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 "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 // Inverse of av1_calculate_scaled_superres_size(): calculates the original
38 // dimensions from the given scaled dimensions and the scale denominator.
calculate_unscaled_superres_size(int * width,int denom)39 void calculate_unscaled_superres_size(int *width, int denom) {
40 if (denom != SCALE_NUMERATOR) {
41 // Note: av1_calculate_scaled_superres_size() rounds *up* after division
42 // when the resulting dimensions are odd. So here, we round *down*.
43 *width = *width * denom / SCALE_NUMERATOR;
44 }
45 }
46
47 template <typename Pixel>
48 class TestImage {
49 public:
TestImage(int w_src,int h,int superres_denom,int x0,int bd)50 TestImage(int w_src, int h, int superres_denom, int x0, int bd)
51 : w_src_(w_src), h_(h), superres_denom_(superres_denom), x0_(x0),
52 bd_(bd) {
53 assert(bd < 16);
54 assert(bd <= 8 * static_cast<int>(sizeof(Pixel)));
55 assert(9 <= superres_denom && superres_denom <= 16);
56 assert(SCALE_NUMERATOR == 8);
57 assert(0 <= x0_ && x0_ <= RS_SCALE_SUBPEL_MASK);
58
59 w_dst_ = w_src_;
60 calculate_unscaled_superres_size(&w_dst_, superres_denom);
61
62 src_stride_ = ALIGN_POWER_OF_TWO(w_src_ + 2 * kHPad, 4);
63 dst_stride_ = ALIGN_POWER_OF_TWO(w_dst_ + 2 * kHPad, 4);
64
65 // Allocate image data
66 src_data_.resize(2 * src_block_size());
67 dst_data_.resize(2 * dst_block_size());
68 }
69
70 void Initialize(ACMRandom *rnd);
71 void Check() const;
72
src_stride() const73 int src_stride() const { return src_stride_; }
dst_stride() const74 int dst_stride() const { return dst_stride_; }
75
src_block_size() const76 int src_block_size() const { return (h_ + 2 * kVPad) * src_stride(); }
dst_block_size() const77 int dst_block_size() const { return (h_ + 2 * kVPad) * dst_stride(); }
78
src_width() const79 int src_width() const { return w_src_; }
dst_width() const80 int dst_width() const { return w_dst_; }
height() const81 int height() const { return h_; }
x0() const82 int x0() const { return x0_; }
83
GetSrcData(bool ref,bool borders) const84 const Pixel *GetSrcData(bool ref, bool borders) const {
85 const Pixel *block = &src_data_[ref ? 0 : src_block_size()];
86 return borders ? block : block + kHPad + src_stride_ * kVPad;
87 }
88
GetDstData(bool ref,bool borders)89 Pixel *GetDstData(bool ref, bool borders) {
90 Pixel *block = &dst_data_[ref ? 0 : dst_block_size()];
91 return borders ? block : block + kHPad + dst_stride_ * kVPad;
92 }
93
94 private:
95 int w_src_, w_dst_, h_, superres_denom_, x0_, bd_;
96 int src_stride_, dst_stride_;
97
98 std::vector<Pixel> src_data_;
99 std::vector<Pixel> dst_data_;
100 };
101
102 template <typename Pixel>
FillEdge(ACMRandom * rnd,int num_pixels,int bd,bool trash,Pixel * data)103 void FillEdge(ACMRandom *rnd, int num_pixels, int bd, bool trash, Pixel *data) {
104 if (!trash) {
105 memset(data, 0, sizeof(*data) * num_pixels);
106 return;
107 }
108 const Pixel mask = (1 << bd) - 1;
109 for (int i = 0; i < num_pixels; ++i) data[i] = rnd->Rand16() & mask;
110 }
111
112 template <typename Pixel>
PrepBuffers(ACMRandom * rnd,int w,int h,int stride,int bd,bool trash_edges,Pixel * data)113 void PrepBuffers(ACMRandom *rnd, int w, int h, int stride, int bd,
114 bool trash_edges, Pixel *data) {
115 assert(rnd);
116 const Pixel mask = (1 << bd) - 1;
