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 "Sk4px.h"
9#include "SkNx.h"
10#include "SkRandom.h"
11#include "Test.h"
12
13template <int N>
14static void test_Nf(skiatest::Reporter* r) {
15
16    auto assert_nearly_eq = [&](float eps, const SkNx<N, float>& v,
17                                float a, float b, float c, float d) {
18        auto close = [=](float a, float b) { return fabsf(a-b) <= eps; };
19        float vals[4];
20        v.store(vals);
21        bool ok = close(vals[0], a) && close(vals[1], b)
22               && close(   v[0], a) && close(   v[1], b);
23        REPORTER_ASSERT(r, ok);
24        if (N == 4) {
25            ok = close(vals[2], c) && close(vals[3], d)
26              && close(   v[2], c) && close(   v[3], d);
27            REPORTER_ASSERT(r, ok);
28        }
29    };
30    auto assert_eq = [&](const SkNx<N, float>& v, float a, float b, float c, float d) {
31        return assert_nearly_eq(0, v, a,b,c,d);
32    };
33
34    float vals[] = {3, 4, 5, 6};
35    SkNx<N,float> a = SkNx<N,float>::Load(vals),
36                  b(a),
37                  c = a;
38    SkNx<N,float> d;
39    d = a;
40
41    assert_eq(a, 3, 4, 5, 6);
42    assert_eq(b, 3, 4, 5, 6);
43    assert_eq(c, 3, 4, 5, 6);
44    assert_eq(d, 3, 4, 5, 6);
45
46    assert_eq(a+b, 6, 8, 10, 12);
47    assert_eq(a*b, 9, 16, 25, 36);
48    assert_eq(a*b-b, 6, 12, 20, 30);
49    assert_eq((a*b).sqrt(), 3, 4, 5, 6);
50    assert_eq(a/b, 1, 1, 1, 1);
51    assert_eq(SkNx<N,float>(0)-a, -3, -4, -5, -6);
52
53    SkNx<N,float> fours(4);
54
55    assert_eq(fours.sqrt(), 2,2,2,2);
56    assert_nearly_eq(0.001f, fours.rsqrt(), 0.5, 0.5, 0.5, 0.5);
57
58    assert_nearly_eq(0.001f, fours.invert(), 0.25, 0.25, 0.25, 0.25);
59
60    assert_eq(SkNx<N,float>::Min(a, fours), 3, 4, 4, 4);
61    assert_eq(SkNx<N,float>::Max(a, fours), 4, 4, 5, 6);
62
63    // Test some comparisons.  This is not exhaustive.
64    REPORTER_ASSERT(r, (a == b).allTrue());
65    REPORTER_ASSERT(r, (a+b == a*b-b).anyTrue());
66    REPORTER_ASSERT(r, !(a+b == a*b-b).allTrue());
67    REPORTER_ASSERT(r, !(a+b == a*b).anyTrue());
68    REPORTER_ASSERT(r, !(a != b).anyTrue());
69    REPORTER_ASSERT(r, (a < fours).anyTrue());
70    REPORTER_ASSERT(r, (a <= fours).anyTrue());
71    REPORTER_ASSERT(r, !(a > fours).allTrue());
72    REPORTER_ASSERT(r, !(a >= fours).allTrue());
73}
74
75DEF_TEST(SkNf, r) {
76    test_Nf<2>(r);
77    test_Nf<4>(r);
78}
79
80template <int N, typename T>
81void test_Ni(skiatest::Reporter* r) {
82    auto assert_eq = [&](const SkNx<N,T>& v, T a, T b, T c, T d, T e, T f, T g, T h) {
83        T vals[8];
84        v.store(vals);
85
86        switch (N) {
87          case 8: REPORTER_ASSERT(r, vals[4] == e && vals[5] == f && vals[6] == g && vals[7] == h);
88          case 4: REPORTER_ASSERT(r, vals[2] == c && vals[3] == d);
89          case 2: REPORTER_ASSERT(r, vals[0] == a && vals[1] == b);
