1 /* 2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include <math.h> 12 #include <stdlib.h> 13 #include <string.h> 14 15 #include "third_party/googletest/src/include/gtest/gtest.h" 16 17 #include "./vp9_rtcd.h" 18 #include "./vpx_config.h" 19 #include "./vpx_dsp_rtcd.h" 20 #include "test/acm_random.h" 21 #include "test/buffer.h" 22 #include "test/clear_system_state.h" 23 #include "test/register_state_check.h" 24 #include "test/util.h" 25 #include "vp9/common/vp9_entropy.h" 26 #include "vp9/common/vp9_scan.h" 27 #include "vpx/vpx_codec.h" 28 #include "vpx/vpx_integer.h" 29 #include "vpx_ports/vpx_timer.h" 30 31 using libvpx_test::ACMRandom; 32 using libvpx_test::Buffer; 33 34 namespace { 35 const int number_of_iterations = 100; 36 37 typedef void (*QuantizeFunc)(const tran_low_t *coeff, intptr_t count, 38 int skip_block, const int16_t *zbin, 39 const int16_t *round, const int16_t *quant, 40 const int16_t *quant_shift, tran_low_t *qcoeff, 41 tran_low_t *dqcoeff, const int16_t *dequant, 42 uint16_t *eob, const int16_t *scan, 43 const int16_t *iscan); 44 typedef std::tr1::tuple<QuantizeFunc, QuantizeFunc, vpx_bit_depth_t, 45 int /*max_size*/, bool /*is_fp*/> 46 QuantizeParam; 47 48 // Wrapper for FP version which does not use zbin or quant_shift. 49 typedef void (*QuantizeFPFunc)(const tran_low_t *coeff, intptr_t count, 50 int skip_block, const int16_t *round, 51 const int16_t *quant, tran_low_t *qcoeff, 52 tran_low_t *dqcoeff, const int16_t *dequant, 53 uint16_t *eob, const int16_t *scan, 54 const int16_t *iscan); 55 56 template <QuantizeFPFunc fn> 57 void QuantFPWrapper(const tran_low_t *coeff, intptr_t count, int skip_block, 58 const int16_t *zbin, const int16_t *round, 59 const int16_t *quant, const int16_t *quant_shift, 60 tran_low_t *qcoeff, tran_low_t *dqcoeff, 61 const int16_t *dequant, uint16_t *eob, const int16_t *scan, 62 const int16_t *iscan) { 63 (void)zbin; 64 (void)quant_shift; 65 66 fn(coeff, count, skip_block, round, quant, qcoeff, dqcoeff, dequant, eob, 67 scan, iscan); 68 } 69 70 class VP9QuantizeBase { 71 public: 72 VP9QuantizeBase(vpx_bit_depth_t bit_depth, int max_size, bool is_fp) 73 : bit_depth_(bit_depth), max_size_(max_size), is_fp_(is_fp) { 74 max_value_ = (1 << bit_depth_) - 1; 75 zbin_ptr_ = 76 reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*zbin_ptr_))); 77 round_fp_ptr_ = reinterpret_cast<int16_t *>( 78 vpx_memalign(16, 8 * sizeof(*round_fp_ptr_))); 79 quant_fp_ptr_ = reinterpret_cast<int16_t *>( 80 vpx_memalign(16, 8 * sizeof(*quant_fp_ptr_))); 81 round_ptr_ = 82 reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*round_ptr_))); 83 quant_ptr_ = 84 reinterpret_cast<int16_t *>(vpx_memalign(16, 8 * sizeof(*quant_ptr_))); 85 quant_shift_ptr_ = reinterpret_cast<int16_t *>( 86 vpx_memalign(16, 8 * sizeof(*quant_shift_ptr_))); 87 dequant_ptr_ = reinterpret_cast<int16_t *>( 88 vpx_memalign(16, 8 * sizeof(*dequant_ptr_))); 89 } 90 91 ~VP9QuantizeBase() { 92 vpx_free(zbin_ptr_); 93 vpx_free(round_fp_ptr_); 94 vpx_free(quant_fp_ptr_); 95 vpx_free(round_ptr_); 96 vpx_free(quant_ptr_); 97 vpx_free(quant_shift_ptr_); 98 vpx_free(dequant_ptr_); 99 zbin_ptr_ = NULL; 100 round_fp_ptr_ = NULL; 101 quant_fp_ptr_ = NULL; 102 round_ptr_ = NULL; 103 quant_ptr_ = NULL; 104 quant_shift_ptr_ = NULL; 105 dequant_ptr_ = NULL; 106 libvpx_test::ClearSystemState(); 107 } 108 109 protected: 110 int16_t *zbin_ptr_; 111 int16_t *round_fp_ptr_; 112 int16_t *quant_fp_ptr_; 113 int16_t *round_ptr_; 114 int16_t *quant_ptr_; 115 int16_t *quant_shift_ptr_; 116 int16_t *dequant_ptr_; 117 const vpx_bit_depth_t bit_depth_; 118 int max_value_; 119 const int max_size_; 120 const bool is_fp_; 121 }; 122 123 class VP9QuantizeTest : public VP9QuantizeBase, 124 public ::testing::TestWithParam<QuantizeParam> { 125 public: 126 VP9QuantizeTest() 127 : VP9QuantizeBase(GET_PARAM(2), GET_PARAM(3), GET_PARAM(4)), 128 quantize_op_(GET_PARAM(0)), ref_quantize_op_(GET_PARAM(1)) {} 129 130 protected: 131 const QuantizeFunc quantize_op_; 132 const QuantizeFunc ref_quantize_op_; 133 }; 134 135 // This quantizer compares the AC coefficients to the quantization step size to 136 // determine if further multiplication operations are needed. 