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 <string.h> 13 14 #include "third_party/googletest/src/googletest/include/gtest/gtest.h" 15 16 #include "config/aom_config.h" 17 #include "config/aom_dsp_rtcd.h" 18 19 #include "aom_dsp/aom_dsp_common.h" 20 #include "aom_dsp/aom_filter.h" 21 #include "aom_mem/aom_mem.h" 22 #include "aom_ports/aom_timer.h" 23 #include "aom_ports/mem.h" 24 #include "av1/common/filter.h" 25 #include "test/acm_random.h" 26 #include "test/clear_system_state.h" 27 #include "test/register_state_check.h" 28 #include "test/util.h" 29 30 namespace { 31 32 static const unsigned int kMaxDimension = MAX_SB_SIZE; 33 34 typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride, 35 uint8_t *dst, ptrdiff_t dst_stride, 36 const int16_t *filter_x, int filter_x_stride, 37 const int16_t *filter_y, int filter_y_stride, 38 int w, int h); 39 40 struct ConvolveFunctions { 41 ConvolveFunctions(ConvolveFunc copy, ConvolveFunc h8, ConvolveFunc v8, int bd) 42 : copy_(copy), h8_(h8), v8_(v8), use_highbd_(bd) {} 43 44 ConvolveFunc copy_; 45 ConvolveFunc h8_; 46 ConvolveFunc v8_; 47 int use_highbd_; // 0 if high bitdepth not used, else the actual bit depth. 48 }; 49 50 typedef ::testing::tuple<int, int, const ConvolveFunctions *> ConvolveParam; 51 52 #define ALL_SIZES_64(convolve_fn) \ 53 make_tuple(4, 4, &convolve_fn), make_tuple(8, 4, &convolve_fn), \ 54 make_tuple(4, 8, &convolve_fn), make_tuple(8, 8, &convolve_fn), \ 55 make_tuple(16, 8, &convolve_fn), make_tuple(8, 16, &convolve_fn), \ 56 make_tuple(16, 16, &convolve_fn), make_tuple(32, 16, &convolve_fn), \ 57 make_tuple(16, 32, &convolve_fn), make_tuple(32, 32, &convolve_fn), \ 58 make_tuple(64, 32, &convolve_fn), make_tuple(32, 64, &convolve_fn), \ 59 make_tuple(64, 64, &convolve_fn) 60 61 #define ALL_SIZES(convolve_fn) \ 62 make_tuple(128, 64, &convolve_fn), make_tuple(64, 128, &convolve_fn), \ 63 make_tuple(128, 128, &convolve_fn), ALL_SIZES_64(convolve_fn) 64 65 // Reference 8-tap subpixel filter, slightly modified to fit into this test. 66 #define AV1_FILTER_WEIGHT 128 67 #define AV1_FILTER_SHIFT 7 68 uint8_t clip_pixel(int x) { return x < 0 ? 0 : x > 255 ? 255 : x; } 69 70 void filter_block2d_8_c(const uint8_t *src_ptr, unsigned int src_stride, 71 const int16_t *HFilter, const int16_t *VFilter, 72 uint8_t *dst_ptr, unsigned int dst_stride, 73 unsigned int output_width, unsigned int output_height) { 74 // Between passes, we use an intermediate buffer whose height is extended to 75 // have enough horizontally filtered values as input for the vertical pass. 76 // This buffer is allocated to be big enough for the largest block type we 77 // support. 78 const int kInterp_Extend = 4; 79 const unsigned int intermediate_height = 80 (kInterp_Extend - 1) + output_height + kInterp_Extend; 81 unsigned int i, j; 82 83 assert(intermediate_height > 7); 84 85 // Size of intermediate_buffer is max_intermediate_height * filter_max_width, 86 // where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height 87 // + kInterp_Extend 88 // = 3 + 16 + 4 89 // = 23 90 // and filter_max_width = 16 91 // 92 uint8_t intermediate_buffer[(kMaxDimension + 8) * kMaxDimension]; 93 const int intermediate_next_stride = 94 1 - static_cast<int>(intermediate_height * output_width); 95 96 // Horizontal pass (src -> transposed intermediate). 97 uint8_t *output_ptr = intermediate_buffer; 98 const int src_next_row_stride = src_stride - output_width; 99 src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1); 100 for (i = 0; i < intermediate_height; ++i) { 101 for (j = 0; j < output_width; ++j) { 102 // Apply filter... 