90        }
91        switch (N) {
92          case 8: REPORTER_ASSERT(r, v[4] == e && v[5] == f &&
93                                     v[6] == g && v[7] == h);
94          case 4: REPORTER_ASSERT(r, v[2] == c && v[3] == d);
95          case 2: REPORTER_ASSERT(r, v[0] == a && v[1] == b);
96        }
97    };
98
99    T vals[] = { 1,2,3,4,5,6,7,8 };
100    SkNx<N,T> a = SkNx<N,T>::Load(vals),
101              b(a),
102              c = a;
103    SkNx<N,T> d;
104    d = a;
105
106    assert_eq(a, 1,2,3,4,5,6,7,8);
107    assert_eq(b, 1,2,3,4,5,6,7,8);
108    assert_eq(c, 1,2,3,4,5,6,7,8);
109    assert_eq(d, 1,2,3,4,5,6,7,8);
110
111    assert_eq(a+a, 2,4,6,8,10,12,14,16);
112    assert_eq(a*a, 1,4,9,16,25,36,49,64);
113    assert_eq(a*a-a, 0,2,6,12,20,30,42,56);
114
115    assert_eq(a >> 2, 0,0,0,1,1,1,1,2);
116    assert_eq(a << 1, 2,4,6,8,10,12,14,16);
117
118    REPORTER_ASSERT(r, a[1] == 2);
119}
120
121DEF_TEST(SkNx, r) {
122    test_Ni<2, uint16_t>(r);
123    test_Ni<4, uint16_t>(r);
124    test_Ni<8, uint16_t>(r);
125
126    test_Ni<2, int>(r);
127    test_Ni<4, int>(r);
128    test_Ni<8, int>(r);
129}
130
131DEF_TEST(SkNi_min_lt, r) {
132    // Exhaustively check the 8x8 bit space.
133    for (int a = 0; a < (1<<8); a++) {
134    for (int b = 0; b < (1<<8); b++) {
135        Sk16b aw(a), bw(b);
136        REPORTER_ASSERT(r, Sk16b::Min(aw, bw)[0] == SkTMin(a, b));
137        REPORTER_ASSERT(r, !(aw < bw)[0] == !(a < b));
138    }}
139
140    // Exhausting the 16x16 bit space is kind of slow, so only do that in release builds.
141#ifdef SK_DEBUG
142    SkRandom rand;
143    for (int i = 0; i < (1<<16); i++) {
144        uint16_t a = rand.nextU() >> 16,
145                 b = rand.nextU() >> 16;
146        REPORTER_ASSERT(r, Sk16h::Min(Sk16h(a), Sk16h(b))[0] == SkTMin(a, b));
147    }
148#else
149    for (int a = 0; a < (1<<16); a++) {
150    for (int b = 0; b < (1<<16); b++) {
151        REPORTER_ASSERT(r, Sk16h::Min(Sk16h(a), Sk16h(b))[0] == SkTMin(a, b));
152    }}
153#endif
154}
155
156DEF_TEST(SkNi_saturatedAdd, r) {
157    for (int a = 0; a < (1<<8); a++) {
158    for (int b = 0; b < (1<<8); b++) {
159        int exact = a+b;
160        if (exact > 255) { exact = 255; }
161        if (exact <   0) { exact =   0; }
162
163        REPORTER_ASSERT(r, Sk16b(a).saturatedAdd(Sk16b(b))[0] == exact);
164    }
165    }
166}
167
168DEF_TEST(Sk4px_muldiv255round, r) {
169    for (int a = 0; a < (1<<8); a++) {
170    for (int b = 0; b < (1<<8); b++) {
171        int exact = (a*b+127)/255;
172
173        // Duplicate a and b 16x each.
174        auto av = Sk4px::DupAlpha(a),
175             bv = Sk4px::DupAlpha(b);
176
177        // This way should always be exactly correct.
178        int correct = (av * bv).div255()[0];
179        REPORTER_ASSERT(r, correct == exact);
180
181        // We're a bit more flexible on this method: correct for 0 or 255, otherwise off by <=1.