137 // Based on vp9_quantize_fp_sse2(). 138 void quantize_fp_nz_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, 139 int skip_block, const int16_t *round_ptr, 140 const int16_t *quant_ptr, tran_low_t *qcoeff_ptr, 141 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, 142 uint16_t *eob_ptr, const int16_t *scan, 143 const int16_t *iscan) { 144 int i, eob = -1; 145 const int thr = dequant_ptr[1] >> 1; 146 (void)iscan; 147 (void)skip_block; 148 assert(!skip_block); 149 150 // Quantization pass: All coefficients with index >= zero_flag are 151 // skippable. Note: zero_flag can be zero. 152 for (i = 0; i < n_coeffs; i += 16) { 153 int y; 154 int nzflag_cnt = 0; 155 int abs_coeff[16]; 156 int coeff_sign[16]; 157 158 // count nzflag for each row (16 tran_low_t) 159 for (y = 0; y < 16; ++y) { 160 const int rc = i + y; 161 const int coeff = coeff_ptr[rc]; 162 coeff_sign[y] = (coeff >> 31); 163 abs_coeff[y] = (coeff ^ coeff_sign[y]) - coeff_sign[y]; 164 // The first 16 are skipped in the sse2 code. Do the same here to match. 165 if (i >= 16 && (abs_coeff[y] <= thr)) { 166 nzflag_cnt++; 167 } 168 } 169 170 for (y = 0; y < 16; ++y) { 171 const int rc = i + y; 172 // If all of the AC coeffs in a row has magnitude less than the 173 // quantization step_size/2, quantize to zero. 174 if (nzflag_cnt < 16) { 175 int tmp = 176 clamp(abs_coeff[y] + round_ptr[rc != 0], INT16_MIN, INT16_MAX); 177 tmp = (tmp * quant_ptr[rc != 0]) >> 16; 178 qcoeff_ptr[rc] = (tmp ^ coeff_sign[y]) - coeff_sign[y]; 179 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0]; 180 } else { 181 qcoeff_ptr[rc] = 0; 182 dqcoeff_ptr[rc] = 0; 183 } 184 } 185 } 186 187 // Scan for eob. 188 for (i = 0; i < n_coeffs; i++) { 189 // Use the scan order to find the correct eob. 190 const int rc = scan[i]; 191 if (qcoeff_ptr[rc]) { 192 eob = i; 193 } 194 } 195 *eob_ptr = eob + 1; 196 } 197 198 void GenerateHelperArrays(ACMRandom *rnd, int16_t *zbin, int16_t *round, 199 int16_t *quant, int16_t *quant_shift, 200 int16_t *dequant, int16_t *round_fp, 201 int16_t *quant_fp) { 202 // Max when q == 0. Otherwise, it is 48 for Y and 42 for U/V. 203 const int max_qrounding_factor_fp = 64; 204 205 for (int j = 0; j < 2; j++) { 206 // The range is 4 to 1828 in the VP9 tables. 207 const int qlookup = rnd->RandRange(1825) + 4; 208 round_fp[j] = (max_qrounding_factor_fp * qlookup) >> 7; 209 quant_fp[j] = (1 << 16) / qlookup; 210 211 // Values determined by deconstructing vp9_init_quantizer(). 212 // zbin may be up to 1143 for 8 and 10 bit Y values, or 1200 for 12 bit Y 213 // values or U/V values of any bit depth. This is because y_delta is not 214 // factored into the vp9_ac_quant() call. 215 zbin[j] = rnd->RandRange(1200); 216 217 // round may be up to 685 for Y values or 914 for U/V. 218 round[j] = rnd->RandRange(914); 219 // quant ranges from 1 to -32703 220 quant[j] = static_cast<int>(rnd->RandRange(32704)) - 32703; 221 // quant_shift goes up to 1 << 16. 222 quant_shift[j] = rnd->RandRange(16384); 223 // dequant maxes out at 1828 for all cases. 