117
118 // Fill in the first buffer with random data
119 // Top border
120 FillEdge(rnd, stride * kVPad, bd, trash_edges, data);
121 for (int r = 0; r < h; ++r) {
122 Pixel *row_data = data + (kVPad + r) * stride;
123 // Left border, contents, right border
124 FillEdge(rnd, kHPad, bd, trash_edges, row_data);
125 for (int c = 0; c < w; ++c) row_data[kHPad + c] = rnd->Rand16() & mask;
126 FillEdge(rnd, kHPad, bd, trash_edges, row_data + kHPad + w);
127 }
128 // Bottom border
129 FillEdge(rnd, stride * kVPad, bd, trash_edges, data + stride * (kVPad + h));
130
131 const int bpp = sizeof(*data);
132 const int block_elts = stride * (h + 2 * kVPad);
133 const int block_size = bpp * block_elts;
134
135 // Now copy that to the second buffer
136 memcpy(data + block_elts, data, block_size);
137 }
138
139 template <typename Pixel>
Initialize(ACMRandom * rnd)140 void TestImage<Pixel>::Initialize(ACMRandom *rnd) {
141 PrepBuffers(rnd, w_src_, h_, src_stride_, bd_, false, &src_data_[0]);
142 PrepBuffers(rnd, w_dst_, h_, dst_stride_, bd_, true, &dst_data_[0]);
143 }
144
145 template <typename Pixel>
Check() const146 void TestImage<Pixel>::Check() const {
147 const int num_pixels = dst_block_size();
148 const Pixel *ref_dst = &dst_data_[0];
149 const Pixel *tst_dst = &dst_data_[num_pixels];
150
151 // If memcmp returns 0, there's nothing to do.
152 if (0 == memcmp(ref_dst, tst_dst, sizeof(*ref_dst) * num_pixels)) return;
153
154 // Otherwise, iterate through the buffer looking for differences, *ignoring
155 // the edges*
156 const int stride = dst_stride_;
157 for (int r = kVPad; r < h_ + kVPad; ++r) {
158 for (int c = kVPad; c < w_dst_ + kHPad; ++c) {
159 const int32_t ref_value = ref_dst[r * stride + c];
160 const int32_t tst_value = tst_dst[r * stride + c];
161
162 EXPECT_EQ(tst_value, ref_value)
163 << "Error at row: " << (r - kVPad) << ", col: " << (c - kHPad)
164 << ", superres_denom: " << superres_denom_ << ", height: " << h_
165 << ", src_width: " << w_src_ << ", dst_width: " << w_dst_
166 << ", x0: " << x0_;
167 }
168 }
169 }
170
171 template <typename Pixel>
172 class ConvolveHorizRSTestBase : public ::testing::Test {
173 public:
ConvolveHorizRSTestBase()174 ConvolveHorizRSTestBase() : image_(nullptr) {}
175 ~ConvolveHorizRSTestBase() override = default;
176
177 // Implemented by subclasses (SetUp depends on the parameters passed
178 // in and RunOne depends on the function to be tested. These can't
179 // be templated for low/high bit depths because they have different
180 // numbers of parameters)
181 void SetUp() override = 0;
182 virtual void RunOne(bool ref) = 0;
183
184 protected:
SetBitDepth(int bd)185 void SetBitDepth(int bd) { bd_ = bd; }
186
CorrectnessTest()187 void CorrectnessTest() {
188 ACMRandom rnd(ACMRandom::DeterministicSeed());
189 for (int i = 0; i < kTestIters; ++i) {
190 for (int superres_denom = 9; superres_denom <= 16; superres_denom++) {
191 // Get a random height between 512 and 767
192 int height = rnd.Rand8() + 512;
193
194 // Get a random src width between 128 and 383
195 int width_src = rnd.Rand8() + 128;
196
197 // x0 is normally calculated by get_upscale_convolve_x0 in
198 // av1/common/resize.c. However, this test should work for
199 // any value of x0 between 0 and RS_SCALE_SUBPEL_MASK
200 // (inclusive), so we choose one at random.