182        int fast = av.approxMulDiv255(bv)[0];
183        REPORTER_ASSERT(r, fast-exact >= -1 && fast-exact <= 1);
184        if (a == 0 || a == 255 || b == 0 || b == 255) {
185            REPORTER_ASSERT(r, fast == exact);
186        }
187    }
188    }
189}
190
191DEF_TEST(Sk4px_widening, r) {
192    SkPMColor colors[] = {
193        SkPreMultiplyColor(0xff00ff00),
194        SkPreMultiplyColor(0x40008000),
195        SkPreMultiplyColor(0x7f020406),
196        SkPreMultiplyColor(0x00000000),
197    };
198    auto packed = Sk4px::Load4(colors);
199
200    auto wideLo = packed.widenLo(),
201         wideHi = packed.widenHi(),
202         wideLoHi    = packed.widenLoHi(),
203         wideLoHiAlt = wideLo + wideHi;
204    REPORTER_ASSERT(r, 0 == memcmp(&wideLoHi, &wideLoHiAlt, sizeof(wideLoHi)));
205}
206
207DEF_TEST(SkNx_abs, r) {
208    auto fs = Sk4f(0.0f, -0.0f, 2.0f, -4.0f).abs();
209    REPORTER_ASSERT(r, fs[0] == 0.0f);
210    REPORTER_ASSERT(r, fs[1] == 0.0f);
211    REPORTER_ASSERT(r, fs[2] == 2.0f);
212    REPORTER_ASSERT(r, fs[3] == 4.0f);
213}
214
215DEF_TEST(Sk4i_abs, r) {
216    auto is = Sk4i(0, -1, 2, -2147483647).abs();
217    REPORTER_ASSERT(r, is[0] == 0);
218    REPORTER_ASSERT(r, is[1] == 1);
219    REPORTER_ASSERT(r, is[2] == 2);
220    REPORTER_ASSERT(r, is[3] == 2147483647);
221}
222
223DEF_TEST(Sk4i_minmax, r) {
224    auto a = Sk4i(0, 2, 4, 6);
225    auto b = Sk4i(1, 1, 3, 7);
226    auto min = Sk4i::Min(a, b);
227    auto max = Sk4i::Max(a, b);
228    for(int i = 0; i < 4; ++i) {
229        REPORTER_ASSERT(r, min[i] == SkTMin(a[i], b[i]));
230        REPORTER_ASSERT(r, max[i] == SkTMax(a[i], b[i]));
231    }
232}
233
234DEF_TEST(SkNx_floor, r) {
235    auto fs = Sk4f(0.4f, -0.4f, 0.6f, -0.6f).floor();
236    REPORTER_ASSERT(r, fs[0] ==  0.0f);
237    REPORTER_ASSERT(r, fs[1] == -1.0f);
238    REPORTER_ASSERT(r, fs[2] ==  0.0f);
239    REPORTER_ASSERT(r, fs[3] == -1.0f);
240}
241
242DEF_TEST(SkNx_shuffle, r) {
243    Sk4f f4(0,10,20,30);
244
245    Sk2f f2 = SkNx_shuffle<2,1>(f4);
246    REPORTER_ASSERT(r, f2[0] == 20);
247    REPORTER_ASSERT(r, f2[1] == 10);
248
249    f4 = SkNx_shuffle<0,1,1,0>(f2);
250    REPORTER_ASSERT(r, f4[0] == 20);
251    REPORTER_ASSERT(r, f4[1] == 10);
252    REPORTER_ASSERT(r, f4[2] == 10);
253    REPORTER_ASSERT(r, f4[3] == 20);
254}
255
256DEF_TEST(SkNx_int_float, r) {
257    Sk4f f(-2.3f, 1.0f, 0.45f, 0.6f);
258
259    Sk4i i = SkNx_cast<int>(f);
260    REPORTER_ASSERT(r, i[0] == -2);
261    REPORTER_ASSERT(r, i[1] ==  1);
262    REPORTER_ASSERT(r, i[2] ==  0);
263    REPORTER_ASSERT(r, i[3] ==  0);
264
265    f = SkNx_cast<float>(i);
266    REPORTER_ASSERT(r, f[0] == -2.0f);
267    REPORTER_ASSERT(r, f[1] ==  1.0f);
268    REPORTER_ASSERT(r, f[2] ==  0.0f);
269    REPORTER_ASSERT(r, f[3] ==  0.0f);
270}
271
272#include "SkRandom.h"
273
274DEF_TEST(SkNx_u16_float, r) {
275    {
276        // u16 --> float
277        auto h4 = Sk4h(15, 17, 257, 65535);
278        auto f4 = SkNx_cast<float>(h4);
