1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <array>
13 #include <cstdlib>
14 #include <iostream>
15 #include <string>
16 #include <tuple>
17
18 #include "third_party/googletest/src/googletest/include/gtest/gtest.h"
19
20 #include "config/aom_config.h"
21 #include "config/av1_rtcd.h"
22
23 #include "aom_ports/aom_timer.h"
24 #include "av1/common/cdef_block.h"
25 #include "test/acm_random.h"
26 #include "test/register_state_check.h"
27 #include "test/util.h"
28
29 using libaom_test::ACMRandom;
30
31 namespace {
32
33 using CdefFilterBlockFunctions = std::array<cdef_filter_block_func, 4>;
34
35 typedef std::tuple<CdefFilterBlockFunctions, CdefFilterBlockFunctions,
36 BLOCK_SIZE, int, int>
37 cdef_dir_param_t;
38
39 class CDEFBlockTest : public ::testing::TestWithParam<cdef_dir_param_t> {
40 public:
41 ~CDEFBlockTest() override = default;
SetUp()42 void SetUp() override {
43 cdef = GET_PARAM(0);
44 ref_cdef = GET_PARAM(1);
45 bsize = GET_PARAM(2);
46 boundary = GET_PARAM(3);
47 depth = GET_PARAM(4);
48 }
49
50 protected:
51 BLOCK_SIZE bsize;
52 int boundary;
53 int depth;
54 CdefFilterBlockFunctions cdef;
55 CdefFilterBlockFunctions ref_cdef;
56 };
57 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFBlockTest);
58
59 typedef CDEFBlockTest CDEFBlockHighbdTest;
60 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFBlockHighbdTest);
61
62 typedef CDEFBlockTest CDEFSpeedTest;
63 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFSpeedTest);
64
65 typedef CDEFBlockTest CDEFSpeedHighbdTest;
66 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFSpeedHighbdTest);
67
test_cdef(BLOCK_SIZE bsize,int iterations,CdefFilterBlockFunctions cdef,CdefFilterBlockFunctions ref_cdef,int boundary,int depth)68 int64_t test_cdef(BLOCK_SIZE bsize, int iterations,
69 CdefFilterBlockFunctions cdef,
70 CdefFilterBlockFunctions ref_cdef, int boundary, int depth) {
71 aom_usec_timer ref_timer;
72 int64_t ref_elapsed_time = 0;
73 const int size = 8;
74 const int ysize = size + 2 * CDEF_VBORDER;
75 ACMRandom rnd(ACMRandom::DeterministicSeed());
76 DECLARE_ALIGNED(16, uint16_t, s[ysize * CDEF_BSTRIDE]);
77 DECLARE_ALIGNED(16, static uint16_t, d[size * size]);
78 DECLARE_ALIGNED(16, static uint16_t, ref_d[size * size]);
79 memset(ref_d, 0, sizeof(ref_d));
80 memset(d, 0, sizeof(d));
81
82 int error = 0, pristrength = 0, secstrength, dir;
83 int pridamping, secdamping, bits, level, count,
84 errdepth = 0, errpristrength = 0, errsecstrength = 0, errboundary = 0,
85 errpridamping = 0, errsecdamping = 0;
86 unsigned int pos = 0;
87
88 const int block_width =
89 ((bsize == BLOCK_8X8) || (bsize == BLOCK_8X4)) ? 8 : 4;
90 const int block_height =
91 ((bsize == BLOCK_8X8) || (bsize == BLOCK_4X8)) ? 8 : 4;
92 const unsigned int max_pos = size * size >> static_cast<int>(depth == 8);
93 for (pridamping = 3 + depth - 8; pridamping < 7 - 3 * !!boundary + depth - 8;
94 pridamping++) {
95 for (secdamping = 3 + depth - 8;
96 secdamping < 7 - 3 * !!boundary + depth - 8; secdamping++) {
97 for (count = 0; count < iterations; count++) {
98 for (level = 0; level < (1 << depth) && !error;
99 level += (2 + 6 * !!boundary) << (depth - 8)) {
100 for (bits = 1; bits <= depth && !error; bits += 1 + 3 * !!boundary) {
101 for (unsigned int i = 0; i < sizeof(s) / sizeof(*s); i++)
102 s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
103 (1 << depth) - 1);
104 if (boundary) {
105 if (boundary & 1) { // Left
106 for (int i = 0; i < ysize; i++)
107 for (int j = 0; j < CDEF_HBORDER; j++)
108 s[i * CDEF_BSTRIDE + j] = CDEF_VERY_LARGE;
109 }
110 if (boundary & 2) { // Right
111 for (int i = 0; i < ysize; i++)
112 for (int j = CDEF_HBORDER + size; j < CDEF_BSTRIDE; j++)
113 s[i * CDEF_BSTRIDE + j] = CDEF_VERY_LARGE;
114 }
115 if (boundary & 4) { // Above
116 for (int i = 0; i < CDEF_VBORDER; i++)
117 for (int j = 0; j < CDEF_BSTRIDE; j++)
118 s[i * CDEF_BSTRIDE + j] = CDEF_VERY_LARGE;
119 }
120 if (boundary & 8) { // Below
121 for (int i = CDEF_VBORDER + size; i < ysize; i++)
122 for (int j = 0; j < CDEF_BSTRIDE; j++)
123 s[i * CDEF_BSTRIDE + j] = CDEF_VERY_LARGE;
124 }
125 }
126 for (dir = 0; dir < 8; dir++) {
127 for (pristrength = 0; pristrength <= 19 << (depth - 8) && !error;
128 pristrength += (1 + 4 * !!boundary) << (depth - 8)) {
129 if (pristrength == 16) pristrength = 19;
130 for (secstrength = 0; secstrength <= 4 << (depth - 8) && !error;
131 secstrength += 1 << (depth - 8)) {
132 if (secstrength == 3 << (depth - 8)) continue;
133
134 const int strength_index =
135 (secstrength == 0) | ((pristrength == 0) << 1);
136
