gecko-dev/image/SurfaceFilters.h

/* -*- 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/. */

/**
 * This header contains various SurfaceFilter implementations that apply
 * transformations to image data, for usage with SurfacePipe.
 */

#ifndef mozilla_image_SurfaceFilters_h
#define mozilla_image_SurfaceFilters_h

#include <algorithm>
#include <stdint.h>
#include <string.h>

#include "mozilla/Likely.h"
#include "mozilla/Maybe.h"
#include "mozilla/UniquePtr.h"
#include "mozilla/gfx/2D.h"

#include "DownscalingFilter.h"
#include "SurfaceCache.h"
#include "SurfacePipe.h"

namespace mozilla {
namespace image {

//////////////////////////////////////////////////////////////////////////////
// DeinterlacingFilter
//////////////////////////////////////////////////////////////////////////////

template <typename PixelType, typename Next> class DeinterlacingFilter;

/**
 * A configuration struct for DeinterlacingFilter.
 *
 * The 'PixelType' template parameter should be either uint32_t (for output to a
 * SurfaceSink) or uint8_t (for output to a PalettedSurfaceSink).
 */
template <typename PixelType>
struct DeinterlacingConfig
{
  template <typename Next> using Filter = DeinterlacingFilter<PixelType, Next>;
  bool mProgressiveDisplay; /// If true, duplicate rows during deinterlacing
                            /// to make progressive display look better, at
                            /// the cost of some performance.
};

/**
 * DeinterlacingFilter performs deinterlacing by reordering the rows that are
 * written to it.
 *
 * The 'PixelType' template parameter should be either uint32_t (for output to a
 * SurfaceSink) or uint8_t (for output to a PalettedSurfaceSink).
 *
 * The 'Next' template parameter specifies the next filter in the chain.
 */
template <typename PixelType, typename Next>
class DeinterlacingFilter final : public SurfaceFilter
{
public:
  DeinterlacingFilter()
    : mInputRow(0)
    , mOutputRow(0)
    , mPass(0)
    , mProgressiveDisplay(true)
  { }

  template <typename... Rest>
  nsresult Configure(const DeinterlacingConfig<PixelType>& aConfig, Rest... aRest)
  {
    nsresult rv = mNext.Configure(aRest...);
    if (NS_FAILED(rv)) {
      return rv;
    }

    if (sizeof(PixelType) == 1 && !mNext.IsValidPalettedPipe()) {
      NS_WARNING("Paletted DeinterlacingFilter used with non-paletted pipe?");
      return NS_ERROR_INVALID_ARG;
    }
    if (sizeof(PixelType) == 4 && mNext.IsValidPalettedPipe()) {
      NS_WARNING("Non-paletted DeinterlacingFilter used with paletted pipe?");
      return NS_ERROR_INVALID_ARG;
    }

    gfx::IntSize outputSize = mNext.InputSize();
    mProgressiveDisplay = aConfig.mProgressiveDisplay;

    const uint32_t bufferSize = outputSize.width *
                                outputSize.height *
                                sizeof(PixelType);

    // Use the size of the SurfaceCache as a heuristic to avoid gigantic
    // allocations. Even if DownscalingFilter allowed us to allocate space for
    // the output image, the deinterlacing buffer may still be too big, and
    // fallible allocation won't always save us in the presence of overcommit.
    if (!SurfaceCache::CanHold(bufferSize)) {
      return NS_ERROR_OUT_OF_MEMORY;
    }

    // Allocate the buffer, which contains deinterlaced scanlines of the image.
    // The buffer is necessary so that we can output rows which have already
    // been deinterlaced again on subsequent passes. Since a later stage in the
    // pipeline may be transforming the rows it receives (for example, by
    // downscaling them), the rows may no longer exist in their original form on
    // the surface itself.
    mBuffer.reset(new (fallible) uint8_t[bufferSize]);
    if (MOZ_UNLIKELY(!mBuffer)) {
      return NS_ERROR_OUT_OF_MEMORY;
    }

    // Clear the buffer to avoid writing uninitialized memory to the output.
    memset(mBuffer.get(), 0, bufferSize);

    ConfigureFilter(outputSize, sizeof(PixelType));
    return NS_OK;
  }

  bool IsValidPalettedPipe() const override
  {
    return sizeof(PixelType) == 1 && mNext.IsValidPalettedPipe();
  }

