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 "SkImageInfo.h"
12 #include "SkNx.h"
13 #include "SkTArray.h"
14 #include "SkTypes.h"
15 #include <vector>
16 
17 /**
18  * SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline.
19  *
20  * It's particularly designed for situations where the potential pipeline is extremely
21  * combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ...
22  * No one wants to write specialized routines for all those combinations, and if we did, we'd
23  * end up bloating our code size dramatically.  SkRasterPipeline stages can be chained together
24  * at runtime, so we can scale this problem linearly rather than combinatorically.
25  *
26  * Each stage is represented by a function conforming to a common interface, SkRasterPipeline::Fn,
27  * and by an arbitrary context pointer.  Fn's arguments, and sometimes custom calling convention,
28  * are designed to maximize the amount of data we can pass along the pipeline cheaply.
29  * On many machines all arguments stay in registers the entire time.
30  *
31  * The meaning of the arguments to Fn are sometimes fixed:
32  *    - The Stage* always represents the current stage, mainly providing access to ctx().
33  *    - The first size_t is always the destination x coordinate.
34  *      (If you need y, put it in your context.)
35  *    - The second size_t is always tail: 0 when working on a full 4-pixel slab,
36  *      or 1..3 when using only the bottom 1..3 lanes of each register.
37  *    - By the time the shader's done, the first four vectors should hold source red,
38  *      green, blue, and alpha, up to 4 pixels' worth each.
39  *
40  * Sometimes arguments are flexible:
41  *    - In the shader, the first four vectors can be used for anything, e.g. sample coordinates.
42  *    - The last four vectors are scratch registers that can be used to communicate between
43  *      stages; transfer modes use these to hold the original destination pixel components.
44  *
45  * On some platforms the last four vectors are slower to work with than the other arguments.
46  *
47  * When done mutating its arguments and/or context, a stage can either:
48  *   1) call st->next() with its mutated arguments, chaining to the next stage of the pipeline; or
49  *   2) return, indicating the pipeline is complete for these pixels.
50  *
51  * Some stages that typically return are those that write a color to a destination pointer,
52  * but any stage can short-circuit the rest of the pipeline by returning instead of calling next().
53  */
54 
55 // TODO: There may be a better place to stuff tail, e.g. in the bottom alignment bits of
56 // the Stage*.  This mostly matters on 64-bit Windows where every register is precious.
57 
58 #define SK_RASTER_PIPELINE_STAGES(M)                             \
59     M(trace) M(registers)                                        \
60     M(move_src_dst) M(move_dst_src) M(swap)                      \
61     M(clamp_0) M(clamp_1) M(clamp_a)                             \
62     M(unpremul) M(premul)                                        \
63     M(set_rgb) M(swap_rb)                                        \
64     M(from_srgb) M(to_srgb)                                      \
65     M(from_2dot2) M(to_2dot2)                                    \
66     M(constant_color) M(seed_shader)                             \
67     M(load_a8)   M(store_a8)                                     \
68     M(load_g8)                                                   \
69     M(load_565)  M(store_565)                                    \
70     M(load_4444) M(store_4444)                                   \
71     M(load_f16)  M(store_f16)                                    \
72     M(load_f32)  M(store_f32)                                    \
73     M(load_8888) M(store_8888)                                   \
74     M(load_u16_be) M(load_rgb_u16_be) M(store_u16_be)            \
75     M(load_tables_u16_be) M(load_tables_rgb_u16_be)              \
76     M(load_tables)                                               \
77     M(scale_u8) M(scale_1_float)                                 \
78     M(lerp_u8) M(lerp_565) M(lerp_1_float)                       \
79     M(dstatop) M(dstin) M(dstout) M(dstover)                     \
80     M(srcatop) M(srcin) M(srcout) M(srcover)                     \
81     M(clear) M(modulate) M(multiply) M(plus_) M(screen) M(xor_)  \
82     M(colorburn) M(colordodge) M(darken) M(difference)           \
83     M(exclusion) M(hardlight) M(lighten) M(overlay) M(softlight) \
84     M(luminance_to_alpha)                                        \
85     M(matrix_2x3) M(matrix_3x4) M(matrix_4x5)                    \
86     M(matrix_perspective)                                        \
87     M(parametric_r) M(parametric_g) M(parametric_b)              \
88     M(parametric_a)                                              \
89     M(table_r) M(table_g) M(table_b) M(table_a)                  \
90     M(color_lookup_table) M(lab_to_xyz)                          \
91     M(clamp_x) M(mirror_x) M(repeat_x)                           \
92     M(clamp_y) M(mirror_y) M(repeat_y)                           \
93     M(gather_a8) M(gather_g8) M(gather_i8)                       \
94     M(gather_565) M(gather_4444) M(gather_8888) M(gather_f16)    \
95     M(bilinear_nx) M(bilinear_px) M(bilinear_ny) M(bilinear_py)  \
96     M(bicubic_n3x) M(bicubic_n1x) M(bicubic_p1x) M(bicubic_p3x)  \
97     M(bicubic_n3y) M(bicubic_n1y) M(bicubic_p1y) M(bicubic_p3y)  \
98     M(save_xy) M(accumulate)                                     \
99     M(linear_gradient_2stops)                                    \
100     M(byte_tables) M(byte_tables_rgb)                            \
101     M(shader_adapter)                                            \
102     M(rgb_to_hsl)                                                \
103     M(hsl_to_rgb)
104 
105 class SkRasterPipeline {
106 public:
107     SkRasterPipeline();
108 
109     enum StockStage {
110     #define M(stage) stage,
111         SK_RASTER_PIPELINE_STAGES(M)
112     #undef M
113     };
114     void append(StockStage, void* = nullptr);
append(StockStage stage,const void * ctx)115     void append(StockStage stage, const void* ctx) { this->append(stage, const_cast<void*>(ctx)); }
116 
117     // Append all stages to this pipeline.
118     void extend(const SkRasterPipeline&);
119 
120     // Runs the pipeline walking x through [x,x+n).
121     void run(size_t x, size_t n) const;
122 
123     void dump() const;
124 
125     struct Stage {
126         StockStage stage;
127         void*        ctx;
128     };
129 
130     // Conversion from sRGB can be subtly tricky when premultiplication is involved.
131     // Use these helpers to keep things sane.
132     void append_from_srgb(SkAlphaType);
133 
empty()134     bool empty() const { return fStages.empty(); }
135 
136 private:
137     bool run_with_jumper(size_t x, size_t n) const;
138 
139     std::vector<Stage> fStages;
140 };
141 
142 #endif//SkRasterPipeline_DEFINED
143