103 const int temp = (src_ptr[0] * HFilter[0]) + (src_ptr[1] * HFilter[1]) + 104 (src_ptr[2] * HFilter[2]) + (src_ptr[3] * HFilter[3]) + 105 (src_ptr[4] * HFilter[4]) + (src_ptr[5] * HFilter[5]) + 106 (src_ptr[6] * HFilter[6]) + (src_ptr[7] * HFilter[7]) + 107 (AV1_FILTER_WEIGHT >> 1); // Rounding 108 109 // Normalize back to 0-255... 110 *output_ptr = clip_pixel(temp >> AV1_FILTER_SHIFT); 111 ++src_ptr; 112 output_ptr += intermediate_height; 113 } 114 src_ptr += src_next_row_stride; 115 output_ptr += intermediate_next_stride; 116 } 117 118 // Vertical pass (transposed intermediate -> dst). 119 src_ptr = intermediate_buffer; 120 const int dst_next_row_stride = dst_stride - output_width; 121 for (i = 0; i < output_height; ++i) { 122 for (j = 0; j < output_width; ++j) { 123 // Apply filter... 124 const int temp = (src_ptr[0] * VFilter[0]) + (src_ptr[1] * VFilter[1]) + 125 (src_ptr[2] * VFilter[2]) + (src_ptr[3] * VFilter[3]) + 126 (src_ptr[4] * VFilter[4]) + (src_ptr[5] * VFilter[5]) + 127 (src_ptr[6] * VFilter[6]) + (src_ptr[7] * VFilter[7]) + 128 (AV1_FILTER_WEIGHT >> 1); // Rounding 129 130 // Normalize back to 0-255... 131 *dst_ptr++ = clip_pixel(temp >> AV1_FILTER_SHIFT); 132 src_ptr += intermediate_height; 133 } 134 src_ptr += intermediate_next_stride; 135 dst_ptr += dst_next_row_stride; 136 } 137 } 138 139 void block2d_average_c(uint8_t *src, unsigned int src_stride, 140 uint8_t *output_ptr, unsigned int output_stride, 141 unsigned int output_width, unsigned int output_height) { 142 unsigned int i, j; 143 for (i = 0; i < output_height; ++i) { 144 for (j = 0; j < output_width; ++j) { 145 output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1; 146 } 147 output_ptr += output_stride; 148 } 149 } 150 151 void filter_average_block2d_8_c(const uint8_t *src_ptr, 152 const unsigned int src_stride, 153 const int16_t *HFilter, const int16_t *VFilter, 154 uint8_t *dst_ptr, unsigned int dst_stride, 155 unsigned int output_width, 156 unsigned int output_height) { 157 uint8_t tmp[kMaxDimension * kMaxDimension]; 158 159 assert(output_width <= kMaxDimension); 160 assert(output_height <= kMaxDimension); 161 filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, kMaxDimension, 162 output_width, output_height); 163 block2d_average_c(tmp, kMaxDimension, dst_ptr, dst_stride, output_width, 164 output_height); 165 } 166 167 void highbd_filter_block2d_8_c(const uint16_t *src_ptr, 168 const unsigned int src_stride, 169 const int16_t *HFilter, const int16_t *VFilter, 170 uint16_t *dst_ptr, unsigned int dst_stride, 171 unsigned int output_width, 172 unsigned int output_height, int bd) { 173 // Between passes, we use an intermediate buffer whose height is extended to 174 // have enough horizontally filtered values as input for the vertical pass. 175 // This buffer is allocated to be big enough for the largest block type we 176 // support. 177 const int kInterp_Extend = 4; 178 const unsigned int intermediate_height = 179 (kInterp_Extend - 1) + output_height + kInterp_Extend; 180 181 /* Size of intermediate_buffer is max_intermediate_height * filter_max_width, 182 * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height 183 * + kInterp_Extend 184 * = 3 + 16 + 4 185 * = 23 186 * and filter_max_width = 16 187 */ 188 uint16_t intermediate_buffer[(kMaxDimension + 8) * kMaxDimension] = { 0 }; 189 const int intermediate_next_stride = 190 1 - static_cast<int>(intermediate_height * output_width); 191 192 // Horizontal pass (src -> transposed intermediate). 193 { 194 uint16_t *output_ptr = intermediate_buffer; 195 const int src_next_row_stride = src_stride - output_width; 196 unsigned int i, j; 197 src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1); 198 for (i = 0; i < intermediate_height; ++i) { 199 for (j = 0; j < output_width; ++j) { 200 // Apply filter... 