201 int x0 = rnd.Rand16() % (RS_SCALE_SUBPEL_MASK + 1);
202
203 image_ =
204 new TestImage<Pixel>(width_src, height, superres_denom, x0, bd_);
205 ASSERT_NE(image_, nullptr);
206
207 Prep(&rnd);
208 RunOne(true);
209 RunOne(false);
210 image_->Check();
211
212 delete image_;
213 }
214 }
215 }
216
SpeedTest()217 void SpeedTest() {
218 // Pick some specific parameters to test
219 int height = 767;
220 int width_src = 129;
221 int superres_denom = 13;
222 int x0 = RS_SCALE_SUBPEL_MASK >> 1;
223
224 image_ = new TestImage<Pixel>(width_src, height, superres_denom, x0, bd_);
225 ASSERT_NE(image_, nullptr);
226
227 ACMRandom rnd(ACMRandom::DeterministicSeed());
228 Prep(&rnd);
229
230 aom_usec_timer ref_timer;
231 aom_usec_timer_start(&ref_timer);
232 for (int i = 0; i < kPerfIters; ++i) RunOne(true);
233 aom_usec_timer_mark(&ref_timer);
234 const int64_t ref_time = aom_usec_timer_elapsed(&ref_timer);
235
236 aom_usec_timer tst_timer;
237 aom_usec_timer_start(&tst_timer);
238 for (int i = 0; i < kPerfIters; ++i) RunOne(false);
239 aom_usec_timer_mark(&tst_timer);
240 const int64_t tst_time = aom_usec_timer_elapsed(&tst_timer);
241
242 std::cout << "[ ] C time = " << ref_time / 1000
243 << " ms, SIMD time = " << tst_time / 1000 << " ms\n";
244
245 EXPECT_GT(ref_time, tst_time)
246 << "Error: ConvolveHorizRSTest (Speed Test), SIMD slower than C.\n"
247 << "C time: " << ref_time << " us\n"
248 << "SIMD time: " << tst_time << " us\n";
249 }
250
Prep(ACMRandom * rnd)251 void Prep(ACMRandom *rnd) {
252 assert(rnd);
253 image_->Initialize(rnd);
254 }
255
256 int bd_;
257 TestImage<Pixel> *image_;
258 };
259
260 typedef void (*LowBDConvolveHorizRsFunc)(const uint8_t *src, int src_stride,
261 uint8_t *dst, int dst_stride, int w,
262 int h, const int16_t *x_filters,
263 const int x0_qn, const int x_step_qn);
264
265 // Test parameter list:
266 // <tst_fun_>
267 typedef tuple<LowBDConvolveHorizRsFunc> LowBDParams;
268
269 class LowBDConvolveHorizRSTest
270 : public ConvolveHorizRSTestBase<uint8_t>,
271 public ::testing::WithParamInterface<LowBDParams> {
272 public:
273 ~LowBDConvolveHorizRSTest() override = default;
274
SetUp()275 void SetUp() override {
276 tst_fun_ = GET_PARAM(0);
277 const int bd = 8;
278 SetBitDepth(bd);
279 }
280
RunOne(bool ref)281 void RunOne(bool ref) override {
282 const uint8_t *src = image_->GetSrcData(ref, false);
283 uint8_t *dst = image_->GetDstData(ref, false);
284 const int src_stride = image_->src_stride();
285 const int dst_stride = image_->dst_stride();
286 const int width_src = image_->src_width();
287 const int width_dst = image_->dst_width();
288 const int height = image_->height();
289 const int x0_qn = image_->x0();
290
291 const int32_t x_step_qn =
292 av1_get_upscale_convolve_step(width_src, width_dst);
293
294 if (ref) {
295 av1_convolve_horiz_rs_c(src, src_stride, dst, dst_stride, width_dst,
296 height, &av1_resize_filter_normative[0][0], x0_qn,
297 x_step_qn);
298 } else {
299 tst_fun_(src, src_stride, dst, dst_stride, width_dst, height,
300 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn);
301 }
302 }
303
304 private:
305 LowBDConvolveHorizRsFunc tst_fun_;
306 };
307
