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 "include/core/SkColor.h" 12 #include "include/core/SkImageInfo.h" 13 #include "include/core/SkTileMode.h" 14 #include "include/core/SkTypes.h" 15 #include "include/private/SkNx.h" 16 #include "include/private/SkTArray.h" 17 #include "src/core/SkArenaAlloc.h" 18 #include <functional> 19 #include <vector> // TODO: unused 20 21 /** 22 * SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline. 23 * 24 * It's particularly designed for situations where the potential pipeline is extremely 25 * combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ... 26 * No one wants to write specialized routines for all those combinations, and if we did, we'd 27 * end up bloating our code size dramatically. SkRasterPipeline stages can be chained together 28 * at runtime, so we can scale this problem linearly rather than combinatorically. 29 * 30 * Each stage is represented by a function conforming to a common interface and by an 31 * arbitrary context pointer. The stage funciton arguments and calling convention are 32 * designed to maximize the amount of data we can pass along the pipeline cheaply, and 33 * vary depending on CPU feature detection. 34 */ 35 36 #define SK_RASTER_PIPELINE_STAGES(M) \ 37 M(callback) M(interpreter) \ 38 M(move_src_dst) M(move_dst_src) \ 39 M(clamp_0) M(clamp_1) M(clamp_a) M(clamp_gamut) \ 40 M(unpremul) M(premul) M(premul_dst) \ 41 M(force_opaque) M(force_opaque_dst) \ 42 M(set_rgb) M(unbounded_set_rgb) M(swap_rb) M(swap_rb_dst) \ 43 M(from_srgb) M(to_srgb) \ 44 M(black_color) M(white_color) \ 45 M(uniform_color) M(unbounded_uniform_color) M(uniform_color_dst) \ 46 M(seed_shader) M(dither) \ 47 M(load_a8) M(load_a8_dst) M(store_a8) M(gather_a8) \ 48 M(load_565) M(load_565_dst) M(store_565) M(gather_565) \ 49 M(load_4444) M(load_4444_dst) M(store_4444) M(gather_4444) \ 50 M(load_f16) M(load_f16_dst) M(store_f16) M(gather_f16) \ 51 M(load_af16) M(store_af16) \ 52 M(load_rgf16) M(store_rgf16) \ 53 M(load_f32) M(load_f32_dst) M(store_f32) M(gather_f32) \ 54 M(load_rgf32) M(store_rgf32) \ 55 M(load_8888) M(load_8888_dst) M(store_8888) M(gather_8888) \ 56 M(load_rg88) M(store_rg88) \ 57 M(load_a16) M(store_a16) \ 58 M(load_rg1616) M(store_rg1616) \ 59 M(load_16161616) M(store_16161616) \ 60 M(load_1010102) M(load_1010102_dst) M(store_1010102) M(gather_1010102) \ 61 M(alpha_to_gray) M(alpha_to_gray_dst) M(bt709_luminance_or_luma_to_alpha) \ 62 M(bilerp_clamp_8888) M(bicubic_clamp_8888) \ 63 M(store_u16_be) \ 64 M(load_src) M(store_src) M(load_dst) M(store_dst) \ 65 M(scale_u8) M(scale_565) M(scale_1_float) \ 66 M( lerp_u8) M( lerp_565) M( lerp_1_float) M(lerp_native) \ 67 M(dstatop) M(dstin) M(dstout) M(dstover) \ 68 M(srcatop) M(srcin) M(srcout) M(srcover) \ 69 M(clear) M(modulate) M(multiply) M(plus_) M(screen) M(xor_) \ 70 M(colorburn) M(colordodge) M(darken) M(difference) \ 71 M(exclusion) M(hardlight) M(lighten) M(overlay) M(softlight) \ 72 M(hue) M(saturation) M(color) M(luminosity) \ 73 M(srcover_rgba_8888) \ 74 M(matrix_translate) M(matrix_scale_translate) \ 75 M(matrix_2x3) M(matrix_3x3) M(matrix_3x4) M(matrix_4x5) M(matrix_4x3) \ 76 M(matrix_perspective) \ 77 M(parametric) M(gamma_) \ 78 M(mirror_x) M(repeat_x) \ 79 M(mirror_y) M(repeat_y) \ 80 M(decal_x) M(decal_y) M(decal_x_and_y) \ 81 M(check_decal_mask) \ 82 M(negate_x) \ 83 