137 aom_usec_timer_start(&ref_timer);
138 ref_cdef[strength_index](
139 ref_d, size,
140 s + CDEF_HBORDER + CDEF_VBORDER * CDEF_BSTRIDE,
141 pristrength, secstrength, dir, pridamping, secdamping,
142 depth - 8, block_width, block_height);
143 aom_usec_timer_mark(&ref_timer);
144 ref_elapsed_time += aom_usec_timer_elapsed(&ref_timer);
145 // If cdef and ref_cdef are the same, we're just testing
146 // speed
147 if (cdef[0] != ref_cdef[0])
148 API_REGISTER_STATE_CHECK(cdef[strength_index](
149 d, size, s + CDEF_HBORDER + CDEF_VBORDER * CDEF_BSTRIDE,
150 pristrength, secstrength, dir, pridamping, secdamping,
151 depth - 8, block_width, block_height));
152 if (ref_cdef[0] != cdef[0]) {
153 for (pos = 0; pos < max_pos && !error; pos++) {
154 error = ref_d[pos] != d[pos];
155 errdepth = depth;
156 errpristrength = pristrength;
157 errsecstrength = secstrength;
158 errboundary = boundary;
159 errpridamping = pridamping;
160 errsecdamping = secdamping;
161 }
162 }
163 }
164 }
165 }
166 }
167 }
168 }
169 }
170 }
171
172 pos--;
173 EXPECT_EQ(0, error) << "Error: CDEFBlockTest, SIMD and C mismatch."
174 << std::endl
175 << "First error at " << pos % size << "," << pos / size
176 << " (" << (int16_t)ref_d[pos] << " : " << (int16_t)d[pos]
177 << ") " << std::endl
178 << "pristrength: " << errpristrength << std::endl
179 << "pridamping: " << errpridamping << std::endl
180 << "secstrength: " << errsecstrength << std::endl
181 << "secdamping: " << errsecdamping << std::endl
182 << "depth: " << errdepth << std::endl
183 << "size: " << bsize << std::endl
184 << "boundary: " << errboundary << std::endl
185 << std::endl;
186
187 return ref_elapsed_time;
188 }
189
test_cdef_speed(BLOCK_SIZE bsize,int iterations,CdefFilterBlockFunctions cdef,CdefFilterBlockFunctions ref_cdef,int boundary,int depth)190 void test_cdef_speed(BLOCK_SIZE bsize, int iterations,
191 CdefFilterBlockFunctions cdef,
192 CdefFilterBlockFunctions ref_cdef, int boundary,
193 int depth) {
194 int64_t ref_elapsed_time =
195 test_cdef(bsize, iterations, ref_cdef, ref_cdef, boundary, depth);
196
197 int64_t elapsed_time =
198 test_cdef(bsize, iterations, cdef, cdef, boundary, depth);
199
200 std::cout << "C time: " << ref_elapsed_time << " us" << std::endl
201 << "SIMD time: " << elapsed_time << " us" << std::endl;
202
203 EXPECT_GT(ref_elapsed_time, elapsed_time)
204 << "Error: CDEFSpeedTest, SIMD slower than C." << std::endl
205 << "C time: " << ref_elapsed_time << " us" << std::endl
206 << "SIMD time: " << elapsed_time << " us" << std::endl;
207 }
208
209 typedef int (*find_dir_t)(const uint16_t *img, int stride, int32_t *var,
210 int coeff_shift);
211
212 typedef std::tuple<find_dir_t, find_dir_t> find_dir_param_t;
213
214 class CDEFFindDirTest : public ::testing::TestWithParam<find_dir_param_t> {
215 public:
216 ~CDEFFindDirTest() override = default;
SetUp()217 void SetUp() override {
218 finddir = GET_PARAM(0);
219 ref_finddir = GET_PARAM(1);
220 }
221
222 protected:
223 find_dir_t finddir;
224 find_dir_t ref_finddir;
225 };
226 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFFindDirTest);
227
228 typedef CDEFFindDirTest CDEFFindDirSpeedTest;
229 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFFindDirSpeedTest);
230
test_finddir(int (* finddir)(const uint16_t * img,int stride,int32_t * var,int coeff_shift),int (* ref_finddir)(const uint16_t * img,int stride,int32_t * var,int coeff_shift))231 void test_finddir(int (*finddir)(const uint16_t *img, int stride, int32_t *var,
232 int coeff_shift),
233 int (*ref_finddir)(const uint16_t *img, int stride,
234 int32_t *var, int coeff_shift)) {
235 const int size = 8;
236 ACMRandom rnd(ACMRandom::DeterministicSeed());
237 DECLARE_ALIGNED(16, uint16_t, s[size * size]);
238
239 int error = 0;
240 int depth, bits, level, count, errdepth = 0;
241 int ref_res = 0, res = 0;
242 int32_t ref_var = 0, var = 0;
243
244 for (depth = 8; depth <= 12 && !error; depth += 2) {
245 for (count = 0; count < 512 && !error; count++) {
246 for (level = 0; level < (1 << depth) && !error;
247 level += 1 << (depth - 8)) {
248 for (bits = 1; bits <= depth && !error; bits++) {
249 for (unsigned int i = 0; i < sizeof(s) / sizeof(*s); i++)
250 s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
251 (1 << depth) - 1);
252 for (int c = 0; c < 1 + 9 * (finddir == ref_finddir); c++)
253 ref_res = ref_finddir(s, size, &ref_var, depth - 8);
254 if (finddir != ref_finddir)
255 API_REGISTER_STATE_CHECK(res = finddir(s, size, &var, depth - 8));
256 if (ref_finddir != finddir) {
257 if (res != ref_res || var != ref_var) error = 1;
258 errdepth = depth;