  Maybe<SurfaceInvalidRect> TakeInvalidRect() override
  {
    return mNext.TakeInvalidRect();
  }

protected:
  uint8_t* DoResetToFirstRow() override
  {
    mNext.ResetToFirstRow();
    mPass = 0;
    mInputRow = 0;
    mOutputRow = InterlaceOffset(mPass);
    return GetRowPointer(mOutputRow);
  }

  uint8_t* DoAdvanceRow() override
  {
    if (mPass >= 4) {
      return nullptr;  // We already finished all passes.
    }
    if (mInputRow >= InputSize().height) {
      return nullptr;  // We already got all the input rows we expect.
    }

    // Duplicate from the first Haeberli row to the remaining Haeberli rows
    // within the buffer.
    DuplicateRows(HaeberliOutputStartRow(mPass, mProgressiveDisplay, mOutputRow),
                  HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                         InputSize(), mOutputRow));

    // Write the current set of Haeberli rows (which contains the current row)
    // to the next stage in the pipeline.
    OutputRows(HaeberliOutputStartRow(mPass, mProgressiveDisplay, mOutputRow),
               HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                      InputSize(), mOutputRow));

    // Determine which output row the next input row corresponds to.
    bool advancedPass = false;
    uint32_t stride = InterlaceStride(mPass);
    int32_t nextOutputRow = mOutputRow + stride;
    while (nextOutputRow >= InputSize().height) {
      // Copy any remaining rows from the buffer.
      if (!advancedPass) {
        OutputRows(HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                          InputSize(), mOutputRow),
                   InputSize().height);
      }

      // We finished the current pass; advance to the next one.
      mPass++;
      if (mPass >= 4) {
        return nullptr;  // Finished all passes.
      }

      // Tell the next pipeline stage that we're starting the next pass.
      mNext.ResetToFirstRow();

      // Update our state to reflect the pass change.
      advancedPass = true;
      stride = InterlaceStride(mPass);
      nextOutputRow = InterlaceOffset(mPass);
    }

    MOZ_ASSERT(nextOutputRow >= 0);
    MOZ_ASSERT(nextOutputRow < InputSize().height);

    MOZ_ASSERT(HaeberliOutputStartRow(mPass, mProgressiveDisplay,
                                      nextOutputRow) >= 0);
    MOZ_ASSERT(HaeberliOutputStartRow(mPass, mProgressiveDisplay,
                                      nextOutputRow) < InputSize().height);
    MOZ_ASSERT(HaeberliOutputStartRow(mPass, mProgressiveDisplay,
                                      nextOutputRow) <= nextOutputRow);

    MOZ_ASSERT(HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                      InputSize(), nextOutputRow) >= 0);
    MOZ_ASSERT(HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                      InputSize(), nextOutputRow)
                 <= InputSize().height);
    MOZ_ASSERT(HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                      InputSize(), nextOutputRow)
                 > nextOutputRow);

    int32_t nextHaeberliOutputRow =
      HaeberliOutputStartRow(mPass, mProgressiveDisplay, nextOutputRow);

    // Copy rows from the buffer until we reach the desired output row.
    if (advancedPass) {
      OutputRows(0, nextHaeberliOutputRow);
    } else {
      OutputRows(HaeberliOutputUntilRow(mPass, mProgressiveDisplay,
                                        InputSize(), mOutputRow),
                 nextHaeberliOutputRow);
    }

    // Update our position within the buffer.
    mInputRow++;
    mOutputRow = nextOutputRow;

    // We'll actually write to the first Haeberli output row, then copy it until
    // we reach the last Haeberli output row. The assertions above make sure
    // this always includes mOutputRow.
    return GetRowPointer(nextHaeberliOutputRow);
  }

private:
  static uint32_t InterlaceOffset(uint32_t aPass)
  {
    MOZ_ASSERT(aPass < 4, "Invalid pass");
    static const uint8_t offset[] = { 0, 4, 2, 1 };
    return offset[aPass];
  }

  static uint32_t InterlaceStride(uint32_t aPass)
  {
    MOZ_ASSERT(aPass < 4, "Invalid pass");
    static const uint8_t stride[] = { 8, 8, 4, 2 };
    return stride[aPass];
  }

