B8)
  SWIZZLE_SWAP(SurfaceFormat::B8G8R8X8, SurfaceFormat::A8R8G8B8)
  SWIZZLE_SWAP(SurfaceFormat::A8R8G8B8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_SWAP(SurfaceFormat::A8R8G8B8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_SWAP(SurfaceFormat::X8R8G8B8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_SWAP(SurfaceFormat::X8R8G8B8, SurfaceFormat::B8G8R8A8)

  SWIZZLE_OPAQUE(SurfaceFormat::B8G8R8A8, SurfaceFormat::B8G8R8X8)
  SWIZZLE_OPAQUE(SurfaceFormat::B8G8R8X8, SurfaceFormat::B8G8R8A8)
  SWIZZLE_OPAQUE(SurfaceFormat::R8G8B8A8, SurfaceFormat::R8G8B8X8)
  SWIZZLE_OPAQUE(SurfaceFormat::R8G8B8X8, SurfaceFormat::R8G8B8A8)
  SWIZZLE_OPAQUE(SurfaceFormat::A8R8G8B8, SurfaceFormat::X8R8G8B8)
  SWIZZLE_OPAQUE(SurfaceFormat::X8R8G8B8, SurfaceFormat::A8R8G8B8)

  PACK_RGB(SurfaceFormat::R5G6B5_UINT16, PackToRGB565)
  PACK_RGB(SurfaceFormat::B8G8R8, PackToRGB24)
  PACK_RGB(SurfaceFormat::R8G8B8, PackToRGB24)
  PACK_ALPHA(SurfaceFormat::A8, PackToA8)

  default: break;
  }

  if (aSrcFormat == aDstFormat) {
    // If the formats match, just do a generic copy.
    SwizzleCopy(aSrc, srcGap, aDst, dstGap, size, BytesPerPixel(aSrcFormat));
    return true;
  }

#undef FORMAT_CASE_CALL

  MOZ_ASSERT(false, "Unsupported swizzle formats");
  return false;
}

} // namespace gfx
} // namespace mozilla