1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/gfx/2d/convolver.h Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,191 @@
1.4 +// Copyright (c) 2006-2012 The Chromium Authors. All rights reserved.
1.5 +//
1.6 +// Redistribution and use in source and binary forms, with or without
1.7 +// modification, are permitted provided that the following conditions
1.8 +// are met:
1.9 +// * Redistributions of source code must retain the above copyright
1.10 +// notice, this list of conditions and the following disclaimer.
1.11 +// * Redistributions in binary form must reproduce the above copyright
1.12 +// notice, this list of conditions and the following disclaimer in
1.13 +// the documentation and/or other materials provided with the
1.14 +// distribution.
1.15 +// * Neither the name of Google, Inc. nor the names of its contributors
1.16 +// may be used to endorse or promote products derived from this
1.17 +// software without specific prior written permission.
1.18 +//
1.19 +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
1.20 +// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
1.21 +// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
1.22 +// FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
1.23 +// COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
1.24 +// INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
1.25 +// BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
1.26 +// OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
1.27 +// AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
1.28 +// OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
1.29 +// OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
1.30 +// SUCH DAMAGE.
1.31 +
1.32 +#ifndef SKIA_EXT_CONVOLVER_H_
1.33 +#define SKIA_EXT_CONVOLVER_H_
1.34 +
1.35 +#include <cmath>
1.36 +#include <vector>
1.37 +
1.38 +#include "base/basictypes.h"
1.39 +#include "base/cpu.h"
1.40 +#include "mozilla/Assertions.h"
1.41 +#include "skia/SkTypes.h"
1.42 +
1.43 +// avoid confusion with Mac OS X's math library (Carbon)
1.44 +#if defined(__APPLE__)
1.45 +#undef FloatToFixed
1.46 +#undef FixedToFloat
1.47 +#endif
1.48 +
1.49 +namespace skia {
1.50 +
1.51 +// Represents a filter in one dimension. Each output pixel has one entry in this
1.52 +// object for the filter values contributing to it. You build up the filter
1.53 +// list by calling AddFilter for each output pixel (in order).
1.54 +//
1.55 +// We do 2-dimensional convolution by first convolving each row by one
1.56 +// ConvolutionFilter1D, then convolving each column by another one.
1.57 +//
1.58 +// Entries are stored in fixed point, shifted left by kShiftBits.
1.59 +class ConvolutionFilter1D {
1.60 + public:
1.61 + typedef short Fixed;
1.62 +
1.63 + // The number of bits that fixed point values are shifted by.
1.64 + enum { kShiftBits = 14 };
1.65 +
1.66 + ConvolutionFilter1D();
1.67 + ~ConvolutionFilter1D();
1.68 +
1.69 + // Convert between floating point and our fixed point representation.
1.70 + static Fixed FloatToFixed(float f) {
1.71 + return static_cast<Fixed>(f * (1 << kShiftBits));
1.72 + }
1.73 + static unsigned char FixedToChar(Fixed x) {
1.74 + return static_cast<unsigned char>(x >> kShiftBits);
1.75 + }
1.76 + static float FixedToFloat(Fixed x) {
1.77 + // The cast relies on Fixed being a short, implying that on
1.78 + // the platforms we care about all (16) bits will fit into
1.79 + // the mantissa of a (32-bit) float.
1.80 + static_assert(sizeof(Fixed) == 2,
1.81 + "fixed type should fit in float mantissa");
1.82 + float raw = static_cast<float>(x);
1.83 + return ldexpf(raw, -kShiftBits);
1.84 + }
1.85 +
1.86 + // Returns the maximum pixel span of a filter.
1.87 + int max_filter() const { return max_filter_; }
1.88 +
1.89 + // Returns the number of filters in this filter. This is the dimension of the
1.90 + // output image.
1.91 + int num_values() const { return static_cast<int>(filters_.size()); }
1.92 +
1.93 + // Appends the given list of scaling values for generating a given output
1.94 + // pixel. |filter_offset| is the distance from the edge of the image to where
1.95 + // the scaling factors start. The scaling factors apply to the source pixels
1.96 + // starting from this position, and going for the next |filter_length| pixels.
1.97 + //
1.98 + // You will probably want to make sure your input is normalized (that is,
1.99 + // all entries in |filter_values| sub to one) to prevent affecting the overall
1.100 + // brighness of the image.