201 const int temp = (src_ptr[0] * HFilter[0]) + (src_ptr[1] * HFilter[1]) + 202 (src_ptr[2] * HFilter[2]) + (src_ptr[3] * HFilter[3]) + 203 (src_ptr[4] * HFilter[4]) + (src_ptr[5] * HFilter[5]) + 204 (src_ptr[6] * HFilter[6]) + (src_ptr[7] * HFilter[7]) + 205 (AV1_FILTER_WEIGHT >> 1); // Rounding 206 207 // Normalize back to 0-255... 208 *output_ptr = clip_pixel_highbd(temp >> AV1_FILTER_SHIFT, bd); 209 ++src_ptr; 210 output_ptr += intermediate_height; 211 } 212 src_ptr += src_next_row_stride; 213 output_ptr += intermediate_next_stride; 214 } 215 } 216 217 // Vertical pass (transposed intermediate -> dst). 218 { 219 const uint16_t *interm_ptr = intermediate_buffer; 220 const int dst_next_row_stride = dst_stride - output_width; 221 unsigned int i, j; 222 for (i = 0; i < output_height; ++i) { 223 for (j = 0; j < output_width; ++j) { 224 // Apply filter... 225 const int temp = 226 (interm_ptr[0] * VFilter[0]) + (interm_ptr[1] * VFilter[1]) + 227 (interm_ptr[2] * VFilter[2]) + (interm_ptr[3] * VFilter[3]) + 228 (interm_ptr[4] * VFilter[4]) + (interm_ptr[5] * VFilter[5]) + 229 (interm_ptr[6] * VFilter[6]) + (interm_ptr[7] * VFilter[7]) + 230 (AV1_FILTER_WEIGHT >> 1); // Rounding 231 232 // Normalize back to 0-255... 233 *dst_ptr++ = clip_pixel_highbd(temp >> AV1_FILTER_SHIFT, bd); 234 interm_ptr += intermediate_height; 235 } 236 interm_ptr += intermediate_next_stride; 237 dst_ptr += dst_next_row_stride; 238 } 239 } 240 } 241 242 void highbd_block2d_average_c(uint16_t *src, unsigned int src_stride, 243 uint16_t *output_ptr, unsigned int output_stride, 244 unsigned int output_width, 245 unsigned int output_height) { 246 unsigned int i, j; 247 for (i = 0; i < output_height; ++i) { 248 for (j = 0; j < output_width; ++j) { 249 output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1; 250 } 251 output_ptr += output_stride; 252 } 253 } 254 255 void highbd_filter_average_block2d_8_c( 256 const uint16_t *src_ptr, unsigned int src_stride, const int16_t *HFilter, 257 const int16_t *VFilter, uint16_t *dst_ptr, unsigned int dst_stride, 258 unsigned int output_width, unsigned int output_height, int bd) { 259 uint16_t tmp[kMaxDimension * kMaxDimension]; 260 261 assert(output_width <= kMaxDimension); 262 assert(output_height <= kMaxDimension); 263 highbd_filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 264 kMaxDimension, output_width, output_height, bd); 265 highbd_block2d_average_c(tmp, kMaxDimension, dst_ptr, dst_stride, 266 output_width, output_height); 267 } 268 269 class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> { 270 public: 271 static void SetUpTestCase() { 272 // Force input_ to be unaligned, output to be 16 byte aligned. 273 input_ = reinterpret_cast<uint8_t *>( 274 aom_memalign(kDataAlignment, kInputBufferSize + 1)) + 275 1; 276 ref8_ = reinterpret_cast<uint8_t *>( 277 aom_memalign(kDataAlignment, kOutputStride * kMaxDimension)); 278 output_ = reinterpret_cast<uint8_t *>( 279 aom_memalign(kDataAlignment, kOutputBufferSize)); 280 output_ref_ = reinterpret_cast<uint8_t *>( 281 aom_memalign(kDataAlignment, kOutputBufferSize)); 282 input16_ = reinterpret_cast<uint16_t *>(aom_memalign( 283 kDataAlignment, (kInputBufferSize + 1) * sizeof(uint16_t))) + 284 1; 285 ref16_ = reinterpret_cast<uint16_t *>(aom_memalign( 286 kDataAlignment, kOutputStride * kMaxDimension * sizeof(uint16_t))); 287 output16_ = reinterpret_cast<uint16_t *>( 288 aom_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t))); 289 output16_ref_ = reinterpret_cast<uint16_t *>( 290 aom_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t))); 291 } 292 293 virtual void TearDown() { libaom_test::ClearSystemState(); } 294 295 static void TearDownTestCase() { 296 aom_free(input_ - 1); 297 input_ = NULL; 298 