Modified: trunk/Source/WebCore/platform/graphics/ColorUtilities.h (225171 => 225172)
--- trunk/Source/WebCore/platform/graphics/ColorUtilities.h 2017-11-27 16:30:58 UTC (rev 225171)
+++ trunk/Source/WebCore/platform/graphics/ColorUtilities.h 2017-11-27 16:37:13 UTC (rev 225172)
@@ -93,7 +93,12 @@
struct ColorComponents {
ColorComponents(const FloatComponents&);
-
+
+ static ColorComponents fromRGBA(unsigned pixel)
+ {
+ return ColorComponents((pixel >> 24) & 0xFF, (pixel >> 16) & 0xFF, (pixel >> 8) & 0xFF, pixel & 0xFF);
+ }
+
ColorComponents(uint8_t a = 0, uint8_t b = 0, uint8_t c = 0, uint8_t d = 0)
{
components[0] = a;
@@ -101,9 +106,35 @@
components[2] = c;
components[3] = d;
}
+
+ unsigned toRGBA() const
+ {
+ return components[0] << 24 | components[1] << 16 | components[2] << 8 | components[3];
+ }
+
uint8_t components[4] { };
};
+inline ColorComponents perComponentMax(const ColorComponents& a, const ColorComponents& b)
+{
+ return {
+ std::max(a.components[0], b.components[0]),
+ std::max(a.components[1], b.components[1]),
+ std::max(a.components[2], b.components[2]),
+ std::max(a.components[3], b.components[3])
+ };
+}
+
+inline ColorComponents perComponentMin(const ColorComponents& a, const ColorComponents& b)
+{
+ return {
+ std::min(a.components[0], b.components[0]),
+ std::min(a.components[1], b.components[1]),
+ std::min(a.components[2], b.components[2]),
+ std::min(a.components[3], b.components[3])
+ };
+}
+
inline uint8_t clampedColorComponent(float f)
{
// See also colorFloatToRGBAByte().
Modified: trunk/Source/WebCore/platform/graphics/filters/FEColorMatrix.cpp (225171 => 225172)
--- trunk/Source/WebCore/platform/graphics/filters/FEColorMatrix.cpp 2017-11-27 16:30:58 UTC (rev 225171)
+++ trunk/Source/WebCore/platform/graphics/filters/FEColorMatrix.cpp 2017-11-27 16:37:13 UTC (rev 225172)
@@ -25,17 +25,14 @@
#include "Filter.h"
#include "GraphicsContext.h"
-#include <wtf/text/TextStream.h>
-
#include <runtime/Uint8ClampedArray.h>
#include <wtf/MathExtras.h>
+#include <wtf/text/TextStream.h>
#if USE(ACCELERATE)
#include <Accelerate/Accelerate.h>
#endif
-#define PRINT_FILTER_PERFORMANCE 0
-
namespace WebCore {
FEColorMatrix::FEColorMatrix(Filter& filter, ColorMatrixType type, const Vector<float>& values)
@@ -283,10 +280,6 @@
if (!resultImage)
return;
-#if PRINT_FILTER_PERFORMANCE
- MonotonicTime startTime = MonotonicTime::now();
-#endif
-
ImageBuffer* inBuffer = in->imageBufferResult();
if (inBuffer)
resultImage->context().drawImageBuffer(*inBuffer, drawingRegionOfInputImage(in->absolutePaintRect()));
@@ -318,20 +311,6 @@
}
resultImage->putByteArray(Unmultiplied, *pixelArray, imageRect.size(), imageRect, IntPoint());
-
-#if PRINT_FILTER_PERFORMANCE
- MonotonicTime endTime = MonotonicTime::now();
- Seconds duration = endTime - startTime;
- unsigned pixelCount = imageRect.width() * imageRect.height();
-
- static double totalMegapixels;
- totalMegapixels += (double)pixelCount / 1048576.0;
-
- static Seconds totalTime;
- totalTime += duration;
-
- WTFLogAlways("FEColorMatrix::platformApplySoftware (%dx%d) took %.4fms (ave %.4fms/mp)", imageRect.width(), imageRect.height(), duration.milliseconds(), totalTime.milliseconds() / totalMegapixels);
-#endif
}
static TextStream& operator<<(TextStream& ts, const ColorMatrixType& type)
Modified: trunk/Source/WebCore/platform/graphics/filters/FEMorphology.cpp (225171 => 225172)
--- trunk/Source/WebCore/platform/graphics/filters/FEMorphology.cpp 2017-11-27 16:30:58 UTC (rev 225171)
+++ trunk/Source/WebCore/platform/graphics/filters/FEMorphology.cpp 2017-11-27 16:37:13 UTC (rev 225172)
@@ -4,6 +4,7 @@
* Copyright (C) 2005 Eric Seidel <[email protected]>
* Copyright (C) 2009 Dirk Schulze <[email protected]>
* Copyright (C) Research In Motion Limited 2010. All rights reserved.