TEST_P(LowBDConvolveHorizRSTest,Correctness)308 TEST_P(LowBDConvolveHorizRSTest, Correctness) { CorrectnessTest(); }
TEST_P(LowBDConvolveHorizRSTest,DISABLED_Speed)309 TEST_P(LowBDConvolveHorizRSTest, DISABLED_Speed) { SpeedTest(); }
310
311 INSTANTIATE_TEST_SUITE_P(C, LowBDConvolveHorizRSTest,
312 ::testing::Values(av1_convolve_horiz_rs_c));
313
314 #if HAVE_NEON
315 INSTANTIATE_TEST_SUITE_P(NEON, LowBDConvolveHorizRSTest,
316 ::testing::Values(av1_convolve_horiz_rs_neon));
317 #endif
318
319 #if HAVE_SSE4_1
320 INSTANTIATE_TEST_SUITE_P(SSE4_1, LowBDConvolveHorizRSTest,
321 ::testing::Values(av1_convolve_horiz_rs_sse4_1));
322 #endif
323
324 #if CONFIG_AV1_HIGHBITDEPTH
325 typedef void (*HighBDConvolveHorizRsFunc)(const uint16_t *src, int src_stride,
326 uint16_t *dst, int dst_stride, int w,
327 int h, const int16_t *x_filters,
328 const int x0_qn, const int x_step_qn,
329 int bd);
330
331 // Test parameter list:
332 // <tst_fun_, bd_>
333 typedef tuple<HighBDConvolveHorizRsFunc, int> HighBDParams;
334
335 class HighBDConvolveHorizRSTest
336 : public ConvolveHorizRSTestBase<uint16_t>,
337 public ::testing::WithParamInterface<HighBDParams> {
338 public:
339 ~HighBDConvolveHorizRSTest() override = default;
340
SetUp()341 void SetUp() override {
342 tst_fun_ = GET_PARAM(0);
343 const int bd = GET_PARAM(1);
344 SetBitDepth(bd);
345 }
346
RunOne(bool ref)347 void RunOne(bool ref) override {
348 const uint16_t *src = image_->GetSrcData(ref, false);
349 uint16_t *dst = image_->GetDstData(ref, false);
350 const int src_stride = image_->src_stride();
351 const int dst_stride = image_->dst_stride();
352 const int width_src = image_->src_width();
353 const int width_dst = image_->dst_width();
354 const int height = image_->height();
355 const int x0_qn = image_->x0();
356
357 const int32_t x_step_qn =
358 av1_get_upscale_convolve_step(width_src, width_dst);
359
360 if (ref) {
361 av1_highbd_convolve_horiz_rs_c(
362 src, src_stride, dst, dst_stride, width_dst, height,
363 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn, bd_);
364 } else {
365 tst_fun_(src, src_stride, dst, dst_stride, width_dst, height,
366 &av1_resize_filter_normative[0][0], x0_qn, x_step_qn, bd_);
367 }
368 }
369
370 private:
371 HighBDConvolveHorizRsFunc tst_fun_;
372 };
373
374 const int kBDs[] = { 8, 10, 12 };
375
TEST_P(HighBDConvolveHorizRSTest,Correctness)376 TEST_P(HighBDConvolveHorizRSTest, Correctness) { CorrectnessTest(); }
TEST_P(HighBDConvolveHorizRSTest,DISABLED_Speed)377 TEST_P(HighBDConvolveHorizRSTest, DISABLED_Speed) { SpeedTest(); }
378
379 INSTANTIATE_TEST_SUITE_P(
380 C, HighBDConvolveHorizRSTest,
381 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_c),
382 ::testing::ValuesIn(kBDs)));
383
384 #if HAVE_SSE4_1
385 INSTANTIATE_TEST_SUITE_P(
386 SSE4_1, HighBDConvolveHorizRSTest,
387 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_sse4_1),
388 ::testing::ValuesIn(kBDs)));
389 #endif // HAVE_SSE4_1
390
391 #if HAVE_NEON
392 INSTANTIATE_TEST_SUITE_P(
393 NEON, HighBDConvolveHorizRSTest,
394 ::testing::Combine(::testing::Values(av1_highbd_convolve_horiz_rs_neon),
395 ::testing::ValuesIn(kBDs)));
396 #endif // HAVE_NEON
397
398 #endif // CONFIG_AV1_HIGHBITDEPTH
399
400 } // namespace
401