1 /* 2 * Copyright 2016 Google Inc. 3 * 4 * Use of this source code is governed by a BSD-style license that can be 5 * found in the LICENSE file. 6 */ 7 8 #ifndef SkRasterPipeline_DEFINED 9 #define SkRasterPipeline_DEFINED 10 11 #include "SkArenaAlloc.h" 12 #include "SkColor.h" 13 #include "SkImageInfo.h" 14 #include "SkNx.h" 15 #include "SkTArray.h" // TODO: unused 16 #include "SkTypes.h" 17 #include <functional> 18 #include <vector> // TODO: unused 19 20 /** 21 * SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline. 22 * 23 * It's particularly designed for situations where the potential pipeline is extremely 24 * combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ... 25 * No one wants to write specialized routines for all those combinations, and if we did, we'd 26 * end up bloating our code size dramatically. SkRasterPipeline stages can be chained together 27 * at runtime, so we can scale this problem linearly rather than combinatorically. 28 * 29 * Each stage is represented by a function conforming to a common interface and by an 30 * arbitrary context pointer. The stage funciton arguments and calling convention are 31 * designed to maximize the amount of data we can pass along the pipeline cheaply, and 32 * vary depending on CPU feature detection. 33 */ 34 35 #define SK_RASTER_PIPELINE_STAGES(M) \ 36 M(callback) \ 37 M(move_src_dst) M(move_dst_src) \ 38 M(clamp_0) M(clamp_1) M(clamp_a) M(clamp_a_dst) M(clamp_gamut) \ 39 M(unpremul) M(premul) M(premul_dst) \ 40 M(force_opaque) M(force_opaque_dst) \ 41 M(set_rgb) M(unbounded_set_rgb) M(swap_rb) M(swap_rb_dst) \ 42 M(from_srgb) M(to_srgb) \ 43 M(black_color) M(white_color) M(uniform_color) M(unbounded_uniform_color) \ 44 M(seed_shader) M(dither) \ 45 M(load_a8) M(load_a8_dst) M(store_a8) M(gather_a8) \ 46 M(load_565) M(load_565_dst) M(store_565) M(gather_565) \ 47 M(load_4444) M(load_4444_dst) M(store_4444) M(gather_4444) \ 48 M(load_f16) M(load_f16_dst) M(store_f16) M(gather_f16) \ 49 M(load_f32) M(load_f32_dst) M(store_f32) M(gather_f32) \ 50 M(load_8888) M(load_8888_dst) M(store_8888) M(gather_8888) \ 51 M(load_1010102) M(load_1010102_dst) M(store_1010102) M(gather_1010102) \ 52 M(alpha_to_gray) M(alpha_to_gray_dst) M(luminance_to_alpha) \ 53 M(bilerp_clamp_8888) \ 54 M(store_u16_be) \ 55 M(load_src) M(store_src) M(load_dst) M(store_dst) \ 56 M(scale_u8) M(scale_565) M(scale_1_float) \ 57 M( lerp_u8) M( lerp_565) M( lerp_1_float) \ 58 M(dstatop) M(dstin) M(dstout) M(dstover) \ 59 M(srcatop) M(srcin) M(srcout) M(srcover) \ 60 M(clear) M(modulate) M(multiply) M(plus_) M(screen) M(xor_) \ 61 M(colorburn) M(colordodge) M(darken) M(difference) \ 62 M(exclusion) M(hardlight) M(lighten) M(overlay) M(softlight) \ 63 M(hue) M(saturation) M(color) M(luminosity) \ 64 M(srcover_rgba_8888) \ 65 M(matrix_translate) M(matrix_scale_translate) \ 66 M(matrix_2x3) M(matrix_3x3) M(matrix_3x4) M(matrix_4x5) M(matrix_4x3) \ 67 M(matrix_perspective) \ 68 M(parametric) M(gamma) \ 69 M(mirror_x) M(repeat_x) \ 70 M(mirror_y) M(repeat_y) \ 71 M(decal_x) M(decal_y) M(decal_x_and_y) \ 72 