259 }
260 }
261 }
262 }
263 }
264
265 EXPECT_EQ(0, error) << "Error: CDEFFindDirTest, SIMD and C mismatch."
266 << std::endl
267 << "return: " << res << " : " << ref_res << std::endl
268 << "var: " << var << " : " << ref_var << std::endl
269 << "depth: " << errdepth << std::endl
270 << std::endl;
271 }
272
test_finddir_speed(int (* finddir)(const uint16_t * img,int stride,int32_t * var,int coeff_shift),int (* ref_finddir)(const uint16_t * img,int stride,int32_t * var,int coeff_shift))273 void test_finddir_speed(int (*finddir)(const uint16_t *img, int stride,
274 int32_t *var, int coeff_shift),
275 int (*ref_finddir)(const uint16_t *img, int stride,
276 int32_t *var, int coeff_shift)) {
277 aom_usec_timer ref_timer;
278 aom_usec_timer timer;
279
280 aom_usec_timer_start(&ref_timer);
281 test_finddir(ref_finddir, ref_finddir);
282 aom_usec_timer_mark(&ref_timer);
283 int64_t ref_elapsed_time = aom_usec_timer_elapsed(&ref_timer);
284
285 aom_usec_timer_start(&timer);
286 test_finddir(finddir, finddir);
287 aom_usec_timer_mark(&timer);
288 int64_t elapsed_time = aom_usec_timer_elapsed(&timer);
289
290 EXPECT_GT(ref_elapsed_time, elapsed_time)
291 << "Error: CDEFFindDirSpeedTest, SIMD slower than C." << std::endl
292 << "C time: " << ref_elapsed_time << " us" << std::endl
293 << "SIMD time: " << elapsed_time << " us" << std::endl;
294 }
295
296 typedef void (*find_dir_dual_t)(const uint16_t *img1, const uint16_t *img2,
297 int stride, int32_t *var1, int32_t *var2,
298 int coeff_shift, int *out1, int *out2);
299
300 typedef std::tuple<find_dir_dual_t, find_dir_dual_t> find_dir_dual_param_t;
301
302 class CDEFFindDirDualTest
303 : public ::testing::TestWithParam<find_dir_dual_param_t> {
304 public:
305 ~CDEFFindDirDualTest() override = default;
SetUp()306 void SetUp() override {
307 finddir = GET_PARAM(0);
308 ref_finddir = GET_PARAM(1);
309 }
310
311 protected:
312 find_dir_dual_t finddir;
313 find_dir_dual_t ref_finddir;
314 };
315 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFFindDirDualTest);
316
317 typedef CDEFFindDirDualTest CDEFFindDirDualSpeedTest;
318 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFFindDirDualSpeedTest);
319
test_finddir_dual(void (* finddir)(const uint16_t * img1,const uint16_t * img2,int stride,int32_t * var1,int32_t * var2,int coeff_shift,int * out1,int * out2),void (* ref_finddir)(const uint16_t * img1,const uint16_t * img2,int stride,int32_t * var1,int32_t * var2,int coeff_shift,int * out1,int * out2))320 void test_finddir_dual(
321 void (*finddir)(const uint16_t *img1, const uint16_t *img2, int stride,
322 int32_t *var1, int32_t *var2, int coeff_shift, int *out1,
323 int *out2),
324 void (*ref_finddir)(const uint16_t *img1, const uint16_t *img2, int stride,
325 int32_t *var1, int32_t *var2, int coeff_shift,
326 int *out1, int *out2)) {
327 const int size_wd = 16;
328 const int size_ht = 8;
329 ACMRandom rnd(ACMRandom::DeterministicSeed());
330 DECLARE_ALIGNED(16, uint16_t, s[size_ht * size_wd]);
331
332 int error = 0, errdepth = 0;
333 int32_t ref_var[2] = { 0 };
334 int ref_dir[2] = { 0 };
335 int32_t var[2] = { 0 };
336 int dir[2] = { 0 };
337
338 for (int depth = 8; depth <= 12 && !error; depth += 2) {
339 for (int count = 0; count < 512 && !error; count++) {
340 for (int level = 0; level < (1 << depth) && !error;
341 level += 1 << (depth - 8)) {
342 for (int bits = 1; bits <= depth && !error; bits++) {
343 for (unsigned int i = 0; i < sizeof(s) / sizeof(*s); i++)
344 s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
345 (1 << depth) - 1);
346 for (int c = 0; c < 1 + 9 * (finddir == ref_finddir); c++)
347 ref_finddir(s, s + 8, size_wd, &ref_var[0], &ref_var[1], depth - 8,
348 &ref_dir[0], &ref_dir[1]);
349 if (finddir != ref_finddir)
350 API_REGISTER_STATE_CHECK(finddir(s, s + 8, size_wd, &var[0],
351 &var[1], depth - 8, &dir[0],
352 &dir[1]));
353 if (ref_finddir != finddir) {
354 for (int j = 0; j < 2; j++) {
355 if (ref_dir[j] != dir[j] || ref_var[j] != var[j]) error = 1;
356 }
357 errdepth = depth;