  static int32_t HaeberliOutputStartRow(uint32_t aPass,
                                        bool aProgressiveDisplay,
                                        int32_t aOutputRow)
  {
    MOZ_ASSERT(aPass < 4, "Invalid pass");
    static const uint8_t firstRowOffset[] = { 3, 1, 0, 0 };

    if (aProgressiveDisplay) {
      return std::max(aOutputRow - firstRowOffset[aPass], 0);
    } else {
      return aOutputRow;
    }
  }

  static int32_t HaeberliOutputUntilRow(uint32_t aPass,
                                        bool aProgressiveDisplay,
                                        const gfx::IntSize& aInputSize,
                                        int32_t aOutputRow)
  {
    MOZ_ASSERT(aPass < 4, "Invalid pass");
    static const uint8_t lastRowOffset[] = { 4, 2, 1, 0 };

    if (aProgressiveDisplay) {
      return std::min(aOutputRow + lastRowOffset[aPass],
                      aInputSize.height - 1)
             + 1;  // Add one because this is an open interval on the right.
    } else {
      return aOutputRow + 1;
    }
  }

  void DuplicateRows(int32_t aStart, int32_t aUntil)
  {
    MOZ_ASSERT(aStart >= 0);
    MOZ_ASSERT(aUntil >= 0);

    if (aUntil <= aStart || aStart >= InputSize().height) {
      return;
    }

    // The source row is the first row in the range.
    const uint8_t* sourceRowPointer = GetRowPointer(aStart);

    // We duplicate the source row into each subsequent row in the range.
    for (int32_t destRow = aStart + 1 ; destRow < aUntil ; ++destRow) {
      uint8_t* destRowPointer = GetRowPointer(destRow);
      memcpy(destRowPointer, sourceRowPointer, InputSize().width * sizeof(PixelType));
    }
  }

  void OutputRows(int32_t aStart, int32_t aUntil)
  {
    MOZ_ASSERT(aStart >= 0);
    MOZ_ASSERT(aUntil >= 0);

    if (aUntil <= aStart || aStart >= InputSize().height) {
      return;
    }

    for (int32_t rowToOutput = aStart; rowToOutput < aUntil; ++rowToOutput) {
      mNext.WriteBuffer(reinterpret_cast<PixelType*>(GetRowPointer(rowToOutput)));
    }
  }

  uint8_t* GetRowPointer(uint32_t aRow) const
  {
    uint32_t offset = aRow * InputSize().width * sizeof(PixelType);
    MOZ_ASSERT(offset < InputSize().width * InputSize().height * sizeof(PixelType),
               "Start of row is outside of image");
    MOZ_ASSERT(offset + InputSize().width * sizeof(PixelType)
                 <= InputSize().width * InputSize().height * sizeof(PixelType),
               "End of row is outside of image");
    return mBuffer.get() + offset;
  }

  Next mNext;                    /// The next SurfaceFilter in the chain.

  UniquePtr<uint8_t[]> mBuffer;  /// The buffer used to store reordered rows.
  int32_t mInputRow;             /// The current row we're reading. (0-indexed)
  int32_t mOutputRow;            /// The current row we're writing. (0-indexed)
  uint8_t mPass;                 /// Which pass we're on. (0-indexed)
  bool mProgressiveDisplay;      /// If true, duplicate rows to optimize for
                                 /// progressive display.
};


//////////////////////////////////////////////////////////////////////////////
// RemoveFrameRectFilter
//////////////////////////////////////////////////////////////////////////////

template <typename Next> class RemoveFrameRectFilter;

/**
 * A configuration struct for RemoveFrameRectFilter.
 */
struct RemoveFrameRectConfig
{
  template <typename Next> using Filter = RemoveFrameRectFilter<Next>;
  gfx::IntRect mFrameRect;  /// The surface subrect which contains data.
};

