2 * Copyright 2016 Google Inc.
4 * Use of this source code is governed by a BSD-style license that can be
5 * found in the LICENSE file.
8 #ifndef SkRasterPipeline_DEFINED
9 #define SkRasterPipeline_DEFINED
11 #include "SkArenaAlloc.h"
12 #include "SkImageInfo.h"
19 struct SkJumper_Engine;
22 * SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline.
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.
30 * Each stage is represented by a function conforming to a common interface, SkRasterPipeline::Fn,
31 * and by an arbitrary context pointer. Fn's arguments, and sometimes custom calling convention,
32 * are designed to maximize the amount of data we can pass along the pipeline cheaply.
33 * On many machines all arguments stay in registers the entire time.
35 * The meaning of the arguments to Fn are sometimes fixed:
36 * - The Stage* always represents the current stage, mainly providing access to ctx().
37 * - The first size_t is always the destination x coordinate.
38 * (If you need y, put it in your context.)
39 * - The second size_t is always tail: 0 when working on a full 4-pixel slab,
40 * or 1..3 when using only the bottom 1..3 lanes of each register.
41 * - By the time the shader's done, the first four vectors should hold source red,
42 * green, blue, and alpha, up to 4 pixels' worth each.
44 * Sometimes arguments are flexible:
45 * - In the shader, the first four vectors can be used for anything, e.g. sample coordinates.
46 * - The last four vectors are scratch registers that can be used to communicate between
47 * stages; transfer modes use these to hold the original destination pixel components.
49 * On some platforms the last four vectors are slower to work with than the other arguments.
51 * When done mutating its arguments and/or context, a stage can either:
52 * 1) call st->next() with its mutated arguments, chaining to the next stage of the pipeline; or
53 * 2) return, indicating the pipeline is complete for these pixels.
55 * Some stages that typically return are those that write a color to a destination pointer,
56 * but any stage can short-circuit the rest of the pipeline by returning instead of calling next().
59 // TODO: There may be a better place to stuff tail, e.g. in the bottom alignment bits of
60 // the Stage*. This mostly matters on 64-bit Windows where every register is precious.
62 #define SK_RASTER_PIPELINE_STAGES(M) \
64 M(move_src_dst) M(move_dst_src) M(swap) \
65 M(clamp_0) M(clamp_1) M(clamp_a) \
66 M(unpremul) M(premul) \
67 M(set_rgb) M(swap_rb) \
68 M(from_srgb) M(to_srgb) \
69 M(constant_color) M(seed_shader) M(dither) \
70 M(load_a8) M(store_a8) \
72 M(load_565) M(store_565) \
73 M(load_4444) M(store_4444) \
74 M(load_f16) M(store_f16) \
75 M(load_f32) M(store_f32) \
76 M(load_8888) M(store_8888) \
77 M(load_u16_be) M(load_rgb_u16_be) M(store_u16_be) \
78 M(load_tables_u16_be) M(load_tables_rgb_u16_be) \
79 M(load_tables) M(load_rgba) M(store_rgba) \
80 M(scale_u8) M(scale_1_float) \
81 M(lerp_u8) M(lerp_565) M(lerp_1_float) \
82 M(dstatop) M(dstin) M(dstout) M(dstover) \
83 M(srcatop) M(srcin) M(srcout) M(srcover) \
84 M(clear) M(modulate) M(multiply) M(plus_) M(screen) M(xor_) \
85 M(colorburn) M(colordodge) M(darken) M(difference) \
86 M(exclusion) M(hardlight) M(lighten) M(overlay) M(softlight) \
87 M(hue) M(saturation) M(color) M(luminosity) \
88 M(luminance_to_alpha) \
89 M(matrix_2x3) M(matrix_3x4) M(matrix_4x5) M(matrix_4x3) \
90 M(matrix_perspective) \
91 M(parametric_r) M(parametric_g) M(parametric_b) \
93 M(table_r) M(table_g) M(table_b) M(table_a) \
95 M(clamp_x) M(mirror_x) M(repeat_x) \
96 M(clamp_y) M(mirror_y) M(repeat_y) \
97 M(clamp_x_1) M(mirror_x_1) M(repeat_x_1) \
98 M(gather_a8) M(gather_g8) M(gather_i8) \
99 M(gather_565) M(gather_4444) M(gather_8888) M(gather_f16) \
100 M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py) \
101 M(bicubic_n3x) M(bicubic_n1x) M(bicubic_p1x) M(bicubic_p3x) \
102 M(bicubic_n3y) M(bicubic_n1y) M(bicubic_p1y) M(bicubic_p3y) \
103 M(save_xy) M(accumulate) \
104 M(evenly_spaced_gradient) \
106 M(evenly_spaced_2_stop_gradient) \
107 M(xy_to_unit_angle) \
109 M(byte_tables) M(byte_tables_rgb) \
113 class SkRasterPipeline {
115 explicit SkRasterPipeline(SkArenaAlloc*);
117 SkRasterPipeline(const SkRasterPipeline&) = delete;
118 SkRasterPipeline(SkRasterPipeline&&) = default;
120 SkRasterPipeline& operator=(const SkRasterPipeline&) = delete;
121 SkRasterPipeline& operator=(SkRasterPipeline&&) = default;
126 #define M(stage) stage,
127 SK_RASTER_PIPELINE_STAGES(M)
130 void append(StockStage, void* = nullptr);
131 void append(StockStage stage, const void* ctx) { this->append(stage, const_cast<void*>(ctx)); }
133 // Append all stages to this pipeline.
134 void extend(const SkRasterPipeline&);
136 // Runs the pipeline walking x through [x,x+n).
137 void run(size_t x, size_t n) const;
139 // Allocates a thunk which amortizes run() setup cost in alloc.
140 std::function<void(size_t, size_t)> compile() const;
144 // Conversion from sRGB can be subtly tricky when premultiplication is involved.
145 // Use these helpers to keep things sane.
146 void append_from_srgb(SkAlphaType);
148 bool empty() const { return fStages == nullptr; }
157 static void BuildPipeline(const StageList*, const SkJumper_Engine&, void**);
158 void unchecked_append(StockStage, void*);
159 void extend(const StageList*);
161 SkArenaAlloc* fAlloc;
166 template <size_t bytes>
167 class SkRasterPipeline_ : public SkRasterPipeline {
170 : SkRasterPipeline(&fBuiltinAlloc)
171 , fBuiltinAlloc(fBuffer) {}
175 SkArenaAlloc fBuiltinAlloc;
179 #endif//SkRasterPipeline_DEFINED