1 /*
2 * Copyright (c) 2018, 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 <tuple>
13 #include <vector>
14
15 #include "third_party/googletest/src/googletest/include/gtest/gtest.h"
16
17 #include "config/av1_rtcd.h"
18
19 #include "aom_ports/aom_timer.h"
20 #include "av1/common/convolve.h"
21 #include "av1/common/resize.h"
22 #include "test/acm_random.h"
23 #include "test/register_state_check.h"
24 #include "test/util.h"
25
26 namespace {
27 const int kTestIters = 10;
28 const int kPerfIters = 1000;
29
30 const int kVPad = 32;
31 const int kHPad = 32;
32
33 using libaom_test::ACMRandom;
34 using std::make_tuple;
35 using std::tuple;
36
37 template <typename Pixel>
38 class TestImage {
39 public:
TestImage(int w_src,int h,int superres_denom,int x0,int bd)40 TestImage(int w_src, int h, int superres_denom, int x0, int bd)
41 : w_src_(w_src), h_(h), superres_denom_(superres_denom), x0_(x0),
42 bd_(bd) {
43 assert(bd < 16);
44 assert(bd <= 8 * static_cast<int>(sizeof(Pixel)));
45 assert(9 <= superres_denom && superres_denom <= 16);
46 assert(SCALE_NUMERATOR == 8);
47 assert(0 <= x0_ && x0_ <= RS_SCALE_SUBPEL_MASK);
48
49 w_dst_ = w_src_;
50 av1_calculate_unscaled_superres_size(&w_dst_, nullptr, superres_denom);
51
52 src_stride_ = ALIGN_POWER_OF_TWO(w_src_ + 2 * kHPad, 4);
53 dst_stride_ = ALIGN_POWER_OF_TWO(w_dst_ + 2 * kHPad, 4);
54
55 // Allocate image data
56 src_data_.resize(2 * src_block_size());
57 dst_data_.resize(2 * dst_block_size());
58 }
59
60 void Initialize(ACMRandom *rnd);
61 void Check() const;
62
src_stride() const63 int src_stride() const { return src_stride_; }
dst_stride() const64 int dst_stride() const { return dst_stride_; }
65
src_block_size() const66 int src_block_size() const { return (h_ + 2 * kVPad) * src_stride(); }
dst_block_size() const67 int dst_block_size() const { return (h_ + 2 * kVPad) * dst_stride(); }
68
src_width() const69 int src_width() const { return w_src_; }
dst_width() const70 int dst_width() const { return w_dst_; }
height() const71 int height() const { return h_; }
x0() const72 int x0() const { return x0_; }
73
GetSrcData(bool ref,bool borders) const74 const Pixel *GetSrcData(bool ref, bool borders) const {
75 const Pixel *block = &src_data_[ref ? 0 : src_block_size()];
76 return borders ? block : block + kHPad + src_stride_ * kVPad;
77 }
78
GetDstData(bool ref,bool borders)79 Pixel *GetDstData(bool ref, bool borders) {
80 Pixel *block = &dst_data_[ref ? 0 : dst_block_size()];
81 return borders ? block : block + kHPad + dst_stride_ * kVPad;
82 }
83
84 private:
85 int w_src_, w_dst_, h_, superres_denom_, x0_, bd_;
86 int src_stride_, dst_stride_;
87
88 std::vector<Pixel> src_data_;
89 std::vector<Pixel> dst_data_;
90 };
91
92 template <typename Pixel>
FillEdge(ACMRandom * rnd,int num_pixels,int bd,bool trash,Pixel * data)93 void FillEdge(ACMRandom *rnd, int num_pixels, int bd, bool trash, Pixel *data) {
94 if (!trash) {
95 memset(data, 0, sizeof(*data) * num_pixels);
96 return;
97 }
98 const Pixel mask = (1 << bd) - 1;
99 for (int i = 0; i < num_pixels; ++i) data[i] = rnd->Rand16() & mask;
100 }
101
102 template <typename Pixel>
PrepBuffers(ACMRandom * rnd,int w,int h,int stride,int bd,bool trash_edges,Pixel * data)103 void PrepBuffers(ACMRandom *rnd, int w, int h, int stride, int bd,
104 bool trash_edges, Pixel *data) {
105 assert(rnd);