358 }
359 }
360 }
361 }
362 }
363
364 for (int j = 0; j < 2; j++) {
365 EXPECT_EQ(0, error) << "Error: CDEFFindDirTest, SIMD and C mismatch."
366 << std::endl
367 << "direction: " << dir[j] << " : " << ref_dir[j]
368 << std::endl
369 << "variance: " << var[j] << " : " << ref_var[j]
370 << std::endl
371 << "depth: " << errdepth << std::endl
372 << std::endl;
373 }
374 }
375
test_finddir_dual_speed(void (* finddir)(const uint16_t * img1,const uint16_t * img2,int stride,int32_t * var1,int32_t * var2,int coeff_shift,int * out1,int * out2),void (* ref_finddir)(const uint16_t * img1,const uint16_t * img2,int stride,int32_t * var1,int32_t * var2,int coeff_shift,int * out1,int * out2))376 void test_finddir_dual_speed(
377 void (*finddir)(const uint16_t *img1, const uint16_t *img2, int stride,
378 int32_t *var1, int32_t *var2, int coeff_shift, int *out1,
379 int *out2),
380 void (*ref_finddir)(const uint16_t *img1, const uint16_t *img2, int stride,
381 int32_t *var1, int32_t *var2, int coeff_shift,
382 int *out1, int *out2)) {
383 aom_usec_timer ref_timer;
384 aom_usec_timer timer;
385
386 aom_usec_timer_start(&ref_timer);
387 test_finddir_dual(ref_finddir, ref_finddir);
388 aom_usec_timer_mark(&ref_timer);
389 const double ref_elapsed_time =
390 static_cast<double>(aom_usec_timer_elapsed(&ref_timer));
391
392 aom_usec_timer_start(&timer);
393 test_finddir_dual(finddir, finddir);
394 aom_usec_timer_mark(&timer);
395 const double elapsed_time =
396 static_cast<double>(aom_usec_timer_elapsed(&timer));
397
398 printf(
399 "ref_time=%lf \t simd_time=%lf \t "
400 "gain=%lf \n",
401 ref_elapsed_time, elapsed_time, ref_elapsed_time / elapsed_time);
402 }
403
404 #define MAX_CDEF_BLOCK 256
405
406 constexpr int kIterations = 100;
407
408 using CDEFCopyRect8To16 = void (*)(uint16_t *dst, int dstride,
409 const uint8_t *src, int sstride, int width,
410 int height);
411
412 using CDEFCopyRect8To16Param = std::tuple<CDEFCopyRect8To16, CDEFCopyRect8To16>;
413
414 class CDEFCopyRect8to16Test
415 : public ::testing::TestWithParam<CDEFCopyRect8To16Param> {
416 public:
CDEFCopyRect8to16Test()417 CDEFCopyRect8to16Test()
418 : rnd_(libaom_test::ACMRandom::DeterministicSeed()),
419 test_func_(GET_PARAM(0)), ref_func_(GET_PARAM(1)) {}
420 ~CDEFCopyRect8to16Test() override = default;
SetUp()421 void SetUp() override {
422 src_ = reinterpret_cast<uint8_t *>(
423 aom_memalign(8, sizeof(uint8_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
424 ASSERT_NE(src_, nullptr);
425 ref_dst_ = reinterpret_cast<uint16_t *>(
426 aom_memalign(16, sizeof(uint16_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
427 ASSERT_NE(ref_dst_, nullptr);
428 test_dst_ = reinterpret_cast<uint16_t *>(
429 aom_memalign(16, sizeof(uint16_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
430 ASSERT_NE(test_dst_, nullptr);
431 }
432
TearDown()433 void TearDown() override {
434 aom_free(src_);
435 aom_free(ref_dst_);
436 aom_free(test_dst_);
437 }
438
test_copy_rect_8_to_16(CDEFCopyRect8To16 test_func,CDEFCopyRect8To16 ref_func)439 void test_copy_rect_8_to_16(CDEFCopyRect8To16 test_func,
440 CDEFCopyRect8To16 ref_func) {
441 constexpr int stride = MAX_CDEF_BLOCK;
442 int error = 0;
443 for (int k = 0; k < kIterations && !error; k++) {
444 // This function operates on values of width that are either 4 or a
445 // multiple of 8. For height, generate a random value between 1 and 256,
446 // making sure it is even.
