1/*
2 * Copyright 2015 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#include "Benchmark.h"
9#include "SkPMFloat.h"
10
11// Used to prevent the compiler from optimizing away the whole loop.
12volatile uint32_t blackhole = 0;
13
14// Not a great random number generator, but it's very fast.
15// The code we're measuring is quite fast, so low overhead is essential.
16static uint32_t lcg_rand(uint32_t* seed) {
17    *seed *= 1664525;
18    *seed += 1013904223;
19    return *seed;
20}
21
22// I'm having better luck getting these to constant-propagate away as template parameters.
23template <bool kClamp, bool kWide>
24struct PMFloatGetSetBench : public Benchmark {
25    PMFloatGetSetBench() {}
26
27    const char* onGetName() override {
28        switch (kClamp << 1 | kWide) {
29            case 0: return "SkPMFloat_get_1x";
30            case 1: return "SkPMFloat_get_4x";
31            case 2: return "SkPMFloat_clamp_1x";
32            case 3: return "SkPMFloat_clamp_4x";
33        }
34        SkFAIL("unreachable");
35        return "oh bother";
36    }
37    bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
38
39    void onDraw(const int loops, SkCanvas* canvas) override {
40        // Unlike blackhole, junk can and probably will be a register.
41        uint32_t junk = 0;
42        uint32_t seed = 0;
43        for (int i = 0; i < loops; i++) {
44            SkPMColor colors[4];
45        #ifdef SK_DEBUG
46            for (int i = 0; i < 4; i++) {
47                // Our SkASSERTs will remind us that it's technically required that we premultiply.
48                colors[i] = SkPreMultiplyColor(lcg_rand(&seed));
49            }
50        #else
51            // But it's a lot faster not to, and this code won't really mind the non-PM colors.
52            (void)lcg_rand(&seed);
53            colors[0] = seed + 0;
54            colors[1] = seed + 1;
55            colors[2] = seed + 2;
56            colors[3] = seed + 3;
57        #endif
58
59            SkPMFloat fa,fb,fc,fd;
60            if (kWide) {
61                SkPMFloat::From4PMColors(colors, &fa, &fb, &fc, &fd);
62            } else {
63                fa = SkPMFloat::FromPMColor(colors[0]);
64                fb = SkPMFloat::FromPMColor(colors[1]);
65                fc = SkPMFloat::FromPMColor(colors[2]);
66                fd = SkPMFloat::FromPMColor(colors[3]);
67            }
68
69            SkPMColor back[4];
70            switch (kClamp << 1 | kWide) {
71                case 0: {
72                    back[0] = fa.round();
73                    back[1] = fb.round();
74                    back[2] = fc.round();
75                    back[3] = fd.round();
76                } break;
77                case 1: SkPMFloat::RoundTo4PMColors(fa, fb, fc, fd, back); break;
78                case 2: {
79                    back[0] = fa.roundClamp();
80                    back[1] = fb.roundClamp();
81                    back[2] = fc.roundClamp();
82                    back[3] = fd.roundClamp();
83                } break;
84                case 3: SkPMFloat::RoundClampTo4PMColors(fa, fb, fc, fd, back); break;
85            }
86            for (int i = 0; i < 4; i++) {
87                junk ^= back[i];
88            }
89        }
90        blackhole ^= junk;
91    }
92};
93
94// Extra () help DEF_BENCH not get confused by the comma inside the <>.
95DEF_BENCH(return (new PMFloatGetSetBench< true,  true>);)
96DEF_BENCH(return (new PMFloatGetSetBench<false,  true>);)
97DEF_BENCH(return (new PMFloatGetSetBench< true, false>);)
98DEF_BENCH(return (new PMFloatGetSetBench<false, false>);)
99
100struct PMFloatGradientBench : public Benchmark {
101    const char* onGetName() override { return "PMFloat_gradient"; }
102    bool isSuitableFor(Backend backend) override { return backend == kNonRendering_Backend; }
103
104    SkPMColor fDevice[100];
105    void onDraw(const int loops, SkCanvas*) override {
106        Sk4f c0 = SkPMFloat::FromARGB(255, 255, 0, 0),
107             c1 = SkPMFloat::FromARGB(255, 0, 0, 255),
108             dc = c1 - c0,
109             fx(0.1f),
110             dx(0.002f),
111             dcdx(dc*dx),
112             dcdx4(dcdx+dcdx+dcdx+dcdx);
113
114        for (int n = 0; n < loops; n++) {
115            Sk4f a = c0 + dc*fx + Sk4f(0.5f),  // The +0.5f lets us call trunc() instead of get().
116                 b = a + dcdx,
117                 c = b + dcdx,
118                 d = c + dcdx;
119            for (size_t i = 0; i < SK_ARRAY_COUNT(fDevice); i += 4) {
120                fDevice[i+0] = SkPMFloat(a).trunc();
121                fDevice[i+1] = SkPMFloat(b).trunc();
122                fDevice[i+2] = SkPMFloat(c).trunc();
123                fDevice[i+3] = SkPMFloat(d).trunc();
124                a += dcdx4;
125                b += dcdx4;
126                c += dcdx4;
127                d += dcdx4;
128            }
129        }
130    }
131};
132
133DEF_BENCH(return new PMFloatGradientBench;)
134