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 "third_party/skia/include/core/SkSize.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 #if defined (ARCH_CPU_MIPS_FAMILY) && \ 23 defined(__mips_dsp) && (__mips_dsp_rev >= 2) 24 #define SIMD_MIPS_DSPR2 1 25 #endif 26 // avoid confusion with Mac OS X's math library (Carbon) 27 #if defined(__APPLE__) 28 #undef FloatToFixed 29 #undef FixedToFloat 30 #endif 31 32 namespace skia { 33 34 // Represents a filter in one dimension. Each output pixel has one entry in this 35 // object for the filter values contributing to it. You build up the filter 36 // list by calling AddFilter for each output pixel (in order). 37 // 38 // We do 2-dimensional convolution by first convolving each row by one 39 // ConvolutionFilter1D, then convolving each column by another one. 40 // 41 // Entries are stored in fixed point, shifted left by kShiftBits. 42 class ConvolutionFilter1D { 43 public: 44 typedef short Fixed; 45 46 // The number of bits that fixed point values are shifted by. 47 enum { kShiftBits = 14 }; 48 49 SK_API ConvolutionFilter1D(); 50 SK_API ~ConvolutionFilter1D(); 51 52 // Convert between floating point and our fixed point representation. FloatToFixed(float f)53 static Fixed FloatToFixed(float f) { 54 return static_cast<Fixed>(f * (1 << kShiftBits)); 55 } FixedToChar(Fixed x)56 static unsigned char FixedToChar(Fixed x) { 57 return static_cast<unsigned char>(x >> kShiftBits); 58 } FixedToFloat(Fixed x)59 static float FixedToFloat(Fixed x) { 60 // The cast relies on Fixed being a short, implying that on 61 // the platforms we care about all (16) bits will fit into 62 // the mantissa of a (32-bit) float. 63 COMPILE_ASSERT(sizeof(Fixed) == 2, fixed_type_should_fit_in_float_mantissa); 64 float raw = static_cast<float>(x); 65 return ldexpf(raw, -kShiftBits); 66 } 67 68 // Returns the maximum pixel span of a filter. max_filter()69 int max_filter() const { return max_filter_; } 70 71 // Returns the number of filters in this filter. This is the dimension of the 72 // output image. num_values()73 int num_values() const { return static_cast<int>(filters_.size()); } 74 75 // Appends the given list of scaling values for generating a given output 76 // pixel. |filter_offset| is the distance from the edge of the image to where 77 // the scaling factors start. The scaling factors apply to the source pixels 78 // starting from this position, and going for the next |filter_length| pixels. 79 // 80 // You will probably want to make sure your input is normalized (that is, 81 // all entries in |filter_values| sub to one) to prevent affecting the overall 82 // brighness of the image. 83 // 84 // The filter_length must be > 0. 85 // 86 // This version will automatically convert your input to fixed point. 87 SK_API void AddFilter(int filter_offset, 88 const float* filter_values, 89 int filter_length); 90 91 // Same as the above version, but the input is already fixed point. 92 void AddFilter(int filter_offset, 93 const Fixed* filter_values, 94 int filter_length); 95 96 // Retrieves a filter for the given |value_offset|, a position in the output 97 // image in the direction we're convolving. The offset and length of the 98 // filter values are put into the corresponding out arguments (see AddFilter 99 // above for what these mean), and a pointer to the first scaling factor is 100 // returned. There will be |filter_length| values in this array. FilterForValue(int value_offset,int * filter_offset,int * filter_length)101 inline const Fixed* FilterForValue(int value_offset, 102 int* filter_offset, 103 int* filter_length) const { 104 const FilterInstance& filter = filters_[value_offset]; 105 *filter_offset = filter.offset; 106 *filter_length = filter.trimmed_length; 107 if (filter.trimmed_length == 0) { 108 return NULL; 109 } 110 return &filter_values_[filter.data_location]; 111 } 112 113 // Retrieves the filter for the offset 0, presumed to be the one and only. 114 // The offset and length of the filter values are put into the corresponding 115 // out arguments (see AddFilter). Note that |filter_legth| and 116 // |specified_filter_length| may be different if leading/trailing zeros of the 117 // original floating point form were clipped. 118 // There will be |filter_length| values in the return array. 119 // Returns NULL if the filter is 0-length (for instance when all floating 120 // point values passed to AddFilter were clipped to 0). 