447 const int width = k == 0 ? 4 : (rnd_.Rand8() % 32 + 1) * 8;
448 const int height = k == 0 ? 4 : (rnd_.Rand8() % 128 + 1) * 2;
449 for (int i = 0; i < height; i++) {
450 for (int j = 0; j < width; j++) {
451 src_[i * stride + j] = rnd_.Rand8();
452 }
453 }
454
455 ref_func(ref_dst_, stride, src_, stride, width, height);
456 test_func(test_dst_, stride, src_, stride, width, height);
457
458 int i, j;
459 for (i = 0; i < height; i++) {
460 for (j = 0; j < width; j++) {
461 if (test_dst_[i * stride + j] != ref_dst_[i * stride + j]) {
462 error = 1;
463 break;
464 }
465 }
466 if (error) {
467 break;
468 }
469 }
470 EXPECT_EQ(0, error)
471 << "Error: CDEFCopyRect8to16Test, SIMD and C mismatch." << std::endl
472 << "First error at " << i << "," << j << " ("
473 << ref_dst_[i * stride + j] << " : " << test_dst_[i * stride + j]
474 << ") " << std::endl
475 << "width: " << width << std::endl
476 << "height: " << height << std::endl
477 << std::endl;
478 }
479 }
480
481 protected:
482 libaom_test::ACMRandom rnd_;
483 uint8_t *src_;
484 uint16_t *ref_dst_;
485 uint16_t *test_dst_;
486 CDEFCopyRect8To16 test_func_;
487 CDEFCopyRect8To16 ref_func_;
488 };
489 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFCopyRect8to16Test);
490
491 using CDEFCopyRect16To16 = void (*)(uint16_t *dst, int dstride,
492 const uint16_t *src, int sstride, int width,
493 int height);
494
495 using CDEFCopyRect16To16Param =
496 std::tuple<CDEFCopyRect16To16, CDEFCopyRect16To16>;
497
498 class CDEFCopyRect16to16Test
499 : public ::testing::TestWithParam<CDEFCopyRect16To16Param> {
500 public:
CDEFCopyRect16to16Test()501 CDEFCopyRect16to16Test()
502 : rnd_(libaom_test::ACMRandom::DeterministicSeed()),
503 test_func_(GET_PARAM(0)), ref_func_(GET_PARAM(1)) {}
504 ~CDEFCopyRect16to16Test() override = default;
SetUp()505 void SetUp() override {
506 src_ = reinterpret_cast<uint16_t *>(
507 aom_memalign(16, sizeof(uint16_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
508 ASSERT_NE(src_, nullptr);
509 ref_dst_ = reinterpret_cast<uint16_t *>(
510 aom_memalign(16, sizeof(uint16_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
511 ASSERT_NE(ref_dst_, nullptr);
512 test_dst_ = reinterpret_cast<uint16_t *>(
513 aom_memalign(16, sizeof(uint16_t) * MAX_CDEF_BLOCK * MAX_CDEF_BLOCK));
514 ASSERT_NE(test_dst_, nullptr);
515 }
516
TearDown()517 void TearDown() override {
518 aom_free(src_);
519 aom_free(ref_dst_);
520 aom_free(test_dst_);
521 }
522
test_copy_rect_16_to_16(CDEFCopyRect16To16 test_func,CDEFCopyRect16To16 ref_func)523 void test_copy_rect_16_to_16(CDEFCopyRect16To16 test_func,
524 CDEFCopyRect16To16 ref_func) {
525 constexpr int stride = MAX_CDEF_BLOCK;
526 int error = 0;
527 for (int k = 0; k < kIterations && !error; k++) {
528 // This function operates on values of width that are either 4 or a
529 // multiple of 8. For height, generate a random value between 1 and 256,
530 // making sure it is even.
531 const int width = k == 0 ? 4 : (rnd_.Rand8() % 32 + 1) * 8;
532 const int height = k == 0 ? 4 : (rnd_.Rand8() % 128 + 1) * 2;
533 for (int i = 0; i < height; i++) {
534 for (int j = 0; j < width; j++) {
535 src_[i * stride + j] = rnd_.Rand16();
536 }
537 }
538
539 ref_func(ref_dst_, stride, src_, stride, width, height);
540 test_func(test_dst_, stride, src_, stride, width, height);
541
542 int i, j;
543 for (i = 0; i < height; i++) {
544 for (j = 0; j < width; j++) {
545 if (test_dst_[i * stride + j] != ref_dst_[i * stride + j]) {
546 error = 1;
547 break;
548 }
549 }
550 if (error) {
551 break;
552 }
553 }
554 EXPECT_EQ(0, error)
555 << "Error: CDEFCopyRect16to16Test, SIMD and C mismatch." << std::endl
556 << "First error at " << i << "," << j << " ("
557 << ref_dst_[i * stride + j] << " : " << test_dst_[i * stride + j]
558 << ") " << std::endl
559 << "width: " << width << std::endl
560 << "height: " << height << std::endl
561 << std::endl;
562 }
563 }
564
565 protected:
566 libaom_test::ACMRandom rnd_;
567 uint16_t *src_;
568 uint16_t *ref_dst_;
569 uint16_t *test_dst_;
570 CDEFCopyRect16To16 test_func_;
571 CDEFCopyRect16To16 ref_func_;
572 };
573 GTEST_ALLOW_UNINSTANTIATED_PARAMETERIZED_TEST(CDEFCopyRect16to16Test);
574
