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/SkSize.h" 14 #include "third_party/skia/include/core/SkTypes.h" 15 16 // We can build SSE2 optimized versions for all x86 CPUs 17 // except when building for the IOS emulator. 18 #if defined(ARCH_CPU_X86_FAMILY) && !defined(OS_IOS) 19 #define SIMD_SSE2 1 20 #define SIMD_PADDING 8 // 8 * int16 21 #endif 22 23 #if defined (ARCH_CPU_MIPS_FAMILY) && \ 24 defined(__mips_dsp) && (__mips_dsp_rev >= 2) 25 #define SIMD_MIPS_DSPR2 1 26 #endif 27 // avoid confusion with Mac OS X's math library (Carbon) 28 #if defined(__APPLE__) 29 #undef FloatToFixed 30 #undef FixedToFloat 31 #endif 32 33 namespace skia { 34 35 // Represents a filter in one dimension. Each output pixel has one entry in this 36 // object for the filter values contributing to it. You build up the filter 37 // list by calling AddFilter for each output pixel (in order). 38 // 39 // We do 2-dimensional convolution by first convolving each row by one 40 // ConvolutionFilter1D, then convolving each column by another one. 41 // 42 // Entries are stored in fixed point, shifted left by kShiftBits. 43 class ConvolutionFilter1D { 44 public: 45 typedef short Fixed; 46 47 // The number of bits that fixed point values are shifted by. 48 enum { kShiftBits = 14 }; 49 50 SK_API ConvolutionFilter1D(); 51 SK_API ~ConvolutionFilter1D(); 52 53 // Convert between floating point and our fixed point representation. FloatToFixed(float f)54 static Fixed FloatToFixed(float f) { 55 return static_cast<Fixed>(f * (1 << kShiftBits)); 56 } FixedToChar(Fixed x)57 static unsigned char FixedToChar(Fixed x) { 58 return static_cast<unsigned char>(x >> kShiftBits); 59 } FixedToFloat(Fixed x)60 static float FixedToFloat(Fixed x) { 61 // The cast relies on Fixed being a short, implying that on 62 // the platforms we care about all (16) bits will fit into 63 // the mantissa of a (32-bit) float. 64 COMPILE_ASSERT(sizeof(Fixed) == 2, fixed_type_should_fit_in_float_mantissa); 65 float raw = static_cast<float>(x); 66 return ldexpf(raw, -kShiftBits); 67 } 68 69 // Returns the maximum pixel span of a filter. max_filter()70 int max_filter() const { return max_filter_; } 71 72 // Returns the number of filters in this filter. This is the dimension of the 73 // output image. num_values()74 int num_values() const { return static_cast<int>(filters_.size()); } 75 76 // Appends the given list of scaling values for generating a given output 77 // pixel. |filter_offset| is the distance from the edge of the image to where 78 // the scaling factors start. The scaling factors apply to the source pixels 79 // starting from this position, and going for the next |filter_length| pixels. 80 // 81 // You will probably want to make sure your input is normalized (that is, 82 // all entries in |filter_values| sub to one) to prevent affecting the overall 83 // brighness of the image. 84 // 85 // The filter_length must be > 0. 86 // 87 // This version will automatically convert your input to fixed point. 88 SK_API void AddFilter(int filter_offset, 89 const float* filter_values, 90 int filter_length); 91 92 // Same as the above version, but the input is already fixed point. 93 void AddFilter(int filter_offset, 94 const Fixed* filter_values, 95 int filter_length); 96 97 // Retrieves a filter for the given |value_offset|, a position in the output 98 // image in the direction we're convolving. The offset and length of the 99 // filter values are put into the corresponding out arguments (see AddFilter 100 // above for what these mean), and a pointer to the first scaling factor is 101 // returned. There will be |filter_length| values in this array. FilterForValue(int value_offset,int * filter_offset,int * filter_length)102 inline const Fixed* FilterForValue(int value_offset, 103 int* filter_offset, 104 int* filter_length) const { 105 const FilterInstance& filter = filters_[value_offset]; 106 *filter_offset = filter.offset; 107 *filter_length = filter.trimmed_length; 108 if (filter.trimmed_length == 0) { 109 return NULL; 110 } 111 return &filter_values_[filter.data_location]; 112 } 113 114 // Retrieves the filter for the offset 0, presumed to be the one and only. 115 // The offset and length of the filter values are put into the corresponding 116 // out arguments (see AddFilter). Note that |filter_legth| and 117 // |specified_filter_length| may be different if leading/trailing zeros of the 118 // original floating point form were clipped. 119 // There will be |filter_length| values in the return array. 120 // Returns NULL if the filter is 0-length (for instance when all floating 121 // point values passed to AddFilter were clipped to 0). 