/**
 * RemoveFrameRectFilter turns an image with a frame rect that does not match
 * its logical size into an image with no frame rect. It does this by writing
 * transparent pixels into any padding regions and throwing away excess data.
 *
 * The 'Next' template parameter specifies the next filter in the chain.
 */
template <typename Next>
class RemoveFrameRectFilter final : public SurfaceFilter
{
public:
  RemoveFrameRectFilter()
    : mRow(0)
  { }

  template <typename... Rest>
  nsresult Configure(const RemoveFrameRectConfig& aConfig, Rest... aRest)
  {
    nsresult rv = mNext.Configure(aRest...);
    if (NS_FAILED(rv)) {
      return rv;
    }

    if (mNext.IsValidPalettedPipe()) {
      NS_WARNING("RemoveFrameRectFilter used with paletted pipe?");
      return NS_ERROR_INVALID_ARG;
    }

    mFrameRect = mUnclampedFrameRect = aConfig.mFrameRect;
    gfx::IntSize outputSize = mNext.InputSize();

    // Forbid frame rects with negative size.
    if (aConfig.mFrameRect.width < 0 || aConfig.mFrameRect.height < 0) {
      return NS_ERROR_INVALID_ARG;
    }

    // Clamp mFrameRect to the output size.
    gfx::IntRect outputRect(0, 0, outputSize.width, outputSize.height);
    mFrameRect = mFrameRect.Intersect(outputRect);

    // If there's no intersection, |mFrameRect| will be an empty rect positioned
    // at the maximum of |inputRect|'s and |aFrameRect|'s coordinates, which is
    // not what we want. Force it to (0, 0) in that case.
    if (mFrameRect.IsEmpty()) {
      mFrameRect.MoveTo(0, 0);
    }

    // We don't need an intermediate buffer unless the unclamped frame rect
    // width is larger than the clamped frame rect width. In that case, the
    // caller will end up writing data that won't end up in the final image at
    // all, and we'll need a buffer to give that data a place to go.
    if (mFrameRect.width < mUnclampedFrameRect.width) {
      mBuffer.reset(new (fallible) uint8_t[mUnclampedFrameRect.width *
                                           sizeof(uint32_t)]);
      if (MOZ_UNLIKELY(!mBuffer)) {
        return NS_ERROR_OUT_OF_MEMORY;
      }

      memset(mBuffer.get(), 0, mUnclampedFrameRect.width * sizeof(uint32_t));
    }

    ConfigureFilter(mUnclampedFrameRect.Size(), sizeof(uint32_t));
    return NS_OK;
  }

  Maybe<SurfaceInvalidRect> TakeInvalidRect() override
  {
    return mNext.TakeInvalidRect();
  }

protected:
  uint8_t* DoResetToFirstRow() override
  {
    uint8_t* rowPtr = mNext.ResetToFirstRow();
    if (rowPtr == nullptr) {
      mRow = mFrameRect.YMost();
      return nullptr;
    }

    mRow = mUnclampedFrameRect.y;

    // Advance the next pipeline stage to the beginning of the frame rect,
    // outputting blank rows.
    if (mFrameRect.y > 0) {
      for (int32_t rowToOutput = 0; rowToOutput < mFrameRect.y ; ++rowToOutput) {
        mNext.WriteEmptyRow();
      }
    }

    // We're at the beginning of the frame rect now, so return if we're either
    // ready for input or we're already done.
    rowPtr = mBuffer ? mBuffer.get() : mNext.CurrentRowPointer();
    if (!mFrameRect.IsEmpty() || rowPtr == nullptr) {
      // Note that the pointer we're returning is for the next row we're
      // actually going to write to, but we may discard writes before that point
      // if mRow < mFrameRect.y.
      return AdjustRowPointer(rowPtr);
    }

    // We've finished the region specified by the frame rect, but the frame rect
    // is empty, so we need to output the rest of the image immediately. Advance
    // to the end of the next pipeline stage's buffer, outputting blank rows.
    while (mNext.WriteEmptyRow() == WriteState::NEED_MORE_DATA) { }

    mRow = mFrameRect.YMost();
    return nullptr;  // We're done.
  }

  uint8_t* DoAdvanceRow() override
  {
    uint8_t* rowPtr = nullptr;

    const int32_t currentRow = mRow;
    mRow++;

    if (currentRow < mFrameRect.y) {
      // This row is outside of the frame rect, so just drop it on the floor.
      rowPtr = mBuffer ? mBuffer.get() : mNext.CurrentRowPointer();
      return AdjustRowPointer(rowPtr);
    } else if (currentRow >= mFrameRect.YMost()) {
      NS_WARNING("RemoveFrameRectFilter: Advancing past end of frame rect");
      return nullptr;
    }