1.101 + //
1.102 + // The filter_length must be > 0.
1.103 + //
1.104 + // This version will automatically convert your input to fixed point.
1.105 + void AddFilter(int filter_offset,
1.106 + const float* filter_values,
1.107 + int filter_length);
1.108 +
1.109 + // Same as the above version, but the input is already fixed point.
1.110 + void AddFilter(int filter_offset,
1.111 + const Fixed* filter_values,
1.112 + int filter_length);
1.113 +
1.114 + // Retrieves a filter for the given |value_offset|, a position in the output
1.115 + // image in the direction we're convolving. The offset and length of the
1.116 + // filter values are put into the corresponding out arguments (see AddFilter
1.117 + // above for what these mean), and a pointer to the first scaling factor is
1.118 + // returned. There will be |filter_length| values in this array.
1.119 + inline const Fixed* FilterForValue(int value_offset,
1.120 + int* filter_offset,
1.121 + int* filter_length) const {
1.122 + const FilterInstance& filter = filters_[value_offset];
1.123 + *filter_offset = filter.offset;
1.124 + *filter_length = filter.length;
1.125 + if (filter.length == 0) {
1.126 + return NULL;
1.127 + }
1.128 + return &filter_values_[filter.data_location];
1.129 + }
1.130 +
1.131 +
1.132 + inline void PaddingForSIMD(int padding_count) {
1.133 + // Padding |padding_count| of more dummy coefficients after the coefficients
1.134 + // of last filter to prevent SIMD instructions which load 8 or 16 bytes
1.135 + // together to access invalid memory areas. We are not trying to align the
1.136 + // coefficients right now due to the opaqueness of <vector> implementation.
1.137 + // This has to be done after all |AddFilter| calls.
1.138 + for (int i = 0; i < padding_count; ++i)
1.139 + filter_values_.push_back(static_cast<Fixed>(0));
1.140 + }
1.141 +
1.142 + private:
1.143 + struct FilterInstance {
1.144 + // Offset within filter_values for this instance of the filter.
1.145 + int data_location;
1.146 +
1.147 + // Distance from the left of the filter to the center. IN PIXELS
1.148 + int offset;
1.149 +
1.150 + // Number of values in this filter instance.
1.151 + int length;
1.152 + };
1.153 +
1.154 + // Stores the information for each filter added to this class.
1.155 + std::vector<FilterInstance> filters_;
1.156 +
1.157 + // We store all the filter values in this flat list, indexed by
1.158 + // |FilterInstance.data_location| to avoid the mallocs required for storing
1.159 + // each one separately.
1.160 + std::vector<Fixed> filter_values_;
1.161 +
1.162 + // The maximum size of any filter we've added.
1.163 + int max_filter_;
1.164 +};
1.165 +
1.166 +// Does a two-dimensional convolution on the given source image.
1.167 +//
1.168 +// It is assumed the source pixel offsets referenced in the input filters
1.169 +// reference only valid pixels, so the source image size is not required. Each
1.170 +// row of the source image starts |source_byte_row_stride| after the previous
1.171 +// one (this allows you to have rows with some padding at the end).
1.172 +//
1.173 +// The result will be put into the given output buffer. The destination image
1.174 +// size will be xfilter.num_values() * yfilter.num_values() pixels. It will be
1.175 +// in rows of exactly xfilter.num_values() * 4 bytes.
1.176 +//
1.177 +// |source_has_alpha| is a hint that allows us to avoid doing computations on
1.178 +// the alpha channel if the image is opaque. If you don't know, set this to
1.179 +// true and it will work properly, but setting this to false will be a few
1.180 +// percent faster if you know the image is opaque.
1.181 +//
1.182 +// The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order
1.183 +// (this is ARGB when loaded into 32-bit words on a little-endian machine).
1.184 +void BGRAConvolve2D(const unsigned char* source_data,
1.185 + int source_byte_row_stride,
1.186 + bool source_has_alpha,
1.187 + const ConvolutionFilter1D& xfilter,
1.188 + const ConvolutionFilter1D& yfilter,
1.189 + int output_byte_row_stride,
1.190 + unsigned char* output,
1.191 + bool use_sse2);
1.192 +} // namespace skia
1.193 +
1.194 +#endif // SKIA_EXT_CONVOLVER_H_
