blob: ce7b5bc48f822f0f8c8bfbf3f381cee135e2012f [file] [log] [blame]
mtkleinc9adb052015-03-30 10:50:27 -07001/*
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
mtklein37478752015-06-15 10:58:42 -07008#include "Sk4px.h"
mtkleinc9adb052015-03-30 10:50:27 -07009#include "SkNx.h"
mtklein27e517a2015-05-14 17:53:04 -070010#include "SkRandom.h"
mtkleinc9adb052015-03-30 10:50:27 -070011#include "Test.h"
12
mtklein6f797092015-11-09 08:33:53 -080013template <int N>
mtkleinc9adb052015-03-30 10:50:27 -070014static void test_Nf(skiatest::Reporter* r) {
15
mtklein6c221b42015-11-20 13:53:19 -080016 auto assert_nearly_eq = [&](float eps, const SkNx<N, float>& v,
17 float a, float b, float c, float d) {
mtklein6f797092015-11-09 08:33:53 -080018 auto close = [=](float a, float b) { return fabsf(a-b) <= eps; };
19 float vals[4];
mtkleinc9adb052015-03-30 10:50:27 -070020 v.store(vals);
mtkleina156a8f2015-04-03 06:16:13 -070021 bool ok = close(vals[0], a) && close(vals[1], b)
mtklein7c249e52016-02-21 10:54:19 -080022 && close( v[0], a) && close( v[1], b);
mtkleina156a8f2015-04-03 06:16:13 -070023 REPORTER_ASSERT(r, ok);
mtkleinc9adb052015-03-30 10:50:27 -070024 if (N == 4) {
mtkleina156a8f2015-04-03 06:16:13 -070025 ok = close(vals[2], c) && close(vals[3], d)
mtklein7c249e52016-02-21 10:54:19 -080026 && close( v[2], c) && close( v[3], d);
mtkleina156a8f2015-04-03 06:16:13 -070027 REPORTER_ASSERT(r, ok);
mtkleinc9adb052015-03-30 10:50:27 -070028 }
29 };
mtklein6c221b42015-11-20 13:53:19 -080030 auto assert_eq = [&](const SkNx<N, float>& v, float a, float b, float c, float d) {
mtkleinc9adb052015-03-30 10:50:27 -070031 return assert_nearly_eq(0, v, a,b,c,d);
32 };
33
mtklein6f797092015-11-09 08:33:53 -080034 float vals[] = {3, 4, 5, 6};
mtklein6c221b42015-11-20 13:53:19 -080035 SkNx<N,float> a = SkNx<N,float>::Load(vals),
36 b(a),
37 c = a;
38 SkNx<N,float> d;
mtkleinc9adb052015-03-30 10:50:27 -070039 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);
mtklein6c221b42015-11-20 13:53:19 -080051 assert_eq(SkNx<N,float>(0)-a, -3, -4, -5, -6);
mtkleinc9adb052015-03-30 10:50:27 -070052
mtklein6c221b42015-11-20 13:53:19 -080053 SkNx<N,float> fours(4);
mtkleinc9adb052015-03-30 10:50:27 -070054
55 assert_eq(fours.sqrt(), 2,2,2,2);
mtkleinf8f90e42016-03-21 10:04:46 -070056 assert_nearly_eq(0.001f, fours.rsqrt(), 0.5, 0.5, 0.5, 0.5);
mtkleinc9adb052015-03-30 10:50:27 -070057
mtkleinf8f90e42016-03-21 10:04:46 -070058 assert_nearly_eq(0.001f, fours.invert(), 0.25, 0.25, 0.25, 0.25);
mtkleinc9adb052015-03-30 10:50:27 -070059
mtklein6c221b42015-11-20 13:53:19 -080060 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);
mtkleinc9adb052015-03-30 10:50:27 -070062
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) {
mtklein6f797092015-11-09 08:33:53 -080076 test_Nf<2>(r);
77 test_Nf<4>(r);
mtkleinc9adb052015-03-30 10:50:27 -070078}
mtklein115acee2015-04-14 14:02:52 -070079
80template <int N, typename T>
81void test_Ni(skiatest::Reporter* r) {