M(bilinear) M(bicubic) \ 84 M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py) \ 85 M(bicubic_n3x) M(bicubic_n1x) M(bicubic_p1x) M(bicubic_p3x) \ 86 M(bicubic_n3y) M(bicubic_n1y) M(bicubic_p1y) M(bicubic_p3y) \ 87 M(save_xy) M(accumulate) \ 88 M(clamp_x_1) M(mirror_x_1) M(repeat_x_1) \ 89 M(evenly_spaced_gradient) \ 90 M(gradient) \ 91 M(evenly_spaced_2_stop_gradient) \ 92 M(xy_to_unit_angle) \ 93 M(xy_to_radius) \ 94 M(xy_to_2pt_conical_strip) \ 95 M(xy_to_2pt_conical_focal_on_circle) \ 96 M(xy_to_2pt_conical_well_behaved) \ 97 M(xy_to_2pt_conical_smaller) \ 98 M(xy_to_2pt_conical_greater) \ 99 M(alter_2pt_conical_compensate_focal) \ 100 M(alter_2pt_conical_unswap) \ 101 M(mask_2pt_conical_nan) \ 102 M(mask_2pt_conical_degenerates) M(apply_vector_mask) \ 103 M(byte_tables) \ 104 M(rgb_to_hsl) M(hsl_to_rgb) \ 105 M(gauss_a_to_rgba) \ 106 M(emboss) \ 107 M(swizzle) 108 109 // The largest number of pixels we handle at a time. 110 static const int SkRasterPipeline_kMaxStride = 16; 111 112 // Structs representing the arguments to some common stages. 113 114 struct SkRasterPipeline_MemoryCtx { 115 void* pixels; 116 int stride; 117 }; 118 119 struct SkRasterPipeline_GatherCtx { 120 const void* pixels; 121 int stride; 122 float width; 123 float height; 124 }; 125 126 // State shared by save_xy, accumulate, and bilinear_* / bicubic_*. 127 struct SkRasterPipeline_SamplerCtx { 128 float x[SkRasterPipeline_kMaxStride]; 129 float y[SkRasterPipeline_kMaxStride]; 130 float fx[SkRasterPipeline_kMaxStride]; 131 float fy[SkRasterPipeline_kMaxStride]; 132 float scalex[SkRasterPipeline_kMaxStride]; 133 float scaley[SkRasterPipeline_kMaxStride]; 134 }; 135 136 struct SkRasterPipeline_TileCtx { 137 float scale; 138 float invScale; // cache of 1/scale 139 }; 140 141 struct SkRasterPipeline_DecalTileCtx { 142 uint32_t mask[SkRasterPipeline_kMaxStride]; 143 float limit_x; 144 float limit_y; 145 }; 146 147 struct SkRasterPipeline_SamplerCtx2 : public SkRasterPipeline_GatherCtx { 148 SkColorType ct; 149 SkTileMode tileX, tileY; 150 float invWidth, invHeight; 151 }; 152 153 struct SkRasterPipeline_CallbackCtx { 154 void (*fn)(SkRasterPipeline_CallbackCtx* self, int active_pixels/*<= SkRasterPipeline_kMaxStride*/); 155 156 // When called, fn() will have our active pixels available in rgba. 157 // When fn() returns, the pipeline will read back those active pixels from read_from. 158 float rgba[4*SkRasterPipeline_kMaxStride]; 159 float* read_from = rgba; 160 }; 161 162 namespace SkSL { 163 struct ByteCode; 164 struct ByteCodeFunction; 165 } 166 167 struct SkRasterPipeline_InterpreterCtx { 168 const SkSL::ByteCode* byteCode; 169 const SkSL::ByteCodeFunction* fn; 170 171 SkColor4f paintColor; 172 const void* inputs; 173 int ninputs; 174 bool shaderConvention; // if false, we're a colorfilter 175 }; 176 177 struct SkRasterPipeline_GradientCtx { 178 size_t stopCount; 179 float* fs[4]; 180 float* bs[4]; 181 float* ts; 182 bool interpolatedInPremul; 183 }; 184 185 struct SkRasterPipeline_EvenlySpaced2StopGradientCtx { 186 float f[4]; 187 float b[4]; 188 bool interpolatedInPremul; 189 }; 190 191 struct SkRasterPipeline_2PtConicalCtx { 192 uint32_t fMask[SkRasterPipeline_kMaxStride]; 193 float fP0, 194 fP1; 195 }; 196 197 struct SkRasterPipeline_UniformColorCtx { 198 float r,g,b,a; 199 uint16_t rgba[4]; // [0,255] in a 16-bit lane. 200 }; 201 202 