+ * Copyright (C) Apple Inc. 2017. All rights reserved.
*
* This library is free software; you can redistribute it and/or
* modify it under the terms of the GNU Library General Public
@@ -24,12 +25,12 @@
#include "config.h"
#include "FEMorphology.h"
+#include "ColorUtilities.h"
#include "Filter.h"
-#include <wtf/text/TextStream.h>
-
#include <runtime/Uint8ClampedArray.h>
#include <wtf/ParallelJobs.h>
#include <wtf/Vector.h>
+#include <wtf/text/TextStream.h>
namespace WebCore {
@@ -83,78 +84,96 @@
setAbsolutePaintRect(enclosingIntRect(paintRect));
}
-static inline bool shouldSupersedeExtremum(unsigned char newValue, unsigned char currentValue, MorphologyOperatorType type)
+static inline int pixelArrayIndex(int x, int y, int width)
{
- return (type == FEMORPHOLOGY_OPERATOR_ERODE && newValue < currentValue)
- || (type == FEMORPHOLOGY_OPERATOR_DILATE && newValue > currentValue);
+ return (y * width + x) * 4;
}
-static inline int pixelArrayIndex(int x, int y, int width, int colorChannel)
+template<MorphologyOperatorType type>
+ALWAYS_INLINE ColorComponents minOrMax(const ColorComponents& a, const ColorComponents& b)
{
- return (y * width + x) * 4 + colorChannel;
+ if (type == FEMORPHOLOGY_OPERATOR_ERODE)
+ return perComponentMin(a, b);
+
+ return perComponentMax(a, b);
}
-static inline unsigned char columnExtremum(const Uint8ClampedArray& srcPixelArray, int x, int yStart, int yEnd, int width, unsigned colorChannel, MorphologyOperatorType type)
+template<MorphologyOperatorType type>
+ALWAYS_INLINE ColorComponents columnExtremum(const Uint8ClampedArray& srcPixelArray, int x, int yStart, int yEnd, int width)
{
- unsigned char extremum = srcPixelArray.item(pixelArrayIndex(x, yStart, width, colorChannel));
+ auto extremum = ColorComponents::fromRGBA(*reinterpret_cast<const unsigned*>(srcPixelArray.data() + pixelArrayIndex(x, yStart, width)));
+
for (int y = yStart + 1; y < yEnd; ++y) {
- unsigned char pixel = srcPixelArray.item(pixelArrayIndex(x, y, width, colorChannel));
- if (shouldSupersedeExtremum(pixel, extremum, type))
- extremum = pixel;
+ auto pixel = ColorComponents::fromRGBA(*reinterpret_cast<const unsigned*>(srcPixelArray.data() + pixelArrayIndex(x, y, width)));
+ extremum = minOrMax<type>(extremum, pixel);
}
return extremum;
}
-static inline unsigned char kernelExtremum(const Vector<unsigned char>& kernel, MorphologyOperatorType type)
+ALWAYS_INLINE ColorComponents columnExtremum(const Uint8ClampedArray& srcPixelArray, int x, int yStart, int yEnd, int width, MorphologyOperatorType type)
{
- Vector<unsigned char>::const_iterator iter = kernel.begin();
- unsigned char extremum = *iter;
- for (Vector<unsigned char>::const_iterator end = kernel.end(); ++iter != end; ) {
- if (shouldSupersedeExtremum(*iter, extremum, type))
- extremum = *iter;
- }
+ if (type == FEMORPHOLOGY_OPERATOR_ERODE)
+ return columnExtremum<FEMORPHOLOGY_OPERATOR_ERODE>(srcPixelArray, x, yStart, yEnd, width);