    // If we had to buffer, copy the data. Otherwise, just advance the row.
    if (mBuffer) {
      // We write from the beginning of the buffer unless |mUnclampedFrameRect.x|
      // is negative; if that's the case, we have to skip the portion of the
      // unclamped frame rect that's outside the row.
      uint32_t* source = reinterpret_cast<uint32_t*>(mBuffer.get()) -
                         std::min(mUnclampedFrameRect.x, 0);

      // We write |mFrameRect.width| columns starting at |mFrameRect.x|; we've
      // already clamped these values to the size of the output, so we don't
      // have to worry about bounds checking here (though WriteBuffer() will do
      // it for us in any case).
      WriteState state = mNext.WriteBuffer(source, mFrameRect.x, mFrameRect.width);

      rowPtr = state == WriteState::NEED_MORE_DATA ? mBuffer.get()
                                                   : nullptr;
    } else {
      rowPtr = mNext.AdvanceRow();
    }

    // If there's still more data coming or we're already done, just adjust the
    // pointer and return.
    if (mRow < mFrameRect.YMost() || rowPtr == nullptr) {
      return AdjustRowPointer(rowPtr);
    }

    // We've finished the region specified by the frame rect. Advance to the end
    // of the next pipeline stage's buffer, outputting blank rows.
    while (mNext.WriteEmptyRow() == WriteState::NEED_MORE_DATA) { }

    mRow = mFrameRect.YMost();
    return nullptr;  // We're done.
  }

private:
  uint8_t* AdjustRowPointer(uint8_t* aNextRowPointer) const
  {
    if (mBuffer) {
      MOZ_ASSERT(aNextRowPointer == mBuffer.get() || aNextRowPointer == nullptr);
      return aNextRowPointer;  // No adjustment needed for an intermediate buffer.
    }

    if (mFrameRect.IsEmpty() ||
        mRow >= mFrameRect.YMost() ||
        aNextRowPointer == nullptr) {
      return nullptr;  // Nothing left to write.
    }

    return aNextRowPointer + mFrameRect.x * sizeof(uint32_t);
  }

  Next mNext;                        /// The next SurfaceFilter in the chain.

  gfx::IntRect mFrameRect;           /// The surface subrect which contains data,
                                     /// clamped to the image size.
  gfx::IntRect mUnclampedFrameRect;  /// The frame rect before clamping.
  UniquePtr<uint8_t[]> mBuffer;      /// The intermediate buffer, if one is
                                     /// necessary because the frame rect width
                                     /// is larger than the image's logical width.
  int32_t  mRow;                     /// The row in unclamped frame rect space
                                     /// that we're currently writing.
};


//////////////////////////////////////////////////////////////////////////////
// ADAM7InterpolatingFilter
//////////////////////////////////////////////////////////////////////////////

template <typename Next> class ADAM7InterpolatingFilter;

/**
 * A configuration struct for ADAM7InterpolatingFilter.
 */
struct ADAM7InterpolatingConfig
{
  template <typename Next> using Filter = ADAM7InterpolatingFilter<Next>;
};

/**
 * ADAM7InterpolatingFilter performs bilinear interpolation over an ADAM7
 * interlaced image.
 *
 * ADAM7 breaks up the image into 8x8 blocks. On each of the 7 passes, a new set
 * of pixels in each block receives their final values, according to the
 * following pattern:
 *
 *    1 6 4 6 2 6 4 6
 *    7 7 7 7 7 7 7 7
 *    5 6 5 6 5 6 5 6
 *    7 7 7 7 7 7 7 7
 *    3 6 4 6 3 6 4 6
 *    7 7 7 7 7 7 7 7
 *    5 6 5 6 5 6 5 6
 *    7 7 7 7 7 7 7 7
 *
 * When rendering the pixels that have not yet received their final values, we
 * can get much better intermediate results if we interpolate between
 * the pixels we *have* gotten so far. This filter performs bilinear
 * interpolation by first performing linear interpolation horizontally for each
 * "important" row (which we'll define as a row that has received any pixels
 * with final values at all) and then performing linear interpolation vertically
 * to produce pixel values for rows which aren't important on the current pass.
 *
 * Note that this filter totally ignores the data which is written to rows which
 * aren't important on the current pass! It's fine to write nothing at all for
 * these rows, although doing so won't cause any harm.
 *
 * XXX(seth): In bug 1280552 we'll add a SIMD implementation for this filter.
 *
 * The 'Next' template parameter specifies the next filter in the chain.
 */
template <typename Next>
class ADAM7InterpolatingFilter final : public SurfaceFilter
{
public:
  ADAM7InterpolatingFilter()
    : mPass(0)  // The current pass, in the range 1..7. Starts at 0 so that
                // DoResetToFirstRow() doesn't have to special case the first pass.
    , mRow(0)
  { }