mtklein6c221b42015-11-20 13:53:19 -080082 auto assert_eq = [&](const SkNx<N,T>& v, T a, T b, T c, T d, T e, T f, T g, T h) {
mtklein115acee2015-04-14 14:02:52 -070083 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 }
mtklein1113da72015-04-27 12:08:01 -070091 switch (N) {
mtklein7c249e52016-02-21 10:54:19 -080092 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);
mtklein1113da72015-04-27 12:08:01 -070096 }
mtklein115acee2015-04-14 14:02:52 -070097 };
98
99 T vals[] = { 1,2,3,4,5,6,7,8 };
mtklein6c221b42015-11-20 13:53:19 -0800100 SkNx<N,T> a = SkNx<N,T>::Load(vals),
mtklein115acee2015-04-14 14:02:52 -0700101 b(a),
102 c = a;
mtklein6c221b42015-11-20 13:53:19 -0800103 SkNx<N,T> d;
mtklein115acee2015-04-14 14:02:52 -0700104 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
mtklein7c249e52016-02-21 10:54:19 -0800118 REPORTER_ASSERT(r, a[1] == 2);
mtklein115acee2015-04-14 14:02:52 -0700119}
120
mtklein6c221b42015-11-20 13:53:19 -0800121DEF_TEST(SkNx, r) {
mtklein115acee2015-04-14 14:02:52 -0700122 test_Ni<2, uint16_t>(r);
123 test_Ni<4, uint16_t>(r);
124 test_Ni<8, uint16_t>(r);
mtklein1113da72015-04-27 12:08:01 -0700125
126 test_Ni<2, int>(r);
127 test_Ni<4, int>(r);
128 test_Ni<8, int>(r);
mtklein115acee2015-04-14 14:02:52 -0700129}
mtklein27e517a2015-05-14 17:53:04 -0700130
mtkleine20633e2015-07-13 12:06:33 -0700131DEF_TEST(SkNi_min_lt, r) {
mtklein27e517a2015-05-14 17:53:04 -0700132 // Exhaustively check the 8x8 bit space.
133 for (int a = 0; a < (1<<8); a++) {
134 for (int b = 0; b < (1<<8); b++) {
mtkleine20633e2015-07-13 12:06:33 -0700135 Sk16b aw(a), bw(b);
mtklein7c249e52016-02-21 10:54:19 -0800136 REPORTER_ASSERT(r, Sk16b::Min(aw, bw)[0] == SkTMin(a, b));
137 REPORTER_ASSERT(r, !(aw < bw)[0] == !(a < b));
mtklein27e517a2015-05-14 17:53:04 -0700138 }}
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;
mtklein7c249e52016-02-21 10:54:19 -0800146 REPORTER_ASSERT(r, Sk16h::Min(Sk16h(a), Sk16h(b))[0] == SkTMin(a, b));
mtklein27e517a2015-05-14 17:53:04 -0700147 }
148#else
149 for (int a = 0; a < (1<<16); a++) {
150 for (int b = 0; b < (1<<16); b++) {
mtklein7c249e52016-02-21 10:54:19 -0800151 REPORTER_ASSERT(r, Sk16h::Min(Sk16h(a), Sk16h(b))[0] == SkTMin(a, b));
mtklein27e517a2015-05-14 17:53:04 -0700152 }}
153#endif
154}
mtklein37478752015-06-15 10:58:42 -0700155
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
mtklein7c249e52016-02-21 10:54:19 -0800163 REPORTER_ASSERT(r, Sk16b(a).saturatedAdd(Sk16b(b))[0] == exact);
mtklein37478752015-06-15 10:58:42 -0700164 }
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.
mtklein059ac002015-06-22 10:39:38 -0700174 auto av = Sk4px::DupAlpha(a),
175 bv = Sk4px::DupAlpha(b);
mtklein37478752015-06-15 10:58:42 -0700176
177 // This way should always be exactly correct.
mtklein7c249e52016-02-21 10:54:19 -0800178 int correct = (av * bv).div255()[0];
mtklein37478752015-06-15 10:58:42 -0700179 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.