struct SkRasterPipeline_EmbossCtx { 203 SkRasterPipeline_MemoryCtx mul, 204 add; 205 }; 206 207 class SkRasterPipeline { 208 public: 209 explicit SkRasterPipeline(SkArenaAlloc*); 210 211 SkRasterPipeline(const SkRasterPipeline&) = delete; 212 SkRasterPipeline(SkRasterPipeline&&) = default; 213 214 SkRasterPipeline& operator=(const SkRasterPipeline&) = delete; 215 SkRasterPipeline& operator=(SkRasterPipeline&&) = default; 216 217 void reset(); 218 219 enum StockStage { 220 #define M(stage) stage, 221 SK_RASTER_PIPELINE_STAGES(M) 222 #undef M 223 }; 224 void append(StockStage, void* = nullptr); append(StockStage stage,const void * ctx)225 void append(StockStage stage, const void* ctx) { this->append(stage, const_cast<void*>(ctx)); } 226 void append(StockStage, uintptr_t ctx); 227 // For raw functions (i.e. from a JIT). Don't use this unless you know exactly what fn needs to 228 // be. :) 229 void append(void* fn, void* ctx); 230 231 // Append all stages to this pipeline. 232 void extend(const SkRasterPipeline&); 233 234 // Runs the pipeline in 2d from (x,y) inclusive to (x+w,y+h) exclusive. 235 void run(size_t x, size_t y, size_t w, size_t h) const; 236 237 // Allocates a thunk which amortizes run() setup cost in alloc. 238 std::function<void(size_t, size_t, size_t, size_t)> compile() const; 239 240 void dump() const; 241 242 // Appends a stage for the specified matrix. 243 // Tries to optimize the stage by analyzing the type of matrix. 244 void append_matrix(SkArenaAlloc*, const SkMatrix&); 245 246 // Appends a stage for a constant uniform color. 247 // Tries to optimize the stage based on the color. 248 void append_constant_color(SkArenaAlloc*, const float rgba[4]); 249 append_constant_color(SkArenaAlloc * alloc,const SkColor4f & color)250 void append_constant_color(SkArenaAlloc* alloc, const SkColor4f& color) { 251 this->append_constant_color(alloc, color.vec()); 252 } 253 254 // Like append_constant_color() but only affecting r,g,b, ignoring the alpha channel. 255 void append_set_rgb(SkArenaAlloc*, const float rgb[3]); 256 append_set_rgb(SkArenaAlloc * alloc,const SkColor4f & color)257 void append_set_rgb(SkArenaAlloc* alloc, const SkColor4f& color) { 258 this->append_set_rgb(alloc, color.vec()); 259 } 260 261 void append_load (SkColorType, const SkRasterPipeline_MemoryCtx*); 262 void append_load_dst(SkColorType, const SkRasterPipeline_MemoryCtx*); 263 void append_store (SkColorType, const SkRasterPipeline_MemoryCtx*); 264 265 void append_gamut_clamp_if_normalized(const SkImageInfo&); 266 empty()267 bool empty() const { return fStages == nullptr; } 268 269 private: 270 struct StageList { 271 StageList* prev; 272 uint64_t stage; 273 void* ctx; 274 bool rawFunction; 275 }; 276 277 using StartPipelineFn = void(*)(size_t,size_t,size_t,size_t, void** program); 278 StartPipelineFn build_pipeline(void**) const; 279 280 void unchecked_append(StockStage, void*); 281 282 // Used by old single-program void** style execution. 283 SkArenaAlloc* fAlloc; 284 StageList* fStages; 285 int fNumStages; 286 int fSlotsNeeded; 287 }; 288 289 template <size_t bytes> 290 class SkRasterPipeline_ : public SkRasterPipeline { 291 public: SkRasterPipeline_()292 SkRasterPipeline_() 293 : SkRasterPipeline(&fBuiltinAlloc) {} 294 295 private: 296 SkSTArenaAlloc<bytes> fBuiltinAlloc; 297 }; 298 299 300 #endif//SkRasterPipeline_DEFINED 301