  template <typename... Rest>
  nsresult Configure(const ADAM7InterpolatingConfig& aConfig, Rest... aRest)
  {
    nsresult rv = mNext.Configure(aRest...);
    if (NS_FAILED(rv)) {
      return rv;
    }

    if (mNext.IsValidPalettedPipe()) {
      NS_WARNING("ADAM7InterpolatingFilter used with paletted pipe?");
      return NS_ERROR_INVALID_ARG;
    }

    // We have two intermediate buffers, one for the previous row with final
    // pixel values and one for the row that the previous filter in the chain is
    // currently writing to.
    size_t inputWidthInBytes = mNext.InputSize().width * sizeof(uint32_t);
    mPreviousRow.reset(new (fallible) uint8_t[inputWidthInBytes]);
    if (MOZ_UNLIKELY(!mPreviousRow)) {
      return NS_ERROR_OUT_OF_MEMORY;
    }

    mCurrentRow.reset(new (fallible) uint8_t[inputWidthInBytes]);
    if (MOZ_UNLIKELY(!mCurrentRow)) {
      return NS_ERROR_OUT_OF_MEMORY;
    }

    memset(mPreviousRow.get(), 0, inputWidthInBytes);
    memset(mCurrentRow.get(), 0, inputWidthInBytes);

    ConfigureFilter(mNext.InputSize(), sizeof(uint32_t));
    return NS_OK;
  }

  Maybe<SurfaceInvalidRect> TakeInvalidRect() override
  {
    return mNext.TakeInvalidRect();
  }

protected:
  uint8_t* DoResetToFirstRow() override
  {
    mRow = 0;
    mPass = std::min(mPass + 1, 7);

    uint8_t* rowPtr = mNext.ResetToFirstRow();
    if (mPass == 7) {
      // Short circuit this filter on the final pass, since all pixels have
      // their final values at that point.
      return rowPtr;
    }

    return mCurrentRow.get();
  }

  uint8_t* DoAdvanceRow() override
  {
    MOZ_ASSERT(0 < mPass && mPass <= 7, "Invalid pass");

    int32_t currentRow = mRow;
    ++mRow;

    if (mPass == 7) {
      // On the final pass we short circuit this filter totally.
      return mNext.AdvanceRow();
    }

    const int32_t lastImportantRow = LastImportantRow(InputSize().height, mPass);
    if (currentRow > lastImportantRow) {
      return nullptr;  // This pass is already complete.
    }

    if (!IsImportantRow(currentRow, mPass)) {
      // We just ignore whatever the caller gives us for these rows. We'll
      // interpolate them in later.
      return mCurrentRow.get();
    }

    // This is an important row. We need to perform horizontal interpolation for
    // these rows.
    InterpolateHorizontally(mCurrentRow.get(), InputSize().width, mPass);

    // Interpolate vertically between the previous important row and the current
    // important row. We skip this if the current row is 0 (which is always an
    // important row), because in that case there is no previous important row
    // to interpolate with.
    if (currentRow != 0) {
      InterpolateVertically(mPreviousRow.get(), mCurrentRow.get(), mPass, mNext);
    }

    // Write out the current row itself, which, being an important row, does not
    // need vertical interpolation.
    uint32_t* currentRowAsPixels = reinterpret_cast<uint32_t*>(mCurrentRow.get());
    mNext.WriteBuffer(currentRowAsPixels);

    if (currentRow == lastImportantRow) {
      // This is the last important row, which completes this pass. Note that
      // for very small images, this may be the first row! Since there won't be
      // another important row, there's nothing to interpolate with vertically,
      // so we just duplicate this row until the end of the image.
      while (mNext.WriteBuffer(currentRowAsPixels) == WriteState::NEED_MORE_DATA) { }

      // All of the remaining rows in the image were determined above, so we're done.
      return nullptr;
    }