aom_free(ref8_); 299 ref8_ = NULL; 300 aom_free(output_); 301 output_ = NULL; 302 aom_free(output_ref_); 303 output_ref_ = NULL; 304 aom_free(input16_ - 1); 305 input16_ = NULL; 306 aom_free(ref16_); 307 ref16_ = NULL; 308 aom_free(output16_); 309 output16_ = NULL; 310 aom_free(output16_ref_); 311 output16_ref_ = NULL; 312 } 313 314 protected: 315 static const int kDataAlignment = 16; 316 static const int kOuterBlockSize = 4 * kMaxDimension; 317 static const int kInputStride = kOuterBlockSize; 318 static const int kOutputStride = kOuterBlockSize; 319 static const int kInputBufferSize = kOuterBlockSize * kOuterBlockSize; 320 static const int kOutputBufferSize = kOuterBlockSize * kOuterBlockSize; 321 322 int Width() const { return GET_PARAM(0); } 323 int Height() const { return GET_PARAM(1); } 324 int BorderLeft() const { 325 const int center = (kOuterBlockSize - Width()) / 2; 326 return (center + (kDataAlignment - 1)) & ~(kDataAlignment - 1); 327 } 328 int BorderTop() const { return (kOuterBlockSize - Height()) / 2; } 329 330 bool IsIndexInBorder(int i) { 331 return (i < BorderTop() * kOuterBlockSize || 332 i >= (BorderTop() + Height()) * kOuterBlockSize || 333 i % kOuterBlockSize < BorderLeft() || 334 i % kOuterBlockSize >= (BorderLeft() + Width())); 335 } 336 337 virtual void SetUp() { 338 UUT_ = GET_PARAM(2); 339 if (UUT_->use_highbd_ != 0) 340 mask_ = (1 << UUT_->use_highbd_) - 1; 341 else 342 mask_ = 255; 343 /* Set up guard blocks for an inner block centered in the outer block */ 344 for (int i = 0; i < kOutputBufferSize; ++i) { 345 if (IsIndexInBorder(i)) { 346 output_[i] = 255; 347 output16_[i] = mask_; 348 } else { 349 output_[i] = 0; 350 output16_[i] = 0; 351 } 352 } 353 354 ::libaom_test::ACMRandom prng; 355 for (int i = 0; i < kInputBufferSize; ++i) { 356 if (i & 1) { 357 input_[i] = 255; 358 input16_[i] = mask_; 359 } else { 360 input_[i] = prng.Rand8Extremes(); 361 input16_[i] = prng.Rand16() & mask_; 362 } 363 } 364 } 365 366 void SetConstantInput(int value) { 367 memset(input_, value, kInputBufferSize); 368 aom_memset16(input16_, value, kInputBufferSize); 369 } 370 371 void CopyOutputToRef() { 372 memcpy(output_ref_, output_, kOutputBufferSize); 373 // Copy 16-bit pixels values. The effective number of bytes is double. 374 memcpy(output16_ref_, output16_, sizeof(output16_[0]) * kOutputBufferSize); 375 } 376 377 void CheckGuardBlocks() { 378 for (int i = 0; i < kOutputBufferSize; ++i) { 379 if (IsIndexInBorder(i)) { 380 EXPECT_EQ(255, output_[i]); 381 } 382 } 383 } 384 385 uint8_t *input() const { 386 const int offset = BorderTop() * kOuterBlockSize + BorderLeft(); 387 if (UUT_->use_highbd_ == 0) { 388 return input_ + offset; 389 } else { 390 return CONVERT_TO_BYTEPTR(input16_) + offset; 391 } 392 } 393 394 uint8_t *output() const { 395 const int offset = BorderTop() * kOuterBlockSize + BorderLeft(); 396 if (UUT_->use_highbd_ == 0) { 397 return output_ + offset; 398 } else { 399 return CONVERT_TO_BYTEPTR(output16_) + offset; 400 } 401 } 402 403 uint8_t *output_ref() const { 404 const int offset = BorderTop() * kOuterBlockSize + BorderLeft(); 405 if (UUT_->use_highbd_ == 0) { 406 return output_ref_ + offset; 407 } else { 408 return CONVERT_TO_BYTEPTR(output16_ref_) + offset; 409 } 410 } 411 412 uint16_t lookup(uint8_t *list, int index) const { 413 if (UUT_->use_highbd_ == 0) { 414 return list[index]; 415 } else { 416 return CONVERT_TO_SHORTPTR(list)[index]; 417 } 418 } 419 420 void assign_val(uint8_t *list, int index, uint16_t val) const { 421 if (UUT_->use_highbd_ == 0) { 422 list[index] = (uint8_t)val; 423 } else { 424 CONVERT_TO_SHORTPTR(list)[index] = val; 425 } 426 } 427 428 void wrapper_filter_average_block2d_8_c( 429 const uint8_t *src_ptr, unsigned