122 SK_API const Fixed* GetSingleFilter(int* specified_filter_length, 123 int* filter_offset, 124 int* filter_length) const; 125 PaddingForSIMD()126 inline void PaddingForSIMD() { 127 // Padding |padding_count| of more dummy coefficients after the coefficients 128 // of last filter to prevent SIMD instructions which load 8 or 16 bytes 129 // together to access invalid memory areas. We are not trying to align the 130 // coefficients right now due to the opaqueness of <vector> implementation. 131 // This has to be done after all |AddFilter| calls. 132 #ifdef SIMD_PADDING 133 for (int i = 0; i < SIMD_PADDING; ++i) 134 filter_values_.push_back(static_cast<Fixed>(0)); 135 #endif 136 } 137 138 private: 139 struct FilterInstance { 140 // Offset within filter_values for this instance of the filter. 141 int data_location; 142 143 // Distance from the left of the filter to the center. IN PIXELS 144 int offset; 145 146 // Number of values in this filter instance. 147 int trimmed_length; 148 149 // Filter length as specified. Note that this may be different from 150 // 'trimmed_length' if leading/trailing zeros of the original floating 151 // point form were clipped differently on each tail. 152 int length; 153 }; 154 155 // Stores the information for each filter added to this class. 156 std::vector<FilterInstance> filters_; 157 158 // We store all the filter values in this flat list, indexed by 159 // |FilterInstance.data_location| to avoid the mallocs required for storing 160 // each one separately. 161 std::vector<Fixed> filter_values_; 162 163 // The maximum size of any filter we've added. 164 int max_filter_; 165 }; 166 167 // Does a two-dimensional convolution on the given source image. 168 // 169 // It is assumed the source pixel offsets referenced in the input filters 170 // reference only valid pixels, so the source image size is not required. Each 171 // row of the source image starts |source_byte_row_stride| after the previous 172 // one (this allows you to have rows with some padding at the end). 173 // 174 // The result will be put into the given output buffer. The destination image 175 // size will be xfilter.num_values() * yfilter.num_values() pixels. It will be 176 // in rows of exactly xfilter.num_values() * 4 bytes. 177 // 178 // |source_has_alpha| is a hint that allows us to avoid doing computations on 179 // the alpha channel if the image is opaque. If you don't know, set this to 180 // true and it will work properly, but setting this to false will be a few 181 // percent faster if you know the image is opaque. 182 // 183 // The layout in memory is assumed to be 4-bytes per pixel in B-G-R-A order 184 // (this is ARGB when loaded into 32-bit words on a little-endian machine). 185 SK_API void BGRAConvolve2D(const unsigned char* source_data, 186 int source_byte_row_stride, 187 bool source_has_alpha, 188 const ConvolutionFilter1D& xfilter, 189 const ConvolutionFilter1D& yfilter, 190 int output_byte_row_stride, 191 unsigned char* output, 192 bool use_simd_if_possible); 193 194 // Does a 1D convolution of the given source image along the X dimension on 195 // a single channel of the bitmap. 196 // 197 // The function uses the same convolution kernel for each pixel. That kernel 198 // must be added to |filter| at offset 0. This is a most straightforward 199 // implementation of convolution, intended chiefly for development purposes. 200 SK_API void SingleChannelConvolveX1D(const unsigned char* source_data, 201 int source_byte_row_stride, 202 int input_channel_index, 203 int input_channel_count, 204 const ConvolutionFilter1D& filter, 205 const SkISize& image_size, 206 unsigned char* output, 207 int output_byte_row_stride, 208 int output_channel_index, 209 int output_channel_count, 210 bool absolute_values); 211 212 // Does a 1D convolution of the given source image along the Y dimension on 213 // a single channel of the bitmap. 214 SK_API void SingleChannelConvolveY1D(const unsigned char* source_data, 215 int source_byte_row_stride, 216 int input_channel_index, 217 int input_channel_count, 218 const ConvolutionFilter1D& filter, 219 const SkISize& image_size, 220 unsigned char* output, 221 int output_byte_row_stride, 222 int output_channel_index, 223 int output_channel_count, 224 bool absolute_values); 225 226 // Set up the |filter| instance with a gaussian kernel. |kernel_sigma| is the 227 // parameter of gaussian. If |derivative| is true, the kernel will be that of 228 // the first derivative. Intended for use with the two routines above. 229 SK_API void SetUpGaussianConvolutionKernel(ConvolutionFilter1D* filter, 230 float kernel_sigma, 231 bool derivative); 232 233 } // namespace skia 234 235 #endif // SKIA_EXT_CONVOLVER_H_ 236