skia/ext/convolver.h

   1 // Copyright (c) 2012 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #ifndef SKIA_EXT_CONVOLVER_H_
   6 #define SKIA_EXT_CONVOLVER_H_
   7
   8 #include <cmath>
   9 #include <vector>
  10
  11 #include "base/basictypes.h"
  12 #include "base/cpu.h"
  13 #include "third_party/skia/include/core/SkTypes.h"
  14
  15 // We can build SSE2 optimized versions for all x86 CPUs
  16 // except when building for the IOS emulator.
  17 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_IOS)
  18 #define SIMD_SSE2 1
  19 #define SIMD_PADDING 8  // 8 * int16
  20 #endif
  21
  22 // avoid confusion with Mac OS X's math library (Carbon)
  23 #if defined(__APPLE__)
  24 #undef FloatToFixed
  25 #undef FixedToFloat
  26 #endif
  27
  28 namespace skia {
  29
  30 // Represents a filter in one dimension. Each output pixel has one entry in this
  31 // object for the filter values contributing to it. You build up the filter
  32 // list by calling AddFilter for each output pixel (in order).
  33 //
  34 // We do 2-dimensional convolution by first convolving each row by one
  35 // ConvolutionFilter1D, then convolving each column by another one.
  36 //
  37 // Entries are stored in fixed point, shifted left by kShiftBits.
  38 class ConvolutionFilter1D {
  39  public:
  40   typedef short Fixed;
  41
  42   // The number of bits that fixed point values are shifted by.
  43   enum { kShiftBits = 14 };
  44
  45   SK_API ConvolutionFilter1D();
  46   SK_API ~ConvolutionFilter1D();
  47
  48   // Convert between floating point and our fixed point representation.
  49   static Fixed FloatToFixed(float f) {
  50     return static_cast<Fixed>(f * (1 << kShiftBits));
  51   }
  52   static unsigned char FixedToChar(Fixed x) {
  53     return static_cast<unsigned char>(x >> kShiftBits);
  54   }
  55   static float FixedToFloat(Fixed x) {
  56     // The cast relies on Fixed being a short, implying that on
  57     // the platforms we care about all (16) bits will fit into
  58     // the mantissa of a (32-bit) float.
  59     COMPILE_ASSERT(sizeof(Fixed) == 2, fixed_type_should_fit_in_float_mantissa);
  60     float raw = static_cast<float>(x);
  61     return ldexpf(raw, -kShiftBits);
  62   }
  63
  64   // Returns the maximum pixel span of a filter.
  65   int max_filter() const { return max_filter_; }
  66
  67   // Returns the number of filters in this filter. This is the dimension of the
  68   // output image.
  69   int num_values() const { return static_cast<int>(filters_.size()); }
  70
  71   // Appends the given list of scaling values for generating a given output
  72   // pixel. |filter_offset| is the distance from the edge of the image to where
  73   // the scaling factors start. The scaling factors apply to the source pixels
  74   // starting from this position, and going for the next |filter_length| pixels.
  75   //
  76   // You will probably want to make sure your input is normalized (that is,
  77   // all entries in |filter_values| sub to one) to prevent affecting the overall
  78   // brighness of the image.
  79   //
  80   // The filter_length must be > 0.
  81   //
  82   // This version will automatically convert your input to fixed point.
  83   SK_API void AddFilter(int filter_offset,
  84                         const float* filter_values,
  85                         int filter_length);
  86
  87   // Same as the above version, but the input is already fixed point.
  88   void AddFilter(int filter_offset,
  89                  const Fixed* filter_values,
  90                  int filter_length);
  91
  92   // Retrieves a filter for the given |value_offset|, a position in the output
  93   // image in the direction we're convolving. The offset and length of the
  94   // filter values are put into the corresponding out arguments (see AddFilter
  95   // above for what these mean), and a pointer to the first scaling factor is
  96   // returned. There will be |filter_length| values in this array.
  97   inline const Fixed* FilterForValue(int value_offset,
  98                                      int* filter_offset,
  99                                      int* filter_length) const {
 100     const FilterInstance& filter = filters_[value_offset];
 101     *filter_offset = filter.offset;
 102     *filter_length = filter.length;
 103     if (filter.length == 0) {
 104       return NULL;
 105     }
 106     return &filter_values_[filter.data_location];
 107   }
 108
 109
 110   inline void PaddingForSIMD() {
 111     // Padding |padding_count| of more dummy coefficients after the coefficients
 112     // of last filter to prevent SIMD instructions which load 8 or 16 bytes
 113     // together to access invalid memory areas. We are not trying to align the
 114     // coefficients right now due to the opaqueness of <vector> implementation.
 115     // This has to be done after all |AddFilter| calls.
 116 #ifdef SIMD_PADDING
 117     for (int i = 0; i < SIMD_PADDING; ++i)
 118       filter_values_.push_back(static_cast<Fixed>(0));
 119 #endif
 120   }
 121
 122  private:
 123   struct FilterInstance {
 124     // Offset within filter_values for this instance of the filter.
 125     int data_location;
 126
 127     // Distance from the left of the filter to the center. IN PIXELS
 128     int offset;
 129
 130     // Number of values in this filter instance.
 131     int length;
 132   };
 133
 134   // Stores the information for each filter added to this class.
 135   std::vector<FilterInstance> filters_;
 136
 137   // We store all the filter values in this flat list, indexed by
 138   // |FilterInstance.data_location| to avoid the mallocs required for storing
 139   // each one separately.
 140   std::vector<Fixed> filter_values_;
 141
 142   // The maximum size of any filter we've added.
 143   int max_filter_;
 144 };
 145
 146 // Does a two-dimensional convolution on the given source image.
 147 //
 148 // It is assumed the source pixel offsets referenced in the input filters
 149 // reference only valid pixels, so the source image size is not required. Each
 150 // row of the source image starts |source_byte_row_stride| after the previous
 151 // one (this allows you to have rows with some padding at the end).
 152 //
 153 // The result will be put into the given output buffer. The destination image
 154 // size will be xfilter.num_values() * yfilter.num_values() pixels. It will be
 155 // in rows of exactly xfilter.num_values() * 4 bytes.
 156 //
 157 // |source_has_alpha| is a hint that allows us to avoid doing computations on
 158 // the alpha channel if the image is opaque. If you don't know, set this to
 159 // true and it will work properly, but setting this to false will be a few
 160 // percent faster if you know the image is opaque.
 161 //
 162 // The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order
 163 // (this is ARGB when loaded into 32-bit words on a little-endian machine).
 164 SK_API void BGRAConvolve2D(const unsigned char* source_data,
 165                            int source_byte_row_stride,
 166                            bool source_has_alpha,
 167                            const ConvolutionFilter1D& xfilter,
 168                            const ConvolutionFilter1D& yfilter,
 169                            int output_byte_row_stride,
 170                            unsigned char* output,
 171                            bool use_simd_if_possible);
 172 }  // namespace skia
 173
 174 #endif  // SKIA_EXT_CONVOLVER_H_