121 SK_API const Fixed* GetSingleFilter(int* specified_filter_length, 122 int* filter_offset, 123 int* filter_length) const; 124 PaddingForSIMD()125 inline void PaddingForSIMD() { 126 // Padding |padding_count| of more dummy coefficients after the coefficients 127 // of last filter to prevent SIMD instructions which load 8 or 16 bytes 128 // together to access invalid memory areas. We are not trying to align the 129 // coefficients right now due to the opaqueness of <vector> implementation. 130 // This has to be done after all |AddFilter| calls. 131 #ifdef SIMD_PADDING 132 for (int i = 0; i < SIMD_PADDING; ++i) 133 filter_values_.push_back(static_cast<Fixed>(0)); 134 #endif 135 } 136 137 private: 138 struct FilterInstance { 139 // Offset within filter_values for this instance of the filter. 140 int data_location; 141 142 // Distance from the left of the filter to the center. IN PIXELS 143 int offset; 144 145 // Number of values in this filter instance. 146 int trimmed_length; 147 148 // Filter length as specified. Note that this may be different from 149 // 'trimmed_length' if leading/trailing zeros of the original floating 150 // point form were clipped differently on each tail. 151 int length; 152 }; 153 154 // Stores the information for each filter added to this class. 155 std::vector<FilterInstance> filters_; 156 157 // We store all the filter values in this flat list, indexed by 158 // |FilterInstance.data_location| to avoid the mallocs required for storing 159 // each one separately. 160 std::vector<Fixed> filter_values_; 161 162 // The maximum size of any filter we've added. 163 int max_filter_; 164 }; 165 166 // Does a two-dimensional convolution on the given source image. 167 // 168 // It is assumed the source pixel offsets referenced in the input filters 169 // reference only valid pixels, so the source image size is not required. Each 170 // row of the source image starts |source_byte_row_stride| after the previous 171 // one (this allows you to have rows with some padding at the end). 172 // 173 // The result will be put into the given output buffer. The destination image 174 // size will be xfilter.num_values() * yfilter.num_values() pixels. It will be 175 // in rows of exactly xfilter.num_values() * 4 bytes. 176 // 177 // |source_has_alpha| is a hint that allows us to avoid doing computations on 178 // the alpha channel if the image is opaque. If you don't know, set this to 179 // true and it will work properly, but setting this to false will be a few 180 // percent faster if you know the image is opaque. 181 // 182 // The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order 183 // (this is ARGB when loaded into 32-bit words on a little-endian machine). 184 SK_API void BGRAConvolve2D(const unsigned char* source_data, 185 int source_byte_row_stride, 186 bool source_has_alpha, 187 const ConvolutionFilter1D& xfilter, 188 const ConvolutionFilter1D& yfilter, 189 int output_byte_row_stride, 190 unsigned char* output, 191 bool use_simd_if_possible); 192 193 // Does a 1D convolution of the given source image along the X dimension on 194 // a single channel of the bitmap. 195 // 196 // The function uses the same convolution kernel for each pixel. That kernel 197 // must be added to |filter| at offset 0. This is a most straightforward 198 // implementation of convolution, intended chiefly for development purposes. 199 SK_API void SingleChannelConvolveX1D(const unsigned char* source_data, 200 int source_byte_row_stride, 201 int input_channel_index, 202 int input_channel_count, 203 const ConvolutionFilter1D& filter, 204 const SkISize& image_size, 205 unsigned char* output, 206 int output_byte_row_stride, 207 int output_channel_index, 208 int output_channel_count, 209 bool absolute_values); 210 211 // Does a 1D convolution of the given source image along the Y dimension on 212 // a single channel of the bitmap. 213 SK_API void SingleChannelConvolveY1D(const unsigned char* source_data, 214 int source_byte_row_stride, 215 int input_channel_index, 216 int input_channel_count, 217 const ConvolutionFilter1D& filter, 218 const SkISize& image_size, 219 unsigned char* output, 220 int output_byte_row_stride, 221 int output_channel_index, 222 int output_channel_count, 223 bool absolute_values); 224 225 // Set up the |filter| instance with a gaussian kernel. |kernel_sigma| is the 226 // parameter of gaussian. If |derivative| is true, the kernel will be that of 227 // the first derivative. Intended for use with the two routines above. 228 SK_API void SetUpGaussianConvolutionKernel(ConvolutionFilter1D* filter, 229 float kernel_sigma, 230 bool derivative); 231 232 } // namespace skia 233 234 #endif // SKIA_EXT_CONVOLVER_H_ 235