int src_stride, const int16_t *HFilter, 430 const int16_t *VFilter, uint8_t *dst_ptr, unsigned int dst_stride, 431 unsigned int output_width, unsigned int output_height) { 432 if (UUT_->use_highbd_ == 0) { 433 filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr, 434 dst_stride, output_width, output_height); 435 } else { 436 highbd_filter_average_block2d_8_c( 437 CONVERT_TO_SHORTPTR(src_ptr), src_stride, HFilter, VFilter, 438 CONVERT_TO_SHORTPTR(dst_ptr), dst_stride, output_width, output_height, 439 UUT_->use_highbd_); 440 } 441 } 442 443 void wrapper_filter_block2d_8_c( 444 const uint8_t *src_ptr, unsigned int src_stride, const int16_t *HFilter, 445 const int16_t *VFilter, uint8_t *dst_ptr, unsigned int dst_stride, 446 unsigned int output_width, unsigned int output_height) { 447 if (UUT_->use_highbd_ == 0) { 448 filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr, 449 dst_stride, output_width, output_height); 450 } else { 451 highbd_filter_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride, 452 HFilter, VFilter, CONVERT_TO_SHORTPTR(dst_ptr), 453 dst_stride, output_width, output_height, 454 UUT_->use_highbd_); 455 } 456 } 457 458 const ConvolveFunctions *UUT_; 459 static uint8_t *input_; 460 static uint8_t *ref8_; 461 static uint8_t *output_; 462 static uint8_t *output_ref_; 463 static uint16_t *input16_; 464 static uint16_t *ref16_; 465 static uint16_t *output16_; 466 static uint16_t *output16_ref_; 467 int mask_; 468 }; 469 470 uint8_t *ConvolveTest::input_ = NULL; 471 uint8_t *ConvolveTest::ref8_ = NULL; 472 uint8_t *ConvolveTest::output_ = NULL; 473 uint8_t *ConvolveTest::output_ref_ = NULL; 474 uint16_t *ConvolveTest::input16_ = NULL; 475 uint16_t *ConvolveTest::ref16_ = NULL; 476 uint16_t *ConvolveTest::output16_ = NULL; 477 uint16_t *ConvolveTest::output16_ref_ = NULL; 478 479 TEST_P(ConvolveTest, GuardBlocks) { CheckGuardBlocks(); } 480 481 TEST_P(ConvolveTest, Copy) { 482 uint8_t *const in = input(); 483 uint8_t *const out = output(); 484 485 ASM_REGISTER_STATE_CHECK(UUT_->copy_(in, kInputStride, out, kOutputStride, 486 NULL, 0, NULL, 0, Width(), Height())); 487 488 CheckGuardBlocks(); 489 490 for (int y = 0; y < Height(); ++y) 491 for (int x = 0; x < Width(); ++x) 492 ASSERT_EQ(lookup(out, y * kOutputStride + x), 493 lookup(in, y * kInputStride + x)) 494 << "(" << x << "," << y << ")"; 495 } 496 497 const int kNumFilterBanks = SWITCHABLE_FILTERS; 498 const int kNumFilters = 16; 499 500 TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) { 501 int subpel_search; 502 for (subpel_search = USE_4_TAPS; subpel_search <= USE_8_TAPS; 503 ++subpel_search) { 504 for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) { 505 const InterpFilter filter = (InterpFilter)filter_bank; 506 const InterpKernel *filters = 507 (const InterpKernel *)av1_get_interp_filter_kernel(filter, 508 subpel_search); 509 for (int i = 0; i < kNumFilters; i++) { 510 const int p0 = filters[i][0] + filters[i][1]; 511 const int p1 = filters[i][2] + filters[i][3]; 512 const int p2 = filters[i][4] + filters[i][5]; 513 const int p3 = filters[i][6] + filters[i][7]; 514 EXPECT_LE(p0, 128); 515 EXPECT_LE(p1, 128); 516 EXPECT_LE(p2, 128); 517 EXPECT_LE(p3, 128); 518 EXPECT_LE(p0 + p3, 128); 519 EXPECT_LE(p0 + p3 + p1, 128); 520 EXPECT_LE(p0 + p3 + p1 + p2, 128); 521 EXPECT_EQ(p0 + p1 + p2 + p3, 128); 522 } 523 } 524 } 525 } 526 527 const int16_t kInvalidFilter[8] = { 0 }; 528 529 TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) { 530 uint8_t *const in = input(); 531 uint8_t *const out = output(); 532 uint8_t *ref; 533 if (UUT_->use_highbd_ == 0) { 534 ref = ref8_; 535 } else { 536 ref = CONVERT_TO_BYTEPTR(ref16_); 537 } 538 int subpel_search; 539 for (subpel_search = USE_4_TAPS; subpel_search <= USE_8_TAPS; 540 ++subpel_search) { 541 for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) { 542 const InterpFilter filter = (InterpFilter)filter_bank; 543 const InterpKernel *filters = 544 (const InterpKernel *)av1_get_interp_filter_kernel(filter, 545 subpel_search); 546 for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) { 547 for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) { 548 wrapper_filter_block2d_8_c(in, kInputStride, filters[filter_x], 549 filters[filter_y], ref, kOutputStride, 550 Width(), Height()); 551 552 if (filter_x && filter_y) 553 continue; 554 else if (filter_y) 555 ASM_REGISTER_STATE_CHECK( 556 UUT_->v8_(in, kInputStride, out, kOutputStride, kInvalidFilter, 557 16, filters[filter_y], 16, Width(), Height())); 558 else if (filter_x) 559 ASM_REGISTER_STATE_CHECK(UUT_->h8_( 560 in, kInputStride, out, kOutputStride, filters[filter_x], 16, 561 kInvalidFilter, 16, Width(), Height())); 562 else 563 ASM_REGISTER_STATE_CHECK(UUT_->copy_( 564 in, kInputStride, out, kOutputStride, kInvalidFilter, 0, 565 kInvalidFilter, 0, Width(), Height())); 566 567 CheckGuardBlocks(); 568 569 for (int y = 0; y < Height(); ++y) 570 for (int x = 0; x < Width(); ++x) 571 ASSERT_EQ(lookup(ref, y * kOutputStride + x), 572 lookup(out, y * kOutputStride + x)) 573 << "mismatch at (" << x << "," << y << "), " 574 << "filters (" << filter_bank << "," << filter_x << "," 575 << filter_y << ")"; 576 } 577 } 578 } 579 } 580 } 581 582 TEST_P(ConvolveTest, FilterExtremes) { 583 uint8_t *const in = input(); 584 uint8_t *const out = output(); 585 uint8_t *ref; 586 if (UUT_->use_highbd_ == 0) { 587 ref = ref8_; 588 } else { 589 ref = CONVERT_TO_BYTEPTR(ref16_); 590 } 591 592 // Populate ref and out with some random data 593 ::libaom_test::ACMRandom prng; 594 for (int y = 0; y < Height(); ++y) { 595 for (int x = 0; x < Width(); ++x) { 596 uint16_t r; 597 if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) { 598 r = prng.Rand8Extremes(); 599 } else { 600 r = prng.Rand16() & mask_; 601 } 602 assign_val(out, y * kOutputStride + x, r); 603 assign_val(ref, y * kOutputStride + x, r); 604 } 605 } 606 607 for (int axis = 0; axis < 2; axis++) { 608 int seed_val = 0; 609 while (seed_val < 256) { 610 for (int y = 0; y < 8; ++y) { 611 for (int x = 0; x < 8; ++x) { 612 assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1, 613 ((seed_val >> (axis ? y : x)) & 1) * mask_); 614 if (axis) seed_val++; 615 } 616 if (axis) 617 seed_val -= 8; 618 else 619 seed_val++; 620 } 621 if (axis) seed_val += 8; 622 int subpel_search; 623 for (subpel_search = USE_4_TAPS; subpel_search <= USE_8_TAPS; 624 ++subpel_search) { 625 for (int filter_bank = 0; filter_bank < kNumFilterBanks; 626 ++filter_bank) { 627 const InterpFilter filter = (InterpFilter)filter_bank; 628 const InterpKernel *filters = 629 (const InterpKernel *)av1_get_interp_filter_kernel(filter, 630 subpel_search); 631 for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) { 632 for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) { 633 wrapper_filter_block2d_8_c(in, kInputStride, filters[filter_x], 634 filters[filter_y], ref, kOutputStride, 635 Width(), Height()); 636 if (filter_x && filter_y) 637 continue; 638 else if (filter_y) 639 ASM_REGISTER_STATE_CHECK(UUT_->v8_( 640 in, kInputStride, out, kOutputStride, kInvalidFilter, 16, 641 filters[filter_y], 16, Width(), Height())); 642 else if (filter_x) 643 ASM_REGISTER_STATE_CHECK(UUT_->h8_( 644 in, kInputStride, out, kOutputStride, filters[filter_x], 16, 645 kInvalidFilter, 16, Width(), Height())); 646 else 647 ASM_REGISTER_STATE_CHECK(UUT_->copy_( 648 in, kInputStride, out, kOutputStride, kInvalidFilter, 0, 649 kInvalidFilter, 