TEST_P(CDEFBlockTest,TestSIMDNoMismatch)575 TEST_P(CDEFBlockTest, TestSIMDNoMismatch) {
576 test_cdef(bsize, 1, cdef, ref_cdef, boundary, depth);
577 }
578
TEST_P(CDEFBlockHighbdTest,TestSIMDHighbdNoMismatch)579 TEST_P(CDEFBlockHighbdTest, TestSIMDHighbdNoMismatch) {
580 test_cdef(bsize, 1, cdef, ref_cdef, boundary, depth);
581 }
582
TEST_P(CDEFSpeedTest,DISABLED_TestSpeed)583 TEST_P(CDEFSpeedTest, DISABLED_TestSpeed) {
584 test_cdef_speed(bsize, 4, cdef, ref_cdef, boundary, depth);
585 }
586
TEST_P(CDEFSpeedHighbdTest,DISABLED_TestSpeed)587 TEST_P(CDEFSpeedHighbdTest, DISABLED_TestSpeed) {
588 test_cdef_speed(bsize, 4, cdef, ref_cdef, boundary, depth);
589 }
590
TEST_P(CDEFFindDirTest,TestSIMDNoMismatch)591 TEST_P(CDEFFindDirTest, TestSIMDNoMismatch) {
592 test_finddir(finddir, ref_finddir);
593 }
594
TEST_P(CDEFFindDirSpeedTest,DISABLED_TestSpeed)595 TEST_P(CDEFFindDirSpeedTest, DISABLED_TestSpeed) {
596 test_finddir_speed(finddir, ref_finddir);
597 }
598
TEST_P(CDEFFindDirDualTest,TestSIMDNoMismatch)599 TEST_P(CDEFFindDirDualTest, TestSIMDNoMismatch) {
600 test_finddir_dual(finddir, ref_finddir);
601 }
602
TEST_P(CDEFFindDirDualSpeedTest,DISABLED_TestSpeed)603 TEST_P(CDEFFindDirDualSpeedTest, DISABLED_TestSpeed) {
604 test_finddir_dual_speed(finddir, ref_finddir);
605 }
606
TEST_P(CDEFCopyRect8to16Test,TestSIMDNoMismatch)607 TEST_P(CDEFCopyRect8to16Test, TestSIMDNoMismatch) {
608 test_copy_rect_8_to_16(test_func_, ref_func_);
609 }
610
TEST_P(CDEFCopyRect16to16Test,TestSIMDNoMismatch)611 TEST_P(CDEFCopyRect16to16Test, TestSIMDNoMismatch) {
612 test_copy_rect_16_to_16(test_func_, ref_func_);
613 }
614
615 using std::make_tuple;
616
617 #if ((AOM_ARCH_X86 && HAVE_SSSE3) || HAVE_SSE4_1 || HAVE_AVX2 || HAVE_NEON)
618 static const CdefFilterBlockFunctions kCdefFilterFuncC[] = {
619 { &cdef_filter_8_0_c, &cdef_filter_8_1_c, &cdef_filter_8_2_c,
620 &cdef_filter_8_3_c }
621 };
622
623 static const CdefFilterBlockFunctions kCdefFilterHighbdFuncC[] = {
624 { &cdef_filter_16_0_c, &cdef_filter_16_0_c, &cdef_filter_16_0_c,
625 &cdef_filter_16_0_c }
626 };
627 #endif
628
629 #if AOM_ARCH_X86 && HAVE_SSSE3
630 static const CdefFilterBlockFunctions kCdefFilterFuncSsse3[] = {
631 { &cdef_filter_8_0_ssse3, &cdef_filter_8_1_ssse3, &cdef_filter_8_2_ssse3,
632 &cdef_filter_8_3_ssse3 }
633 };
634
635 static const CdefFilterBlockFunctions kCdefFilterHighbdFuncSsse3[] = {
636 { &cdef_filter_16_0_ssse3, &cdef_filter_16_1_ssse3, &cdef_filter_16_2_ssse3,
637 &cdef_filter_16_3_ssse3 }
638 };
639
640 INSTANTIATE_TEST_SUITE_P(
641 SSSE3, CDEFBlockTest,
642 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncSsse3),
643 ::testing::ValuesIn(kCdefFilterFuncC),
644 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
645 BLOCK_8X8),
646 ::testing::Range(0, 16), ::testing::Values(8)));
647 INSTANTIATE_TEST_SUITE_P(
648 SSSE3, CDEFBlockHighbdTest,
649 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncSsse3),
650 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
651 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
652 BLOCK_8X8),
653 ::testing::Range(0, 16), ::testing::Range(10, 13, 2)));
654 INSTANTIATE_TEST_SUITE_P(SSSE3, CDEFFindDirTest,
655 ::testing::Values(make_tuple(&cdef_find_dir_ssse3,
656 &cdef_find_dir_c)));
657 INSTANTIATE_TEST_SUITE_P(SSSE3, CDEFFindDirDualTest,
658 ::testing::Values(make_tuple(&cdef_find_dir_dual_ssse3,
659 &cdef_find_dir_dual_c)));
660
661 INSTANTIATE_TEST_SUITE_P(
662 SSSE3, CDEFCopyRect8to16Test,
663 ::testing::Values(make_tuple(&cdef_copy_rect8_8bit_to_16bit_c,
664 &cdef_copy_rect8_8bit_to_16bit_ssse3)));
665
666 INSTANTIATE_TEST_SUITE_P(
667 SSSE3, CDEFCopyRect16to16Test,
668 ::testing::Values(make_tuple(&cdef_copy_rect8_16bit_to_16bit_c,
669 &cdef_copy_rect8_16bit_to_16bit_ssse3)));
670 #endif
671
672 #if HAVE_SSE4_1
673 static const CdefFilterBlockFunctions kCdefFilterFuncSse4_1[] = {
674 { &cdef_filter_8_0_sse4_1, &cdef_filter_8_1_sse4_1, &cdef_filter_8_2_sse4_1,
675 &cdef_filter_8_3_sse4_1 }
676 };
677
678 static const CdefFilterBlockFunctions kCdefFilterHighbdFuncSse4_1[] = {
679 { &cdef_filter_16_0_sse4_1, &cdef_filter_16_1_sse4_1,
680 &cdef_filter_16_2_sse4_1, &cdef_filter_16_3_sse4_1 }
681 };
682