106 const Pixel mask = (1 << bd) - 1;
107
108 // Fill in the first buffer with random data
109 // Top border
110 FillEdge(rnd, stride * kVPad, bd, trash_edges, data);
111 for (int r = 0; r < h; ++r) {
112 Pixel *row_data = data + (kVPad + r) * stride;
113 // Left border, contents, right border
114 FillEdge(rnd, kHPad, bd, trash_edges, row_data);
115 for (int c = 0; c < w; ++c) row_data[kHPad + c] = rnd->Rand16() & mask;
116 FillEdge(rnd, kHPad, bd, trash_edges, row_data + kHPad + w);
117 }
118 // Bottom border
119 FillEdge(rnd, stride * kVPad, bd, trash_edges, data + stride * (kVPad + h));
120
121 const int bpp = sizeof(*data);
122 const int block_elts = stride * (h + 2 * kVPad);
123 const int block_size = bpp * block_elts;
124
125 // Now copy that to the second buffer
126 memcpy(data + block_elts, data, block_size);
127 }
128
129 template <typename Pixel>
Initialize(ACMRandom * rnd)130 void TestImage<Pixel>::Initialize(ACMRandom *rnd) {
131 PrepBuffers(rnd, w_src_, h_, src_stride_, bd_, false, &src_data_[0]);
132 PrepBuffers(rnd, w_dst_, h_, dst_stride_, bd_, true, &dst_data_[0]);
133 }
134
135 template <typename Pixel>
Check() const136 void TestImage<Pixel>::Check() const {
137 const int num_pixels = dst_block_size();
138 const Pixel *ref_dst = &dst_data_[0];
139 const Pixel *tst_dst = &dst_data_[num_pixels];
140
141 // If memcmp returns 0, there's nothing to do.
142 if (0 == memcmp(ref_dst, tst_dst, sizeof(*ref_dst) * num_pixels)) return;
143
144 // Otherwise, iterate through the buffer looking for differences, *ignoring
145 // the edges*
146 const int stride = dst_stride_;
147 for (int r = kVPad; r < h_ + kVPad; ++r) {
148 for (int c = kVPad; c < w_dst_ + kHPad; ++c) {
149 const int32_t ref_value = ref_dst[r * stride + c];
150 const int32_t tst_value = tst_dst[r * stride + c];
151
152 EXPECT_EQ(tst_value, ref_value)
153 << "Error at row: " << (r - kVPad) << ", col: " << (c - kHPad)
154 << ", superres_denom: " << superres_denom_ << ", height: " << h_
155 << ", src_width: " << w_src_ << ", dst_width: " << w_dst_
156 << ", x0: " << x0_;
157 }
158 }
159 }
160
161 template <typename Pixel>
162 class ConvolveHorizRSTestBase : public ::testing::Test {
163 public:
ConvolveHorizRSTestBase()164 ConvolveHorizRSTestBase() : image_(nullptr) {}
165 ~ConvolveHorizRSTestBase() override = default;
166
167 // Implemented by subclasses (SetUp depends on the parameters passed
168 // in and RunOne depends on the function to be tested. These can't
169 // be templated for low/high bit depths because they have different
170 // numbers of parameters)
171 void SetUp() override = 0;
172 virtual void RunOne(bool ref) = 0;
173
174 protected:
SetBitDepth(int bd)175 void SetBitDepth(int bd) { bd_ = bd; }
176
CorrectnessTest()177 void CorrectnessTest() {
178 ACMRandom rnd(ACMRandom::DeterministicSeed());
179 for (int i = 0; i < kTestIters; ++i) {
180 for (int superres_denom = 9; superres_denom <= 16; superres_denom++) {
181 // Get a random height between 512 and 767
182 int height = rnd.Rand8() + 512;
183
184 // Get a random src width between 128 and 383
185 int width_src = rnd.Rand8() + 128;
186
187 // x0 is normally calculated by get_upscale_convolve_x0 in
188 // av1/common/resize.c. However, this test should work for
189 // any value of x0 between 0 and RS_SCALE_SUBPEL_MASK
190 // (inclusive), so we choose one at random.
