1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
|
/*
* Copyright 2016 Google Inc.
*
* Use of this source code is governed by a BSD-style license that can be
* found in the LICENSE file.
*/
#ifndef SkRasterPipeline_DEFINED
#define SkRasterPipeline_DEFINED
#include "SkNx.h"
#include "SkTArray.h"
#include "SkTypes.h"
/**
* SkRasterPipeline provides a cheap way to chain together a pixel processing pipeline.
*
* It's particularly designed for situations where the potential pipeline is extremely
* combinatoric: {N dst formats} x {M source formats} x {K mask formats} x {C transfer modes} ...
* No one wants to write specialized routines for all those combinations, and if we did, we'd
* end up bloating our code size dramatically. SkRasterPipeline stages can be chained together
* at runtime, so we can scale this problem linearly rather than combinatorically.
*
* Each stage is represented by a function conforming to a common interface, SkRasterPipeline::Fn,
* and by an arbitrary context pointer. Fn's arguments, and sometimes custom calling convention,
* are designed to maximize the amount of data we can pass along the pipeline cheaply.
* On many machines all arguments stay in registers the entire time.
*
* The meaning of the arguments to Fn are sometimes fixed:
* - The Stage* always represents the current stage, mainly providing access to ctx().
* - The first size_t is always the destination x coordinate.
* (If you need y, put it in your context.)
* - The second size_t is always tail: 0 when working on a full 4-pixel slab,
* or 1..3 when using only the bottom 1..3 lanes of each register.
* - By the time the shader's done, the first four vectors should hold source red,
* green, blue, and alpha, up to 4 pixels' worth each.
*
* Sometimes arguments are flexible:
* - In the shader, the first four vectors can be used for anything, e.g. sample coordinates.
* - The last four vectors are scratch registers that can be used to communicate between
* stages; transfer modes use these to hold the original destination pixel components.
*
* On some platforms the last four vectors are slower to work with than the other arguments.
*
* When done mutating its arguments and/or context, a stage can either:
* 1) call st->next() with its mutated arguments, chaining to the next stage of the pipeline; or
* 2) return, indicating the pipeline is complete for these pixels.
*
* Some stages that typically return are those that write a color to a destination pointer,
* but any stage can short-circuit the rest of the pipeline by returning instead of calling next().
*/
// TODO: There may be a better place to stuff tail, e.g. in the bottom alignment bits of
// the Stage*. This mostly matters on 64-bit Windows where every register is precious.
class SkRasterPipeline {
public:
struct Stage;
using Fn = void(SK_VECTORCALL *)(Stage*, size_t, size_t, Sk4f,Sk4f,Sk4f,Sk4f,
Sk4f,Sk4f,Sk4f,Sk4f);
struct Stage {
template <typename T>
T ctx() { return static_cast<T>(fCtx); }
void SK_VECTORCALL next(size_t x, size_t tail, Sk4f v0, Sk4f v1, Sk4f v2, Sk4f v3,
Sk4f v4, Sk4f v5, Sk4f v6, Sk4f v7) {
// Stages are logically a pipeline, and physically are contiguous in an array.
// To get to the next stage, we just increment our pointer to the next array element.
fNext(this+1, x,tail, v0,v1,v2,v3, v4,v5,v6,v7);
}
// It makes next() a good bit cheaper if we hold the next function to call here,
// rather than logically simpler choice of the function implementing this stage.
Fn fNext;
void* fCtx;
};
SkRasterPipeline();
// Run the pipeline constructed with append(), walking x through [x,x+n),
// generally in 4-pixel steps, with perhaps one jagged tail step.
void run(size_t x, size_t n);
void run(size_t n) { this->run(0, n); }
// body() will only be called with tail=0, indicating it always works on a full 4 pixels.
// tail() will only be called with tail=1..3 to handle the jagged end of n%4 pixels.
void append(Fn body, Fn tail, const void* ctx = nullptr);
void append(Fn fn, const void* ctx = nullptr) { this->append(fn, fn, ctx); }
enum StockStage {
store_565,
store_srgb,
store_f16,
load_s_565,
load_s_srgb,
load_s_f16,
load_d_565,
load_d_srgb,
load_d_f16,
scale_u8,
lerp_u8,
lerp_565,
lerp_constant_float,
constant_color,
srcover,
kNumStockStages,
};
void append(StockStage, const void* ctx=nullptr);
// Append all stages to this pipeline.
void extend(const SkRasterPipeline&);
private:
using Stages = SkSTArray<10, Stage, /*MEM_COPY=*/true>;
// This no-op default makes fBodyStart and fTailStart unconditionally safe to call,
// and is always the last stage's fNext as a sort of safety net to make sure even a
// buggy pipeline can't walk off its own end.
static void SK_VECTORCALL JustReturn(Stage*, size_t, size_t, Sk4f,Sk4f,Sk4f,Sk4f,
Sk4f,Sk4f,Sk4f,Sk4f);
Stages fBody,
fTail;
Fn fBodyStart = &JustReturn,
fTailStart = &JustReturn;
};
#endif//SkRasterPipeline_DEFINED
|
