image/SurfaceFilters.h
author Ryan VanderMeulen <ryanvm@gmail.com>
Wed, 17 Jan 2018 15:55:38 -0500
changeset 399709 ef04f3ad847540c876c130b7d140a960d922a024
parent 397223 570db0492055ac9bd86af8c13cbdb16148578ff9
child 420213 87a0a6e103d8151d0e40929580545e3fbcc6b490
permissions -rw-r--r--
Backed out 20 changesets (bug 1411654) for incorrect android:debuggable. r=nalexander, a=RyanVM Backed out changeset cfad693be918 (bug 1411654) Backed out changeset 55776829a744 (bug 1411654) Backed out changeset c5bf85d56fed (bug 1411654) Backed out changeset c270f97bb0da (bug 1411654) Backed out changeset fde9bf9c14c3 (bug 1411654) Backed out changeset 01836fd98c63 (bug 1411654) Backed out changeset 730a70767743 (bug 1411654) Backed out changeset 690e265c684c (bug 1411654) Backed out changeset f918500d9cf5 (bug 1411654) Backed out changeset cec2b8828cc8 (bug 1411654) Backed out changeset 76085ddd5ac7 (bug 1411654) Backed out changeset 2b37201606f5 (bug 1411654) Backed out changeset d0d513d1c379 (bug 1411654) Backed out changeset e7b0cc801cf1 (bug 1411654) Backed out changeset 901b304603d9 (bug 1411654) Backed out changeset 373c9a71d945 (bug 1411654) Backed out changeset 3dc3beab95f8 (bug 1411654) Backed out changeset 22a861db1573 (bug 1411654) Backed out changeset 0850b319efd4 (bug 1411654) Backed out changeset d276d3deba05 (bug 1411654)

/* -*- 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