191 int x0 = rnd.Rand16() % (RS_SCALE_SUBPEL_MASK + 1);
192
193 image_ =
194 new TestImage<Pixel>(width_src, height, superres_denom, x0, bd_);
195 ASSERT_NE(image_, nullptr);
196
197 Prep(&rnd);
198 RunOne(true);
199 RunOne(false);
200 image_->Check();
201
202 delete image_;
203 }
204 }
205 }
206
SpeedTest()207 void SpeedTest() {
208 // Pick some specific parameters to test
209 int height = 767;
210 int width_src = 129;
211 int superres_denom = 13;
212 int x0 = RS_SCALE_SUBPEL_MASK >> 1;
213
214 image_ = new TestImage<Pixel>(width_src, height, superres_denom, x0, bd_);
215 ASSERT_NE(image_, nullptr);
216
217 ACMRandom rnd(ACMRandom::DeterministicSeed());
218 Prep(&rnd);
219
220 aom_usec_timer ref_timer;
221 aom_usec_timer_start(&ref_timer);
222 for (int i = 0; i < kPerfIters; ++i) RunOne(true);
223 aom_usec_timer_mark(&ref_timer);
224 const int64_t ref_time = aom_usec_timer_elapsed(&ref_timer);
225
226 aom_usec_timer tst_timer;
227 aom_usec_timer_start(&tst_timer);
228 for (int i = 0; i < kPerfIters; ++i) RunOne(false);
229 aom_usec_timer_mark(&tst_timer);
230 const int64_t tst_time = aom_usec_timer_elapsed(&tst_timer);
231
232 std::cout << "[ ] C time = " << ref_time / 1000
233 << " ms, SIMD time = " << tst_time / 1000 << " ms\n";
234
235 EXPECT_GT(ref_time, tst_time)
236 << "Error: ConvolveHorizRSTest (Speed Test), SIMD slower than C.\n"
237 << "C time: " << ref_time << " us\n"
238 << "SIMD time: " << tst_time << " us\n";
239 }
240
Prep(ACMRandom * rnd)241 void Prep(ACMRandom *rnd) {
242 assert(rnd);
243 image_->Initialize(rnd);
244 }
245
246 int bd_;
247 TestImage<Pixel> *image_;
248 };
249
250 typedef void (*LowBDConvolveHorizRsFunc)(const uint8_t *src, int src_stride,
251 uint8_t *dst, int dst_stride, int w,
252 int h, const int16_t *x_filters,
253 const int x0_qn, const int x_step_qn);
254
255 // Test parameter list:
256 // <tst_fun_>
257 typedef tuple<LowBDConvolveHorizRsFunc> LowBDParams;
258
259 class LowBDConvolveHorizRSTest
260 : public ConvolveHorizRSTestBase<uint8_t>,
261 public ::testing::WithParamInterface<LowBDParams> {
262 public:
263 ~LowBDConvolveHorizRSTest() override = default;
264
SetUp()265 void SetUp() override {
266 tst_fun_ = GET_PARAM(0);
267 const int bd = 8;
268 SetBitDepth(bd);
269 }
270
RunOne(bool ref)271 void RunOne(bool ref) override {
272 const uint8_t *src = image_->GetSrcData(ref, false);
273 uint8_t *dst = image_->GetDstData(ref, false);
274 const int src_stride = image_->src_stride();
275 const int dst_stride = image_->dst_stride();
276 const int width_src = image_->src_width();
277 const int width_dst = image_->dst_width();
278 const int height = image_->height();
279 const int x0_qn = image_->x0();
280
281 const int32_t x_step_qn =
282 av1_get_upscale_convolve_step(width_src, width_dst);
283
284 if (ref) {
285 av1_convolve_horiz_rs_c(src, src_stride, dst, dst_stride, width_dst,
286 height, &av1_resize_filter_normative[0][0], x0_qn,
287 x_step_qn);
288 } else {
289 tst_fun_(src, src_stride, dst, dst_stride, width_dst, height,
290 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn);
291 }
292 }
293
294 private:
295 LowBDConvolveHorizRsFunc tst_fun_;
296 };
297
TEST_P(LowBDConvolveHorizRSTest,Correctness)298 TEST_P(LowBDConvolveHorizRSTest, Correctness) { CorrectnessTest(); }
TEST_P(LowBDConvolveHorizRSTest,DISABLED_Speed)299 TEST_P(LowBDConvolveHorizRSTest, DISABLED_Speed) { SpeedTest(); }
300
301 INSTANTIATE_TEST_SUITE_P(C, LowBDConvolveHorizRSTest,
302 ::testing::Values(av1_convolve_horiz_rs_c));
303
304 #if HAVE_SSE4_1
305 INSTANTIATE_TEST_SUITE_P(SSE4_1, LowBDConvolveHorizRSTest,
306 ::testing::Values(av1_convolve_horiz_rs_sse4_1));
307 #endif
308
309 #if CONFIG_AV1_HIGHBITDEPTH
310 typedef void (*HighBDConvolveHorizRsFunc)(const uint16_t *src, int src_stride,
311 uint16_t *dst, int dst_stride, int w,
312 int h, const int16_t *x_filters,
313 const int x0_qn, const int x_step_qn,
314 int bd);
315
316 // Test parameter list:
317 // <tst_fun_, bd_>
318 typedef tuple<HighBDConvolveHorizRsFunc, int> HighBDParams;
319
320 class HighBDConvolveHorizRSTest
321 : public ConvolveHorizRSTestBase<uint16_t>,
322 public ::testing::WithParamInterface<HighBDParams> {
323 public:
324 ~HighBDConvolveHorizRSTest() override = default;
325
SetUp()326 void SetUp() override {
327 tst_fun_ = GET_PARAM(0);
328 const int bd = GET_PARAM(1);
329 SetBitDepth(bd);
330 }
331
RunOne(bool ref)332 void RunOne(bool ref) override {
333 const uint16_t *src = image_->GetSrcData(ref, false);
334 uint16_t *dst = image_->GetDstData(ref, false);
335 const int src_stride = image_->src_stride();
336 const int dst_stride = image_->dst_stride();
337 const int width_src = image_->src_width();
338 const int width_dst = image_->dst_width();
339 const int height = image_->height();
340 const int x0_qn = image_->x0();
341
342 const int32_t x_step_qn =
343 av1_get_upscale_convolve_step(width_src, width_dst);
344
345 if (ref) {
346 av1_highbd_convolve_horiz_rs_c(
347 src, src_stride, dst, dst_stride, width_dst, height,
348 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn, bd_);
349 } else {
350 tst_fun_(src, src_stride, dst, dst_stride, width_dst, height,
351 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn, bd_);
352 }
353 }
354
355 private:
356 HighBDConvolveHorizRsFunc tst_fun_;
357 };
358
359 const int kBDs[] = { 8, 10, 12 };
360
TEST_P(HighBDConvolveHorizRSTest,Correctness)361 TEST_P(HighBDConvolveHorizRSTest, Correctness) { CorrectnessTest(); }
TEST_P(HighBDConvolveHorizRSTest,DISABLED_Speed)362 TEST_P(HighBDConvolveHorizRSTest, DISABLED_Speed) { SpeedTest(); }
363
364 INSTANTIATE_TEST_SUITE_P(
365 C, HighBDConvolveHorizRSTest,
366 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_c),
367 ::testing::ValuesIn(kBDs)));
368
369 #if HAVE_SSE4_1
370 INSTANTIATE_TEST_SUITE_P(
371 SSE4_1, HighBDConvolveHorizRSTest,
372 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_sse4_1),
373 ::testing::ValuesIn(kBDs)));
374 #endif // HAVE_SSE4_1
375
376 #if HAVE_NEON
377 INSTANTIATE_TEST_SUITE_P(
378 NEON, HighBDConvolveHorizRSTest,
379 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_neon),
380 ::testing::ValuesIn(kBDs)));
381 #endif // HAVE_NEON
382
383 #endif // CONFIG_AV1_HIGHBITDEPTH
384
385 } // namespace
386