683 INSTANTIATE_TEST_SUITE_P(
684 SSE4_1, CDEFBlockTest,
685 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncSse4_1),
686 ::testing::ValuesIn(kCdefFilterFuncC),
687 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
688 BLOCK_8X8),
689 ::testing::Range(0, 16), ::testing::Values(8)));
690 INSTANTIATE_TEST_SUITE_P(
691 SSE4_1, CDEFBlockHighbdTest,
692 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncSse4_1),
693 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
694 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
695 BLOCK_8X8),
696 ::testing::Range(0, 16), ::testing::Range(10, 13, 2)));
697 INSTANTIATE_TEST_SUITE_P(SSE4_1, CDEFFindDirTest,
698 ::testing::Values(make_tuple(&cdef_find_dir_sse4_1,
699 &cdef_find_dir_c)));
700 INSTANTIATE_TEST_SUITE_P(
701 SSE4_1, CDEFFindDirDualTest,
702 ::testing::Values(make_tuple(&cdef_find_dir_dual_sse4_1,
703 &cdef_find_dir_dual_c)));
704
705 INSTANTIATE_TEST_SUITE_P(
706 SSE4_1, CDEFCopyRect8to16Test,
707 ::testing::Values(make_tuple(&cdef_copy_rect8_8bit_to_16bit_c,
708 &cdef_copy_rect8_8bit_to_16bit_sse4_1)));
709
710 INSTANTIATE_TEST_SUITE_P(
711 SSE4_1, CDEFCopyRect16to16Test,
712 ::testing::Values(make_tuple(&cdef_copy_rect8_16bit_to_16bit_c,
713 &cdef_copy_rect8_16bit_to_16bit_sse4_1)));
714 #endif
715
716 #if HAVE_AVX2
717 static const CdefFilterBlockFunctions kCdefFilterFuncAvx2[] = {
718 { &cdef_filter_8_0_avx2, &cdef_filter_8_1_avx2, &cdef_filter_8_2_avx2,
719 &cdef_filter_8_3_avx2 }
720 };
721
722 static const CdefFilterBlockFunctions kCdefFilterHighbdFuncAvx2[] = {
723 { &cdef_filter_16_0_avx2, &cdef_filter_16_1_avx2, &cdef_filter_16_2_avx2,
724 &cdef_filter_16_3_avx2 }
725 };
726
727 INSTANTIATE_TEST_SUITE_P(
728 AVX2, CDEFBlockTest,
729 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncAvx2),
730 ::testing::ValuesIn(kCdefFilterFuncC),
731 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
732 BLOCK_8X8),
733 ::testing::Range(0, 16), ::testing::Values(8)));
734 INSTANTIATE_TEST_SUITE_P(
735 AVX2, CDEFBlockHighbdTest,
736 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncAvx2),
737 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
738 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
739 BLOCK_8X8),
740 ::testing::Range(0, 16), ::testing::Range(10, 13, 2)));
741 INSTANTIATE_TEST_SUITE_P(AVX2, CDEFFindDirTest,
742 ::testing::Values(make_tuple(&cdef_find_dir_avx2,
743 &cdef_find_dir_c)));
744 INSTANTIATE_TEST_SUITE_P(AVX2, CDEFFindDirDualTest,
745 ::testing::Values(make_tuple(&cdef_find_dir_dual_avx2,
746 &cdef_find_dir_dual_c)));
747
748 INSTANTIATE_TEST_SUITE_P(
749 AVX2, CDEFCopyRect8to16Test,
750 ::testing::Values(make_tuple(&cdef_copy_rect8_8bit_to_16bit_c,
751 &cdef_copy_rect8_8bit_to_16bit_avx2)));
752
753 INSTANTIATE_TEST_SUITE_P(
754 AVX2, CDEFCopyRect16to16Test,
755 ::testing::Values(make_tuple(&cdef_copy_rect8_16bit_to_16bit_c,
756 &cdef_copy_rect8_16bit_to_16bit_avx2)));
757 #endif
758
759 #if HAVE_NEON
760 static const CdefFilterBlockFunctions kCdefFilterFuncNeon[] = {
761 { &cdef_filter_8_0_neon, &cdef_filter_8_1_neon, &cdef_filter_8_2_neon,
762 &cdef_filter_8_3_neon }
763 };
764
765 static const CdefFilterBlockFunctions kCdefFilterHighbdFuncNeon[] = {
766 { &cdef_filter_16_0_neon, &cdef_filter_16_1_neon, &cdef_filter_16_2_neon,
767 &cdef_filter_16_3_neon }
768 };
769
770 INSTANTIATE_TEST_SUITE_P(
771 NEON, CDEFBlockTest,
772 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncNeon),
773 ::testing::ValuesIn(kCdefFilterFuncC),
774 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
775 BLOCK_8X8),
776 ::testing::Range(0, 16), ::testing::Values(8)));
777 INSTANTIATE_TEST_SUITE_P(
778 NEON, CDEFBlockHighbdTest,
779 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncNeon),
780 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
781 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
782 BLOCK_8X8),
783 ::testing::Range(0, 16), ::testing::Range(10, 13, 2)));
784 INSTANTIATE_TEST_SUITE_P(NEON, CDEFFindDirTest,
785 ::testing::Values(make_tuple(&cdef_find_dir_neon,
786 &cdef_find_dir_c)));
787 INSTANTIATE_TEST_SUITE_P(NEON, CDEFFindDirDualTest,
788 ::testing::Values(make_tuple(&cdef_find_dir_dual_neon,