M(check_decal_mask) \ 73 M(negate_x) \ 74 M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py) \ 75 M(bicubic_n3x) M(bicubic_n1x) M(bicubic_p1x) M(bicubic_p3x) \ 76 M(bicubic_n3y) M(bicubic_n1y) M(bicubic_p1y) M(bicubic_p3y) \ 77 M(save_xy) M(accumulate) \ 78 M(clamp_x_1) M(mirror_x_1) M(repeat_x_1) \ 79 M(evenly_spaced_gradient) \ 80 M(gradient) \ 81 M(evenly_spaced_2_stop_gradient) \ 82 M(xy_to_unit_angle) \ 83 M(xy_to_radius) \ 84 M(xy_to_2pt_conical_strip) \ 85 M(xy_to_2pt_conical_focal_on_circle) \ 86 M(xy_to_2pt_conical_well_behaved) \ 87 M(xy_to_2pt_conical_smaller) \ 88 M(xy_to_2pt_conical_greater) \ 89 M(alter_2pt_conical_compensate_focal) \ 90 M(alter_2pt_conical_unswap) \ 91 M(mask_2pt_conical_nan) \ 92 M(mask_2pt_conical_degenerates) M(apply_vector_mask) \ 93 M(byte_tables) \ 94 M(rgb_to_hsl) M(hsl_to_rgb) \ 95 M(gauss_a_to_rgba) \ 96 M(emboss) 97 98 // The largest number of pixels we handle at a time. 99 static const int SkRasterPipeline_kMaxStride = 16; 100 101 // Structs representing the arguments to some common stages. 102 103 struct SkRasterPipeline_MemoryCtx { 104 void* pixels; 105 int stride; 106 }; 107 108 struct SkRasterPipeline_GatherCtx { 109 const void* pixels; 110 int stride; 111 float width; 112 float height; 113 }; 114 115 // State shared by save_xy, accumulate, and bilinear_* / bicubic_*. 116 struct SkRasterPipeline_SamplerCtx { 117 float x[SkRasterPipeline_kMaxStride]; 118 float y[SkRasterPipeline_kMaxStride]; 119 float fx[SkRasterPipeline_kMaxStride]; 120 float fy[SkRasterPipeline_kMaxStride]; 121 float scalex[SkRasterPipeline_kMaxStride]; 122 float scaley[SkRasterPipeline_kMaxStride]; 123 }; 124 125 struct SkRasterPipeline_TileCtx { 126 float scale; 127 float invScale; // cache of 1/scale 128 }; 129 130 struct SkRasterPipeline_DecalTileCtx { 131 uint32_t mask[SkRasterPipeline_kMaxStride]; 132 float limit_x; 133 float limit_y; 134 }; 135 136 struct SkRasterPipeline_CallbackCtx { 137 void (*fn)(SkRasterPipeline_CallbackCtx* self, int active_pixels/*<= SkRasterPipeline_kMaxStride*/); 138 139 // When called, fn() will have our active pixels available in rgba. 140 // When fn() returns, the pipeline will read back those active pixels from read_from. 141 float rgba[4*SkRasterPipeline_kMaxStride]; 142 float* read_from = rgba; 143 }; 144 145 struct SkRasterPipeline_GradientCtx { 146 size_t stopCount; 147 float* fs[4]; 148 float* bs[4]; 149 float* ts; 150 bool interpolatedInPremul; 151 }; 152 153 struct SkRasterPipeline_EvenlySpaced2StopGradientCtx { 154 float f[4]; 155 float b[4]; 156 bool interpolatedInPremul; 157 }; 158 159 struct SkRasterPipeline_2PtConicalCtx { 160 uint32_t fMask[SkRasterPipeline_kMaxStride]; 161 float fP0, 162 fP1; 163 }; 164 165 struct SkRasterPipeline_UniformColorCtx { 166 float r,g,b,a; 167 uint16_t rgba[4]; // [0,255] in a 16-bit lane. 168 }; 169 170 struct SkRasterPipeline_EmbossCtx { 171 SkRasterPipeline_MemoryCtx mul, 172 add; 173 }; 174 175 176 177 class SkRasterPipeline { 178 public: 179 explicit SkRasterPipeline(SkArenaAlloc*); 180 181 SkRasterPipeline(const SkRasterPipeline&) = delete; 