mtklein7c249e52016-02-21 10:54:19 -0800182 int fast = av.approxMulDiv255(bv)[0];
mtklein37478752015-06-15 10:58:42 -0700183 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}
mtklein4be181e2015-07-14 10:54:19 -0700190
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}
mtkleina508f3c2015-09-01 06:29:45 -0700206
mtkleinc33065a2016-01-15 12:16:40 -0800207DEF_TEST(SkNx_abs, r) {
208 auto fs = Sk4f(0.0f, -0.0f, 2.0f, -4.0f).abs();
mtklein7c249e52016-02-21 10:54:19 -0800209 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);
mtkleinc33065a2016-01-15 12:16:40 -0800213}
mtklein629f25a2016-02-08 05:54:38 -0800214
mtklein126626e2016-02-09 15:41:36 -0800215DEF_TEST(SkNx_floor, r) {
216 auto fs = Sk4f(0.4f, -0.4f, 0.6f, -0.6f).floor();
mtklein7c249e52016-02-21 10:54:19 -0800217 REPORTER_ASSERT(r, fs[0] == 0.0f);
218 REPORTER_ASSERT(r, fs[1] == -1.0f);
219 REPORTER_ASSERT(r, fs[2] == 0.0f);
220 REPORTER_ASSERT(r, fs[3] == -1.0f);
mtklein126626e2016-02-09 15:41:36 -0800221}
222
mtkleine4c0bee2016-02-09 10:35:27 -0800223DEF_TEST(SkNx_shuffle, r) {
224 Sk4f f4(0,10,20,30);
mtklein629f25a2016-02-08 05:54:38 -0800225
mtkleine4c0bee2016-02-09 10:35:27 -0800226 Sk2f f2 = SkNx_shuffle<2,1>(f4);
227 REPORTER_ASSERT(r, f2[0] == 20);
228 REPORTER_ASSERT(r, f2[1] == 10);
229
230 f4 = SkNx_shuffle<0,1,1,0>(f2);
231 REPORTER_ASSERT(r, f4[0] == 20);
232 REPORTER_ASSERT(r, f4[1] == 10);
233 REPORTER_ASSERT(r, f4[2] == 10);
234 REPORTER_ASSERT(r, f4[3] == 20);
mtklein629f25a2016-02-08 05:54:38 -0800235}
236
mtklein0cf795f2016-02-17 07:23:36 -0800237DEF_TEST(SkNx_int_float, r) {
238 Sk4f f(-2.3f, 1.0f, 0.45f, 0.6f);
239
240 Sk4i i = SkNx_cast<int>(f);
241 REPORTER_ASSERT(r, i[0] == -2);
242 REPORTER_ASSERT(r, i[1] == 1);
243 REPORTER_ASSERT(r, i[2] == 0);
244 REPORTER_ASSERT(r, i[3] == 0);
245
246 f = SkNx_cast<float>(i);
247 REPORTER_ASSERT(r, f[0] == -2.0f);
248 REPORTER_ASSERT(r, f[1] == 1.0f);
249 REPORTER_ASSERT(r, f[2] == 0.0f);
250 REPORTER_ASSERT(r, f[3] == 0.0f);
251}
252
mtkleine4c0bee2016-02-09 10:35:27 -0800253#include "SkRandom.h"
254
mtklein629f25a2016-02-08 05:54:38 -0800255DEF_TEST(SkNx_u16_float, r) {
256 {
257 // u16 --> float
258 auto h4 = Sk4h(15, 17, 257, 65535);
259 auto f4 = SkNx_cast<float>(h4);
mtklein7c249e52016-02-21 10:54:19 -0800260 REPORTER_ASSERT(r, f4[0] == 15.0f);
261 REPORTER_ASSERT(r, f4[1] == 17.0f);
262 REPORTER_ASSERT(r, f4[2] == 257.0f);
263 REPORTER_ASSERT(r, f4[3] == 65535.0f);
mtklein629f25a2016-02-08 05:54:38 -0800264 }
265 {
266 // float -> u16
267 auto f4 = Sk4f(15, 17, 257, 65535);
268 auto h4 = SkNx_cast<uint16_t>(f4);
mtklein7c249e52016-02-21 10:54:19 -0800269 REPORTER_ASSERT(r, h4[0] == 15);