+
+ return columnExtremum<FEMORPHOLOGY_OPERATOR_DILATE>(srcPixelArray, x, yStart, yEnd, width);
+}
+
+using ColumnExtrema = Vector<ColorComponents, 16>;
+
+template<MorphologyOperatorType type>
+ALWAYS_INLINE ColorComponents kernelExtremum(const ColumnExtrema& kernel)
+{
+ auto extremum = kernel[0];
+ for (size_t i = 1; i < kernel.size(); ++i)
+ extremum = minOrMax<type>(extremum, kernel[i]);
+
return extremum;
}
-void FEMorphology::platformApplyGeneric(PaintingData* paintingData, int yStart, int yEnd)
+ALWAYS_INLINE ColorComponents kernelExtremum(const ColumnExtrema& kernel, MorphologyOperatorType type)
{
- const Uint8ClampedArray& srcPixelArray = *paintingData->srcPixelArray;
- Uint8ClampedArray& dstPixelArray = *paintingData->dstPixelArray;
- const int radiusX = paintingData->radiusX;
- const int radiusY = paintingData->radiusY;
- const int width = paintingData->width;
- const int height = paintingData->height;
+ if (type == FEMORPHOLOGY_OPERATOR_ERODE)
+ return kernelExtremum<FEMORPHOLOGY_OPERATOR_ERODE>(kernel);
+ return kernelExtremum<FEMORPHOLOGY_OPERATOR_DILATE>(kernel);
+}
+
+void FEMorphology::platformApplyGeneric(const PaintingData& paintingData, int startY, int endY)
+{
+ const auto& srcPixelArray = *paintingData.srcPixelArray;
+ auto& dstPixelArray = *paintingData.dstPixelArray;
+
+ const int radiusX = paintingData.radiusX;
+ const int radiusY = paintingData.radiusY;
+ const int width = paintingData.width;
+ const int height = paintingData.height;
+
ASSERT(radiusX <= width || radiusY <= height);
- ASSERT(yStart >= 0 && yEnd <= height && yStart < yEnd);
+ ASSERT(startY >= 0 && endY <= height && startY < endY);
- Vector<unsigned char> extrema;
- for (int y = yStart; y < yEnd; ++y) {
- int yStartExtrema = std::max(0, y - radiusY);
- int yEndExtrema = std::min(height - 1, y + radiusY);
+ ColumnExtrema extrema;
+ extrema.reserveInitialCapacity(2 * radiusX + 1);
- for (unsigned colorChannel = 0; colorChannel < 4; ++colorChannel) {
- extrema.clear();
- // Compute extremas for each columns
- for (int x = 0; x < radiusX; ++x)
- extrema.append(columnExtremum(srcPixelArray, x, yStartExtrema, yEndExtrema, width, colorChannel, m_type));
+ for (int y = startY; y < endY; ++y) {
+ int yRadiusStart = std::max(0, y - radiusY);
+ int yRadiusEnd = std::min(height, y + radiusY + 1);
- // Kernel is filled, get extrema of next column
- for (int x = 0; x < width; ++x) {
- if (x < width - radiusX) {
- int xEnd = std::min(x + radiusX, width - 1);
- extrema.append(columnExtremum(srcPixelArray, xEnd, yStartExtrema, yEndExtrema + 1, width, colorChannel, m_type));
- }
+ extrema.shrink(0);
- if (x > radiusX)
- extrema.remove(0);
+ // We start at the left edge, so compute extreme for the radiusX columns.
+ for (int x = 0; x < radiusX; ++x)
+ extrema.append(columnExtremum(srcPixelArray, x, yRadiusStart, yRadiusEnd, width, m_type));
- // The extrema original size = radiusX.