182 SkRasterPipeline(SkRasterPipeline&&) = default; 183 184 SkRasterPipeline& operator=(const SkRasterPipeline&) = delete; 185 SkRasterPipeline& operator=(SkRasterPipeline&&) = default; 186 187 void reset(); 188 189 enum StockStage { 190 #define M(stage) stage, 191 SK_RASTER_PIPELINE_STAGES(M) 192 #undef M 193 }; 194 void append(StockStage, void* = nullptr); append(StockStage stage,const void * ctx)195 void append(StockStage stage, const void* ctx) { this->append(stage, const_cast<void*>(ctx)); } 196 // For raw functions (i.e. from a JIT). Don't use this unless you know exactly what fn needs to 197 // be. :) 198 void append(void* fn, void* ctx); 199 200 // Append all stages to this pipeline. 201 void extend(const SkRasterPipeline&); 202 203 // Runs the pipeline in 2d from (x,y) inclusive to (x+w,y+h) exclusive. 204 void run(size_t x, size_t y, size_t w, size_t h) const; 205 206 // Allocates a thunk which amortizes run() setup cost in alloc. 207 std::function<void(size_t, size_t, size_t, size_t)> compile() const; 208 209 void dump() const; 210 211 // Appends a stage for the specified matrix. 212 // Tries to optimize the stage by analyzing the type of matrix. 213 void append_matrix(SkArenaAlloc*, const SkMatrix&); 214 215 // Appends a stage for a constant uniform color. 216 // Tries to optimize the stage based on the color. 217 void append_constant_color(SkArenaAlloc*, const float rgba[4]); 218 append_constant_color(SkArenaAlloc * alloc,const SkColor4f & color)219 void append_constant_color(SkArenaAlloc* alloc, const SkColor4f& color) { 220 this->append_constant_color(alloc, color.vec()); 221 } 222 223 // Like append_constant_color() but only affecting r,g,b, ignoring the alpha channel. 224 void append_set_rgb(SkArenaAlloc*, const float rgb[3]); 225 append_set_rgb(SkArenaAlloc * alloc,const SkColor4f & color)226 void append_set_rgb(SkArenaAlloc* alloc, const SkColor4f& color) { 227 this->append_set_rgb(alloc, color.vec()); 228 } 229 230 void append_load (SkColorType, const SkRasterPipeline_MemoryCtx*); 231 void append_load_dst(SkColorType, const SkRasterPipeline_MemoryCtx*); 232 void append_store (SkColorType, const SkRasterPipeline_MemoryCtx*); 233 234 void append_gamut_clamp_if_normalized(const SkImageInfo&); 235 empty()236 bool empty() const { return fStages == nullptr; } 237 238 239 private: 240 struct StageList { 241 StageList* prev; 242 uint64_t stage; 243 void* ctx; 244 bool rawFunction; 245 }; 246 247 using StartPipelineFn = void(*)(size_t,size_t,size_t,size_t, void** program); 248 StartPipelineFn build_pipeline(void**) const; 249 250 void unchecked_append(StockStage, void*); 251 252 // Used by old single-program void** style execution. 253 SkArenaAlloc* fAlloc; 254 StageList* fStages; 255 int fNumStages; 256 int fSlotsNeeded; 257 }; 258 259 template <size_t bytes> 260 class SkRasterPipeline_ : public SkRasterPipeline { 261 public: SkRasterPipeline_()262 SkRasterPipeline_() 263 : SkRasterPipeline(&fBuiltinAlloc) {} 264 265 private: 266 SkSTArenaAlloc<bytes> fBuiltinAlloc; 267 }; 268 269 270 #endif//SkRasterPipeline_DEFINED 271