0, Width(), Height())); 650 651 for (int y = 0; y < Height(); ++y) 652 for (int x = 0; x < Width(); ++x) 653 ASSERT_EQ(lookup(ref, y * kOutputStride + x), 654 lookup(out, y * kOutputStride + x)) 655 << "mismatch at (" << x << "," << y << "), " 656 << "filters (" << filter_bank << "," << filter_x << "," 657 << filter_y << ")"; 658 } 659 } 660 } 661 } 662 } 663 } 664 } 665 666 TEST_P(ConvolveTest, DISABLED_Copy_Speed) { 667 const uint8_t *const in = input(); 668 uint8_t *const out = output(); 669 const int kNumTests = 5000000; 670 const int width = Width(); 671 const int height = Height(); 672 aom_usec_timer timer; 673 674 aom_usec_timer_start(&timer); 675 for (int n = 0; n < kNumTests; ++n) { 676 UUT_->copy_(in, kInputStride, out, kOutputStride, NULL, 0, NULL, 0, width, 677 height); 678 } 679 aom_usec_timer_mark(&timer); 680 681 const int elapsed_time = static_cast<int>(aom_usec_timer_elapsed(&timer)); 682 printf("convolve_copy_%dx%d_%d: %d us\n", width, height, 683 UUT_->use_highbd_ ? UUT_->use_highbd_ : 8, elapsed_time); 684 } 685 686 TEST_P(ConvolveTest, DISABLED_Speed) { 687 uint8_t *const in = input(); 688 uint8_t *const out = output(); 689 uint8_t *ref; 690 if (UUT_->use_highbd_ == 0) { 691 ref = ref8_; 692 } else { 693 ref = CONVERT_TO_BYTEPTR(ref16_); 694 } 695 696 // Populate ref and out with some random data 697 ::libaom_test::ACMRandom prng; 698 for (int y = 0; y < Height(); ++y) { 699 for (int x = 0; x < Width(); ++x) { 700 uint16_t r; 701 if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) { 702 r = prng.Rand8Extremes(); 703 } else { 704 r = prng.Rand16() & mask_; 705 } 706 assign_val(out, y * kOutputStride + x, r); 707 assign_val(ref, y * kOutputStride + x, r); 708 } 709 } 710 711 const InterpFilter filter = (InterpFilter)1; 712 const InterpKernel *filters = 713 (const InterpKernel *)av1_get_interp_filter_kernel(filter, USE_8_TAPS); 714 wrapper_filter_average_block2d_8_c(in, kInputStride, filters[1], filters[1], 715 out, kOutputStride, Width(), Height()); 716 717 aom_usec_timer timer; 718 int tests_num = 1000; 719 720 aom_usec_timer_start(&timer); 721 while (tests_num > 0) { 722 for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) { 723 const InterpFilter filter = (InterpFilter)filter_bank; 724 const InterpKernel *filters = 725 (const InterpKernel *)av1_get_interp_filter_kernel(filter, 726 USE_8_TAPS); 727 for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) { 728 for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) { 729 if (filter_x && filter_y) continue; 730 if (filter_y) 731 ASM_REGISTER_STATE_CHECK( 732 UUT_->v8_(in, kInputStride, out, kOutputStride, kInvalidFilter, 733 16, filters[filter_y], 16, Width(), Height())); 734 else if (filter_x) 735 ASM_REGISTER_STATE_CHECK(UUT_->h8_( 736 in, kInputStride, out, kOutputStride, filters[filter_x], 16, 737 kInvalidFilter, 16, Width(), Height())); 738 } 739 } 740 } 741 tests_num--; 742 } 743 aom_usec_timer_mark(&timer); 744 745 const int elapsed_time = 746 static_cast<int>(aom_usec_timer_elapsed(&timer) / 1000); 747 printf("%dx%d (bitdepth %d) time: %5d ms\n", Width(), Height(), 748 UUT_->use_highbd_, elapsed_time); 749 } 750 751 using ::testing::make_tuple; 752 753 #define WRAP(func, bd) \ 754 static void wrap_##func##_##bd( \ 755 const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, \ 756 ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride, \ 757 const int16_t *filter_y, int filter_y_stride, int w, int h) { \ 758 aom_highbd_##func(src, src_stride, dst, dst_stride, filter_x, \ 759 filter_x_stride, filter_y, filter_y_stride, w, h, bd); \ 760 } 761 #if HAVE_SSE2 && ARCH_X86_64 762 WRAP(convolve_copy_sse2, 8) 763 WRAP(convolve_copy_sse2, 10) 764 WRAP(convolve_copy_sse2, 12) 765 