- // Number of new addition = width - radiusX.
- // Number of removals = width - radiusX - 1.
- ASSERT(extrema.size() >= static_cast<size_t>(radiusX + 1));
- dstPixelArray.set(pixelArrayIndex(x, y, width, colorChannel), kernelExtremum(extrema, m_type));
- }
+ // Kernel is filled, get extrema of next column
+ for (int x = 0; x < width; ++x) {
+ if (x < width - radiusX)
+ extrema.append(columnExtremum(srcPixelArray, x + radiusX, yRadiusStart, yRadiusEnd, width, m_type));
+
+ if (x > radiusX)
+ extrema.remove(0);
+
+ unsigned* destPixel = reinterpret_cast<unsigned*>(dstPixelArray.data() + pixelArrayIndex(x, y, width));
+ *destPixel = kernelExtremum(extrema, m_type).toRGBA();
}
}
}
@@ -161,12 +180,17 @@
void FEMorphology::platformApplyWorker(PlatformApplyParameters* param)
{
- param->filter->platformApplyGeneric(param->paintingData, param->startY, param->endY);
+ param->filter->platformApplyGeneric(*param->paintingData, param->startY, param->endY);
}
-void FEMorphology::platformApply(PaintingData* paintingData)
+void FEMorphology::platformApply(const PaintingData& paintingData)
{
- int optimalThreadNumber = (paintingData->width * paintingData->height) / s_minimalArea;
+ // Empirically, runtime is approximately linear over reasonable kernel sizes with a slope of about 0.65.
+ float kernelFactor = sqrt(paintingData.radiusX * paintingData.radiusY) * 0.65;
+
+ static const int minimalArea = (160 * 160); // Empirical data limit for parallel jobs
+ int optimalThreadNumber = (paintingData.width * paintingData.height * kernelFactor) / minimalArea;
+
if (optimalThreadNumber > 1) {
ParallelJobs<PlatformApplyParameters> parallelJobs(&WebCore::FEMorphology::platformApplyWorker, optimalThreadNumber);
int numOfThreads = parallelJobs.numberOfJobs();
@@ -173,8 +197,8 @@
if (numOfThreads > 1) {
// Split the job into "jobSize"-sized jobs but there a few jobs that need to be slightly larger since
// jobSize * jobs < total size. These extras are handled by the remainder "jobsWithExtra".
- const int jobSize = paintingData->height / numOfThreads;
- const int jobsWithExtra = paintingData->height % numOfThreads;
+ const int jobSize = paintingData.height / numOfThreads;
+ const int jobsWithExtra = paintingData.height % numOfThreads;
int currentY = 0;
for (int job = numOfThreads - 1; job >= 0; --job) {
PlatformApplyParameters& param = parallelJobs.parameter(job);
@@ -182,7 +206,7 @@
param.startY = currentY;
currentY += job < jobsWithExtra ? jobSize + 1 : jobSize;
param.endY = currentY;
- param.paintingData = paintingData;
+ param.paintingData = &paintingData;
}
parallelJobs.execute();
return;
@@ -190,7 +214,7 @@
// Fallback to single thread model
}
- platformApplyGeneric(paintingData, 0, paintingData->height);
+ platformApplyGeneric(paintingData, 0, paintingData.height);
}
bool FEMorphology::platformApplyDegenerate(Uint8ClampedArray& dstPixelArray, const IntRect& imageRect, int radiusX, int radiusY)
@@ -237,7 +261,7 @@
if (platformApplyDegenerate(*dstPixelArray, effectDrawingRect, radiusX, radiusY))
return;
-
+
PaintingData paintingData;
paintingData.srcPixelArray = srcPixelArray.get();
paintingData.dstPixelArray = dstPixelArray;
@@ -246,7 +270,7 @@
paintingData.radiusX = ceilf(radiusX * filter.filterScale());
paintingData.radiusY = ceilf(radiusY * filter.filterScale());
- platformApply(&paintingData);
+ platformApply(paintingData);
}
static TextStream& operator<<(TextStream& ts, const MorphologyOperatorType& type)