    // The current row is now the previous important row; save it.
    Swap(mPreviousRow, mCurrentRow);

    MOZ_ASSERT(mRow < InputSize().height, "Reached the end of the surface without "
                                          "hitting the last important row?");

    return mCurrentRow.get();
  }

private:
  static void InterpolateVertically(uint8_t* aPreviousRow,
                                    uint8_t* aCurrentRow,
                                    uint8_t aPass,
                                    SurfaceFilter& aNext)
  {
    const float* weights = InterpolationWeights(ImportantRowStride(aPass));

    // We need to interpolate vertically to generate the rows between the
    // previous important row and the next one. Recall that important rows are
    // rows which contain at least some final pixels; see
    // InterpolateHorizontally() for some additional explanation as to what that
    // means. Note that we've already written out the previous important row, so
    // we start the iteration at 1.
    for (int32_t outRow = 1; outRow < ImportantRowStride(aPass); ++outRow) {
      const float weight = weights[outRow];

      // We iterate through the previous and current important row every time we
      // write out an interpolated row, so we need to copy the pointers.
      uint8_t* prevRowBytes = aPreviousRow;
      uint8_t* currRowBytes = aCurrentRow;

      // Write out the interpolated pixels. Interpolation is componentwise.
      aNext.template WritePixelsToRow<uint32_t>([&]{
        uint32_t pixel = 0;
        auto* component = reinterpret_cast<uint8_t*>(&pixel);
        *component++ = InterpolateByte(*prevRowBytes++, *currRowBytes++, weight);
        *component++ = InterpolateByte(*prevRowBytes++, *currRowBytes++, weight);
        *component++ = InterpolateByte(*prevRowBytes++, *currRowBytes++, weight);
        *component++ = InterpolateByte(*prevRowBytes++, *currRowBytes++, weight);
        return AsVariant(pixel);
      });
    }
  }

  static void InterpolateHorizontally(uint8_t* aRow, int32_t aWidth, uint8_t aPass)
  {
    // Collect the data we'll need to perform horizontal interpolation. The
    // terminology here bears some explanation: a "final pixel" is a pixel which
    // has received its final value. On each pass, a new set of pixels receives
    // their final value; see the diagram above of the 8x8 pattern that ADAM7
    // uses. Any pixel which hasn't received its final value on this pass
    // derives its value from either horizontal or vertical interpolation
    // instead.
    const size_t finalPixelStride = FinalPixelStride(aPass);
    const size_t finalPixelStrideBytes = finalPixelStride * sizeof(uint32_t);
    const size_t lastFinalPixel = LastFinalPixel(aWidth, aPass);
    const size_t lastFinalPixelBytes = lastFinalPixel * sizeof(uint32_t);
    const float* weights = InterpolationWeights(finalPixelStride);

    // Interpolate blocks of pixels which lie between two final pixels.
    // Horizontal interpolation is done in place, as we'll need the results
    // later when we vertically interpolate.
    for (size_t blockBytes = 0;
         blockBytes < lastFinalPixelBytes;
         blockBytes += finalPixelStrideBytes) {
      uint8_t* finalPixelA = aRow + blockBytes;
      uint8_t* finalPixelB = aRow + blockBytes + finalPixelStrideBytes;

      MOZ_ASSERT(finalPixelA < aRow + aWidth * sizeof(uint32_t),
                 "Running off end of buffer");
      MOZ_ASSERT(finalPixelB < aRow + aWidth * sizeof(uint32_t),
                 "Running off end of buffer");

      // Interpolate the individual pixels componentwise. Note that we start
      // iteration at 1 since we don't need to apply any interpolation to the
      // first pixel in the block, which has its final value.
      for (size_t pixelIndex = 1; pixelIndex < finalPixelStride; ++pixelIndex) {
        const float weight = weights[pixelIndex];
        uint8_t* pixel = aRow + blockBytes + pixelIndex * sizeof(uint32_t);

        MOZ_ASSERT(pixel < aRow + aWidth * sizeof(uint32_t), "Running off end of buffer");

        for (size_t component = 0; component < sizeof(uint32_t); ++component) {
          pixel[component] =
            InterpolateByte(finalPixelA[component], finalPixelB[component], weight);
        }
      }
    }