WRAP(convolve8_horiz_sse2, 8) 766 WRAP(convolve8_vert_sse2, 8) 767 WRAP(convolve8_horiz_sse2, 10) 768 WRAP(convolve8_vert_sse2, 10) 769 WRAP(convolve8_horiz_sse2, 12) 770 WRAP(convolve8_vert_sse2, 12) 771 #endif // HAVE_SSE2 && ARCH_X86_64 772 773 WRAP(convolve_copy_c, 8) 774 WRAP(convolve8_horiz_c, 8) 775 WRAP(convolve8_vert_c, 8) 776 WRAP(convolve_copy_c, 10) 777 WRAP(convolve8_horiz_c, 10) 778 WRAP(convolve8_vert_c, 10) 779 WRAP(convolve_copy_c, 12) 780 WRAP(convolve8_horiz_c, 12) 781 WRAP(convolve8_vert_c, 12) 782 783 #if HAVE_AVX2 784 WRAP(convolve_copy_avx2, 8) 785 WRAP(convolve8_horiz_avx2, 8) 786 WRAP(convolve8_vert_avx2, 8) 787 788 WRAP(convolve_copy_avx2, 10) 789 WRAP(convolve8_horiz_avx2, 10) 790 WRAP(convolve8_vert_avx2, 10) 791 792 WRAP(convolve_copy_avx2, 12) 793 WRAP(convolve8_horiz_avx2, 12) 794 WRAP(convolve8_vert_avx2, 12) 795 #endif // HAVE_AVX2 796 797 #undef WRAP 798 799 const ConvolveFunctions convolve8_c(wrap_convolve_copy_c_8, 800 wrap_convolve8_horiz_c_8, 801 wrap_convolve8_vert_c_8, 8); 802 const ConvolveFunctions convolve10_c(wrap_convolve_copy_c_10, 803 wrap_convolve8_horiz_c_10, 804 wrap_convolve8_vert_c_10, 10); 805 const ConvolveFunctions convolve12_c(wrap_convolve_copy_c_12, 806 wrap_convolve8_horiz_c_12, 807 wrap_convolve8_vert_c_12, 12); 808 const ConvolveParam kArrayConvolve_c[] = { 809 ALL_SIZES(convolve8_c), ALL_SIZES(convolve10_c), ALL_SIZES(convolve12_c) 810 }; 811 812 INSTANTIATE_TEST_CASE_P(C, ConvolveTest, ::testing::ValuesIn(kArrayConvolve_c)); 813 814 #if HAVE_SSE2 && ARCH_X86_64 815 const ConvolveFunctions convolve8_sse2(aom_convolve_copy_c, 816 aom_convolve8_horiz_sse2, 817 aom_convolve8_vert_sse2, 0); 818 const ConvolveFunctions wrap_convolve8_sse2(wrap_convolve_copy_sse2_8, 819 wrap_convolve8_horiz_sse2_8, 820 wrap_convolve8_vert_sse2_8, 8); 821 const ConvolveFunctions wrap_convolve10_sse2(wrap_convolve_copy_sse2_10, 822 wrap_convolve8_horiz_sse2_10, 823 wrap_convolve8_vert_sse2_10, 10); 824 const ConvolveFunctions wrap_convolve12_sse2(wrap_convolve_copy_sse2_12, 825 wrap_convolve8_horiz_sse2_12, 826 wrap_convolve8_vert_sse2_12, 12); 827 const ConvolveParam kArrayConvolve_sse2[] = { ALL_SIZES(convolve8_sse2), 828 ALL_SIZES(wrap_convolve8_sse2), 829 ALL_SIZES(wrap_convolve10_sse2), 830 ALL_SIZES(wrap_convolve12_sse2) }; 831 INSTANTIATE_TEST_CASE_P(SSE2, ConvolveTest, 832 ::testing::ValuesIn(kArrayConvolve_sse2)); 833 #endif 834 835 #if HAVE_SSSE3 836 const ConvolveFunctions convolve8_ssse3(aom_convolve_copy_c, 837 aom_convolve8_horiz_ssse3, 838 aom_convolve8_vert_ssse3, 0); 839 840 const ConvolveParam kArrayConvolve8_ssse3[] = { ALL_SIZES(convolve8_ssse3) }; 841 INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, 842 ::testing::ValuesIn(kArrayConvolve8_ssse3)); 843 #endif 844 845 #if HAVE_AVX2 846 const ConvolveFunctions convolve8_avx2(aom_convolve_copy_c, 847 aom_convolve8_horiz_avx2, 848 aom_convolve8_vert_avx2, 0); 849 850 const ConvolveFunctions wrap_convolve8_avx2(wrap_convolve_copy_avx2_8, 851 wrap_convolve8_horiz_avx2_8, 852 wrap_convolve8_vert_avx2_8, 8); 853 const ConvolveFunctions wrap_convolve10_avx2(wrap_convolve_copy_avx2_10, 854 wrap_convolve8_horiz_avx2_10, 855 wrap_convolve8_vert_avx2_10, 10); 856 const ConvolveFunctions wrap_convolve12_avx2(wrap_convolve_copy_avx2_12, 857 wrap_convolve8_horiz_avx2_12, 858 wrap_convolve8_vert_avx2_12, 12); 859 const ConvolveParam kArray_Convolve8_avx2[] = { 860 ALL_SIZES_64(wrap_convolve8_avx2), ALL_SIZES_64(wrap_convolve10_avx2), 861 ALL_SIZES_64(wrap_convolve12_avx2), ALL_SIZES(convolve8_avx2) 862 }; 863 INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest, 864 ::testing::ValuesIn(kArray_Convolve8_avx2)); 865 #endif // HAVE_AVX2 866 867 } // namespace 868