279        REPORTER_ASSERT(r, f4[0] == 15.0f);
280        REPORTER_ASSERT(r, f4[1] == 17.0f);
281        REPORTER_ASSERT(r, f4[2] == 257.0f);
282        REPORTER_ASSERT(r, f4[3] == 65535.0f);
283    }
284    {
285        // float -> u16
286        auto f4 = Sk4f(15, 17, 257, 65535);
287        auto h4 = SkNx_cast<uint16_t>(f4);
288        REPORTER_ASSERT(r, h4[0] == 15);
289        REPORTER_ASSERT(r, h4[1] == 17);
290        REPORTER_ASSERT(r, h4[2] == 257);
291        REPORTER_ASSERT(r, h4[3] == 65535);
292    }
293
294    // starting with any u16 value, we should be able to have a perfect round-trip in/out of floats
295    //
296    SkRandom rand;
297    for (int i = 0; i < 10000; ++i) {
298        const uint16_t s16[4] {
299            (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
300            (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
301        };
302        auto u4_0 = Sk4h::Load(s16);
303        auto f4 = SkNx_cast<float>(u4_0);
304        auto u4_1 = SkNx_cast<uint16_t>(f4);
305        uint16_t d16[4];
306        u4_1.store(d16);
307        REPORTER_ASSERT(r, !memcmp(s16, d16, sizeof(s16)));
308    }
309}
310
311// The SSE2 implementation of SkNx_cast<uint16_t>(Sk4i) is non-trivial, so worth a test.
312DEF_TEST(SkNx_int_u16, r) {
313    // These are pretty hard to get wrong.
314    for (int i = 0; i <= 0x7fff; i++) {
315        uint16_t expected = (uint16_t)i;
316        uint16_t actual = SkNx_cast<uint16_t>(Sk4i(i))[0];
317
318        REPORTER_ASSERT(r, expected == actual);
319    }
320
321    // A naive implementation with _mm_packs_epi32 would succeed up to 0x7fff but fail here:
322    for (int i = 0x8000; (1) && i <= 0xffff; i++) {
323        uint16_t expected = (uint16_t)i;
324        uint16_t actual = SkNx_cast<uint16_t>(Sk4i(i))[0];
325
326        REPORTER_ASSERT(r, expected == actual);
327    }
328}
329
330DEF_TEST(SkNx_4fLoad4Store4, r) {
331    float src[] = {
332         0.0f,  1.0f,  2.0f,  3.0f,
333         4.0f,  5.0f,  6.0f,  7.0f,
334         8.0f,  9.0f, 10.0f, 11.0f,
335        12.0f, 13.0f, 14.0f, 15.0f
336    };
337
338    Sk4f a, b, c, d;
339    Sk4f::Load4(src, &a, &b, &c, &d);
340    REPORTER_ASSERT(r,  0.0f == a[0]);
341    REPORTER_ASSERT(r,  4.0f == a[1]);
342    REPORTER_ASSERT(r,  8.0f == a[2]);
343    REPORTER_ASSERT(r, 12.0f == a[3]);
344    REPORTER_ASSERT(r,  1.0f == b[0]);
345    REPORTER_ASSERT(r,  5.0f == b[1]);
346    REPORTER_ASSERT(r,  9.0f == b[2]);
347    REPORTER_ASSERT(r, 13.0f == b[3]);
348    REPORTER_ASSERT(r,  2.0f == c[0]);
349    REPORTER_ASSERT(r,  6.0f == c[1]);
350    REPORTER_ASSERT(r, 10.0f == c[2]);
351    REPORTER_ASSERT(r, 14.0f == c[3]);
352    REPORTER_ASSERT(r,  3.0f == d[0]);
353    REPORTER_ASSERT(r,  7.0f == d[1]);
354    REPORTER_ASSERT(r, 11.0f == d[2]);
355    REPORTER_ASSERT(r, 15.0f == d[3]);
356
357    float dst[16];
358    Sk4f::Store4(dst, a, b, c, d);
359    REPORTER_ASSERT(r, 0 == memcmp(dst, src, 16 * sizeof(float)));
360}
361