    // For the pixels after the last final pixel in the row, there isn't a
    // second final pixel to interpolate with, so just duplicate.
    uint32_t* rowPixels = reinterpret_cast<uint32_t*>(aRow);
    uint32_t pixelToDuplicate = rowPixels[lastFinalPixel];
    for (int32_t pixelIndex = lastFinalPixel + 1;
         pixelIndex < aWidth;
         ++pixelIndex) {
      MOZ_ASSERT(pixelIndex < aWidth, "Running off end of buffer");
      rowPixels[pixelIndex] = pixelToDuplicate;
    }
  }

  static uint8_t InterpolateByte(uint8_t aByteA, uint8_t aByteB, float aWeight)
  {
    return uint8_t(aByteA * aWeight + aByteB * (1.0f - aWeight));
  }

  static int32_t ImportantRowStride(uint8_t aPass)
  {
    MOZ_ASSERT(0 < aPass && aPass <= 7, "Invalid pass");

    // The stride between important rows for each pass, with a dummy value for
    // the nonexistent pass 0.
    static int32_t strides[] = { 1, 8, 8, 4, 4, 2, 2, 1 };

    return strides[aPass];
  }

  static bool IsImportantRow(int32_t aRow, uint8_t aPass)
  {
    MOZ_ASSERT(aRow >= 0);

    // Whether the row is important comes down to divisibility by the stride for
    // this pass, which is always a power of 2, so we can check using a mask.
    int32_t mask = ImportantRowStride(aPass) - 1;
    return (aRow & mask) == 0;
  }

  static int32_t LastImportantRow(int32_t aHeight, uint8_t aPass)
  {
    MOZ_ASSERT(aHeight > 0);

    // We can find the last important row using the same mask trick as above.
    int32_t lastRow = aHeight - 1;
    int32_t mask = ImportantRowStride(aPass) - 1;
    return lastRow - (lastRow & mask);
  }

  static size_t FinalPixelStride(uint8_t aPass)
  {
    MOZ_ASSERT(0 < aPass && aPass <= 7, "Invalid pass");

    // The stride between the final pixels in important rows for each pass, with
    // a dummy value for the nonexistent pass 0.
    static size_t strides[] = { 1, 8, 4, 4, 2, 2, 1, 1 };

    return strides[aPass];
  }

  static size_t LastFinalPixel(int32_t aWidth, uint8_t aPass)
  {
    MOZ_ASSERT(aWidth >= 0);

    // Again, we can use the mask trick above to find the last important pixel.
    int32_t lastColumn = aWidth - 1;
    size_t mask = FinalPixelStride(aPass) - 1;
    return lastColumn - (lastColumn & mask);
  }

  static const float* InterpolationWeights(int32_t aStride)
  {
    // Precalculated interpolation weights. These are used to interpolate
    // between final pixels or between important rows. Although no interpolation
    // is actually applied to the previous final pixel or important row value,
    // the arrays still start with 1.0f, which is always skipped, primarily
    // because otherwise |stride1Weights| would have zero elements.
    static float stride8Weights[] =
      { 1.0f, 7 / 8.0f, 6 / 8.0f, 5 / 8.0f, 4 / 8.0f, 3 / 8.0f, 2 / 8.0f, 1 / 8.0f };
    static float stride4Weights[] = { 1.0f, 3 / 4.0f, 2 / 4.0f, 1 / 4.0f };
    static float stride2Weights[] = { 1.0f, 1 / 2.0f };
    static float stride1Weights[] = { 1.0f };

    switch (aStride) {
      case 8:  return stride8Weights;
      case 4:  return stride4Weights;
      case 2:  return stride2Weights;
      case 1:  return stride1Weights;
      default: MOZ_CRASH();
    }
  }

  Next mNext;                         /// The next SurfaceFilter in the chain.

  UniquePtr<uint8_t[]> mPreviousRow;  /// The last important row (i.e., row with
                                      /// final pixel values) that got written to.
  UniquePtr<uint8_t[]> mCurrentRow;   /// The row that's being written to right
                                      /// now.
  uint8_t mPass;                      /// Which ADAM7 pass we're on. Valid passes
                                      /// are 1..7 during processing and 0 prior
                                      /// to configuraiton.
  int32_t mRow;                       /// The row we're currently reading.
};

} // namespace image
} // namespace mozilla

#endif // mozilla_image_SurfaceFilters_h