270 REPORTER_ASSERT(r, h4[1] == 17);
271 REPORTER_ASSERT(r, h4[2] == 257);
272 REPORTER_ASSERT(r, h4[3] == 65535);
mtklein629f25a2016-02-08 05:54:38 -0800273 }
274
275 // starting with any u16 value, we should be able to have a perfect round-trip in/out of floats
276 //
277 SkRandom rand;
mtkleine4c0bee2016-02-09 10:35:27 -0800278 for (int i = 0; i < 10000; ++i) {
mtklein629f25a2016-02-08 05:54:38 -0800279 const uint16_t s16[4] {
280 (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
281 (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
282 };
283 auto u4_0 = Sk4h::Load(s16);
284 auto f4 = SkNx_cast<float>(u4_0);
285 auto u4_1 = SkNx_cast<uint16_t>(f4);
286 uint16_t d16[4];
287 u4_1.store(d16);
288 REPORTER_ASSERT(r, !memcmp(s16, d16, sizeof(s16)));
289 }
290}
mtklein58e389b2016-07-15 07:00:11 -0700291
292// The SSE2 implementation of SkNx_cast<uint16_t>(Sk4i) is non-trivial, so worth a test.
293DEF_TEST(SkNx_int_u16, r) {
294 // These are pretty hard to get wrong.
295 for (int i = 0; i <= 0x7fff; i++) {
296 uint16_t expected = (uint16_t)i;
297 uint16_t actual = SkNx_cast<uint16_t>(Sk4i(i))[0];
298
299 REPORTER_ASSERT(r, expected == actual);
300 }
301
302 // A naive implementation with _mm_packs_epi32 would succeed up to 0x7fff but fail here:
303 for (int i = 0x8000; (1) && i <= 0xffff; i++) {
304 uint16_t expected = (uint16_t)i;
305 uint16_t actual = SkNx_cast<uint16_t>(Sk4i(i))[0];
306
307 REPORTER_ASSERT(r, expected == actual);
308 }
309}
msarettc0444612016-09-16 11:45:58 -0700310
311DEF_TEST(SkNx_4fLoad4Store4, r) {
312 float src[] = {
313 0.0f, 1.0f, 2.0f, 3.0f,
314 4.0f, 5.0f, 6.0f, 7.0f,
315 8.0f, 9.0f, 10.0f, 11.0f,
316 12.0f, 13.0f, 14.0f, 15.0f
317 };
318
319 Sk4f a, b, c, d;
Mike Klein33cbfd72016-10-06 11:09:27 -0400320 Sk4f::Load4(src, &a, &b, &c, &d);
msarettc0444612016-09-16 11:45:58 -0700321 REPORTER_ASSERT(r, 0.0f == a[0]);
322 REPORTER_ASSERT(r, 4.0f == a[1]);
323 REPORTER_ASSERT(r, 8.0f == a[2]);
324 REPORTER_ASSERT(r, 12.0f == a[3]);
325 REPORTER_ASSERT(r, 1.0f == b[0]);
326 REPORTER_ASSERT(r, 5.0f == b[1]);
327 REPORTER_ASSERT(r, 9.0f == b[2]);
328 REPORTER_ASSERT(r, 13.0f == b[3]);
329 REPORTER_ASSERT(r, 2.0f == c[0]);
330 REPORTER_ASSERT(r, 6.0f == c[1]);
331 REPORTER_ASSERT(r, 10.0f == c[2]);
332 REPORTER_ASSERT(r, 14.0f == c[3]);
333 REPORTER_ASSERT(r, 3.0f == d[0]);
334 REPORTER_ASSERT(r, 7.0f == d[1]);
335 REPORTER_ASSERT(r, 11.0f == d[2]);
336 REPORTER_ASSERT(r, 15.0f == d[3]);
337
338 float dst[16];
Mike Klein33cbfd72016-10-06 11:09:27 -0400339 Sk4f::Store4(dst, a, b, c, d);
msarettc0444612016-09-16 11:45:58 -0700340 REPORTER_ASSERT(r, 0 == memcmp(dst, src, 16 * sizeof(float)));
341}