789 &cdef_find_dir_dual_c)));
790
791 INSTANTIATE_TEST_SUITE_P(
792 NEON, CDEFCopyRect8to16Test,
793 ::testing::Values(make_tuple(&cdef_copy_rect8_8bit_to_16bit_c,
794 &cdef_copy_rect8_8bit_to_16bit_neon)));
795
796 INSTANTIATE_TEST_SUITE_P(
797 NEON, CDEFCopyRect16to16Test,
798 ::testing::Values(make_tuple(&cdef_copy_rect8_16bit_to_16bit_c,
799 &cdef_copy_rect8_16bit_to_16bit_neon)));
800 #endif
801
802 // Test speed for all supported architectures
803 #if AOM_ARCH_X86 && HAVE_SSSE3
804 INSTANTIATE_TEST_SUITE_P(
805 SSSE3, CDEFSpeedTest,
806 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncSsse3),
807 ::testing::ValuesIn(kCdefFilterFuncC),
808 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
809 BLOCK_8X8),
810 ::testing::Range(0, 16), ::testing::Values(8)));
811 INSTANTIATE_TEST_SUITE_P(
812 SSSE3, CDEFSpeedHighbdTest,
813 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncSsse3),
814 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
815 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
816 BLOCK_8X8),
817 ::testing::Range(0, 16), ::testing::Values(10)));
818 INSTANTIATE_TEST_SUITE_P(SSSE3, CDEFFindDirSpeedTest,
819 ::testing::Values(make_tuple(&cdef_find_dir_ssse3,
820 &cdef_find_dir_c)));
821 INSTANTIATE_TEST_SUITE_P(SSSE3, CDEFFindDirDualSpeedTest,
822 ::testing::Values(make_tuple(&cdef_find_dir_dual_ssse3,
823 &cdef_find_dir_dual_c)));
824 #endif
825
826 #if HAVE_SSE4_1
827 INSTANTIATE_TEST_SUITE_P(
828 SSE4_1, CDEFSpeedTest,
829 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncSse4_1),
830 ::testing::ValuesIn(kCdefFilterFuncC),
831 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
832 BLOCK_8X8),
833 ::testing::Range(0, 16), ::testing::Values(8)));
834 INSTANTIATE_TEST_SUITE_P(
835 SSE4_1, CDEFSpeedHighbdTest,
836 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncSse4_1),
837 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
838 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
839 BLOCK_8X8),
840 ::testing::Range(0, 16), ::testing::Values(10)));
841 INSTANTIATE_TEST_SUITE_P(SSE4_1, CDEFFindDirSpeedTest,
842 ::testing::Values(make_tuple(&cdef_find_dir_sse4_1,
843 &cdef_find_dir_c)));
844 INSTANTIATE_TEST_SUITE_P(
845 SSE4_1, CDEFFindDirDualSpeedTest,
846 ::testing::Values(make_tuple(&cdef_find_dir_dual_sse4_1,
847 &cdef_find_dir_dual_c)));
848 #endif
849
850 #if HAVE_AVX2
851 INSTANTIATE_TEST_SUITE_P(
852 AVX2, CDEFSpeedTest,
853 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncAvx2),
854 ::testing::ValuesIn(kCdefFilterFuncC),
855 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
856 BLOCK_8X8),
857 ::testing::Range(0, 16), ::testing::Values(8)));
858 INSTANTIATE_TEST_SUITE_P(
859 AVX2, CDEFSpeedHighbdTest,
860 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncAvx2),
861 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
862 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
863 BLOCK_8X8),
864 ::testing::Range(0, 16), ::testing::Values(10)));
865 INSTANTIATE_TEST_SUITE_P(AVX2, CDEFFindDirSpeedTest,
866 ::testing::Values(make_tuple(&cdef_find_dir_avx2,
867 &cdef_find_dir_c)));
868 INSTANTIATE_TEST_SUITE_P(AVX2, CDEFFindDirDualSpeedTest,
869 ::testing::Values(make_tuple(&cdef_find_dir_dual_avx2,
870 &cdef_find_dir_dual_c)));
871 #endif
872
873 #if HAVE_NEON
874 INSTANTIATE_TEST_SUITE_P(
875 NEON, CDEFSpeedTest,
876 ::testing::Combine(::testing::ValuesIn(kCdefFilterFuncNeon),
877 ::testing::ValuesIn(kCdefFilterFuncC),
878 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
879 BLOCK_8X8),
880 ::testing::Range(0, 16), ::testing::Values(8)));
881 INSTANTIATE_TEST_SUITE_P(
882 NEON, CDEFSpeedHighbdTest,
883 ::testing::Combine(::testing::ValuesIn(kCdefFilterHighbdFuncNeon),
884 ::testing::ValuesIn(kCdefFilterHighbdFuncC),
885 ::testing::Values(BLOCK_4X4, BLOCK_4X8, BLOCK_8X4,
886 BLOCK_8X8),
887 ::testing::Range(0, 16), ::testing::Values(10)));
888 INSTANTIATE_TEST_SUITE_P(NEON, CDEFFindDirSpeedTest,
889 ::testing::Values(make_tuple(&cdef_find_dir_neon,
890 &cdef_find_dir_c)));
891 INSTANTIATE_TEST_SUITE_P(NEON, CDEFFindDirDualSpeedTest,
892 ::testing::Values(make_tuple(&cdef_find_dir_dual_neon,
893 &cdef_find_dir_dual_c)));
894 #endif
895
896 } // namespace
897