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/SkTypes.h" 13 #include "include/private/base/SkMacros.h" 14 #include "include/private/base/SkSpan_impl.h" 15 #include "include/private/base/SkTArray.h" 16 #include "src/base/SkArenaAlloc.h" 17 #include "src/core/SkRasterPipelineOpContexts.h" 18 19 #include <cstddef> 20 #include <cstdint> 21 #include <functional> 22 23 class SkMatrix; 24 enum class SkRasterPipelineOp; 25 enum SkColorType : int; 26 struct SkImageInfo; 27 struct skcms_TransferFunction; 28 29 #if __has_cpp_attribute(clang::musttail) && !defined(__EMSCRIPTEN__) && !defined(SK_CPU_ARM32) && \ 30 !defined(SK_CPU_LOONGARCH) 31 #define SK_HAS_MUSTTAIL 1 32 #else 33 #define SK_HAS_MUSTTAIL 0 34 #endif 35 36 /** 37 * SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline. 38 * 39 * It's particularly designed for situations where the potential pipeline is extremely 40 * combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ... 41 * No one wants to write specialized routines for all those combinations, and if we did, we'd 42 * end up bloating our code size dramatically. SkRasterPipeline stages can be chained together 43 * at runtime, so we can scale this problem linearly rather than combinatorically. 44 * 45 * Each stage is represented by a function conforming to a common interface and by an 46 * arbitrary context pointer. The stage function arguments and calling convention are 47 * designed to maximize the amount of data we can pass along the pipeline cheaply, and 48 * vary depending on CPU feature detection. 49 */ 50 51 // Raster pipeline programs are stored as a contiguous array of SkRasterPipelineStages. 52 SK_BEGIN_REQUIRE_DENSE 53 struct SkRasterPipelineStage { 54 // `fn` holds a function pointer from `ops_lowp` or `ops_highp` in SkOpts.cpp. These functions 55 // correspond to operations from the SkRasterPipelineOp enum in SkRasterPipelineOpList.h. The 56 // exact function pointer type varies depending on architecture (specifically, look for `using 57 // Stage =` in SkRasterPipeline_opts.h). 58 void (*fn)(); 59 60 // `ctx` holds data used by the stage function. 61 // Most context structures are declared in SkRasterPipelineOpContexts.h, and have names ending 62 // in Ctx (e.g. "SkRasterPipeline_SamplerCtx"). Some Raster Pipeline stages pack non-pointer 63 // data into this field using `SkRPCtxUtils::Pack`. 64 void* ctx; 65 }; 66 SK_END_REQUIRE_DENSE 67 68 class SkRasterPipeline { 69 public: 70 explicit SkRasterPipeline(SkArenaAlloc*); 71 72 SkRasterPipeline(const SkRasterPipeline&) = delete; 73 SkRasterPipeline(SkRasterPipeline&&) = default; 74 75 SkRasterPipeline& operator=(const SkRasterPipeline&) = delete; 76 SkRasterPipeline& operator=(SkRasterPipeline&&) = default; 77 78 void reset(); 79 80 void append(SkRasterPipelineOp, void* = nullptr); append(SkRasterPipelineOp op,const void * ctx)81 void append(SkRasterPipelineOp op, const void* ctx) { this->append(op,const_cast<void*>(ctx)); } 82 void append(SkRasterPipelineOp, uintptr_t ctx); 83 84 // Append all stages to this pipeline. 85 void extend(const SkRasterPipeline&); 86 87 // Runs the pipeline in 2d from (x,y) inclusive to (x+w,y+h) exclusive. 88 void run(size_t x, size_t y, size_t w, size_t h) const; 89 90 // Allocates a thunk which amortizes run() setup cost in alloc. 91 std::function<void(size_t, size_t, size_t, size_t)> compile() const; 92 93 // Callers can inspect the stage list for debugging purposes. 94 struct StageList { 95 StageList* prev; 96 SkRasterPipelineOp stage; 97 void* ctx; 98 }; 99 100 static const char* GetOpName(SkRasterPipelineOp op); getStageList()101 const StageList* getStageList() const { return fStages; } getNumStages()102 int getNumStages() const { return fNumStages; } 103 104 // Prints the entire StageList using SkDebugf. 105 void dump() const; 106 107 // Appends a stage for the specified matrix. 108 // Tries to optimize the stage by analyzing the type of matrix. 109 void appendMatrix(SkArenaAlloc*, const SkMatrix&); 110 111 // Appends a stage for a constant uniform color. 112 // Tries to optimize the stage based on the color. 113 void appendConstantColor(SkArenaAlloc*, const float rgba[4]); 114 appendConstantColor(SkArenaAlloc * alloc,const SkColor4f & color)115 void appendConstantColor(SkArenaAlloc* alloc, const SkColor4f& color) { 116 this->appendConstantColor(alloc, color.vec()); 117 } 118 119 // Like appendConstantColor() but only affecting r,g,b, ignoring the alpha channel. 120 void appendSetRGB(SkArenaAlloc*, const float rgb[3]); 121 appendSetRGB(SkArenaAlloc * alloc,const SkColor4f & color)122 void appendSetRGB(SkArenaAlloc* alloc, const SkColor4f& color) { 123 this->appendSetRGB(alloc, color.vec()); 124 } 125 126 void appendLoad (SkColorType, const SkRasterPipeline_MemoryCtx*); 127 void appendLoadDst(SkColorType, const SkRasterPipeline_MemoryCtx*); 128 void appendStore (SkColorType, const SkRasterPipeline_MemoryCtx*); 129 130 void appendClampIfNormalized(const SkImageInfo&); 131 132 void appendTransferFunction(const skcms_TransferFunction&); 133 134 void appendStackRewind(); 135 empty()136 bool empty() const { return fStages == nullptr; } 137 138 private: 139 bool buildLowpPipeline(SkRasterPipelineStage* ip) const; 140 void buildHighpPipeline(SkRasterPipelineStage* ip) const; 141 142 using StartPipelineFn = void (*)(size_t, size_t, size_t, size_t, 143 SkRasterPipelineStage* program, 144 SkSpan<SkRasterPipeline_MemoryCtxPatch>, 145 uint8_t*); 146 StartPipelineFn buildPipeline(SkRasterPipelineStage*) const; 147 148 void uncheckedAppend(SkRasterPipelineOp, void*); 149 int stagesNeeded() const; 150 151 void addMemoryContext(SkRasterPipeline_MemoryCtx*, int bytesPerPixel, bool load, bool store); 152 uint8_t* tailPointer(); 153 154 SkArenaAlloc* fAlloc; 155 SkRasterPipeline_RewindCtx* fRewindCtx; 156 StageList* fStages; 157 uint8_t* fTailPointer; 158 int fNumStages; 159 160 // Only 1 in 2 million CPU-backend pipelines used more than two MemoryCtxs. 161 // (See the comment in SkRasterPipelineOpContexts.h for how MemoryCtx patching works) 162 skia_private::STArray<2, SkRasterPipeline_MemoryCtxInfo> fMemoryCtxInfos; 163 }; 164 165 template <size_t bytes> 166 class SkRasterPipeline_ : public SkRasterPipeline { 167 public: SkRasterPipeline_()168 SkRasterPipeline_() 169 : SkRasterPipeline(&fBuiltinAlloc) {} 170 171 private: 172 SkSTArenaAlloc<bytes> fBuiltinAlloc; 173 }; 174 175 176 #endif//SkRasterPipeline_DEFINED 177