224 dequant[j] = rnd->RandRange(1828); 225 } 226 for (int j = 2; j < 8; j++) { 227 zbin[j] = zbin[1]; 228 round_fp[j] = round_fp[1]; 229 quant_fp[j] = quant_fp[1]; 230 round[j] = round[1]; 231 quant[j] = quant[1]; 232 quant_shift[j] = quant_shift[1]; 233 dequant[j] = dequant[1]; 234 } 235 } 236 237 TEST_P(VP9QuantizeTest, OperationCheck) { 238 ACMRandom rnd(ACMRandom::DeterministicSeed()); 239 Buffer<tran_low_t> coeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 16); 240 ASSERT_TRUE(coeff.Init()); 241 Buffer<tran_low_t> qcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 242 ASSERT_TRUE(qcoeff.Init()); 243 Buffer<tran_low_t> dqcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 244 ASSERT_TRUE(dqcoeff.Init()); 245 Buffer<tran_low_t> ref_qcoeff = 246 Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 247 ASSERT_TRUE(ref_qcoeff.Init()); 248 Buffer<tran_low_t> ref_dqcoeff = 249 Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 250 ASSERT_TRUE(ref_dqcoeff.Init()); 251 uint16_t eob, ref_eob; 252 253 for (int i = 0; i < number_of_iterations; ++i) { 254 // Test skip block for the first three iterations to catch all the different 255 // sizes. 256 const int skip_block = 0; 257 TX_SIZE sz; 258 if (max_size_ == 16) { 259 sz = static_cast<TX_SIZE>(i % 3); // TX_4X4, TX_8X8 TX_16X16 260 } else { 261 sz = TX_32X32; 262 } 263 const TX_TYPE tx_type = static_cast<TX_TYPE>((i >> 2) % 3); 264 const scan_order *scan_order = &vp9_scan_orders[sz][tx_type]; 265 const int count = (4 << sz) * (4 << sz); 266 coeff.Set(&rnd, -max_value_, max_value_); 267 GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_, 268 quant_shift_ptr_, dequant_ptr_, round_fp_ptr_, 269 quant_fp_ptr_); 270 int16_t *r_ptr = (is_fp_) ? round_fp_ptr_ : round_ptr_; 271 int16_t *q_ptr = (is_fp_) ? quant_fp_ptr_ : quant_ptr_; 272 ref_quantize_op_(coeff.TopLeftPixel(), count, skip_block, zbin_ptr_, r_ptr, 273 q_ptr, quant_shift_ptr_, ref_qcoeff.TopLeftPixel(), 274 ref_dqcoeff.TopLeftPixel(), dequant_ptr_, &ref_eob, 275 scan_order->scan, scan_order->iscan); 276 277 ASM_REGISTER_STATE_CHECK(quantize_op_( 278 coeff.TopLeftPixel(), count, skip_block, zbin_ptr_, r_ptr, q_ptr, 279 quant_shift_ptr_, qcoeff.TopLeftPixel(), dqcoeff.TopLeftPixel(), 280 dequant_ptr_, &eob, scan_order->scan, scan_order->iscan)); 281 282 EXPECT_TRUE(qcoeff.CheckValues(ref_qcoeff)); 283 EXPECT_TRUE(dqcoeff.CheckValues(ref_dqcoeff)); 284 285 EXPECT_EQ(eob, ref_eob); 286 287 if (HasFailure()) { 288 printf("Failure on iteration %d.\n", i); 289 qcoeff.PrintDifference(ref_qcoeff); 290 dqcoeff.PrintDifference(ref_dqcoeff); 291 return; 292 } 293 } 294 } 295 296 TEST_P(VP9QuantizeTest, EOBCheck) { 297 ACMRandom rnd(ACMRandom::DeterministicSeed()); 298 Buffer<tran_low_t> coeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 16); 299 ASSERT_TRUE(coeff.Init()); 300 Buffer<tran_low_t> qcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 301 ASSERT_TRUE(qcoeff.Init()); 302 Buffer<tran_low_t> dqcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 303 ASSERT_TRUE(dqcoeff.Init()); 304 Buffer<tran_low_t> ref_qcoeff = 305 Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 306 ASSERT_TRUE(ref_qcoeff.Init()); 307 Buffer<tran_low_t> ref_dqcoeff = 308 Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 309 ASSERT_TRUE(ref_dqcoeff.Init()); 310 uint16_t eob, ref_eob; 311 312 for (int i = 0; i < number_of_iterations; ++i) { 313 const int skip_block = 0; 314 TX_SIZE sz; 315 if (max_size_ == 16) { 316 sz = static_cast<TX_SIZE>(i % 3); // TX_4X4, TX_8X8 TX_16X16 317 } else { 318 sz = TX_32X32; 319 } 320 const TX_TYPE tx_type = static_cast<TX_TYPE>((i >> 2) % 3); 321 const scan_order *scan_order = &vp9_scan_orders[sz][tx_type]; 322 int count = (4 << sz) * (4 << sz); 323 // Two random entries 324 coeff.Set(0); 325 coeff.TopLeftPixel()[rnd(count)] = 326 static_cast<int>(rnd.RandRange(max_value_ * 2)) - max_value_; 327 coeff.TopLeftPixel()[rnd(count)] = 328 static_cast<int>(rnd.RandRange(max_value_ * 2)) - max_value_; 329 GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_, 330 quant_shift_ptr_, dequant_ptr_, round_fp_ptr_, 331 quant_fp_ptr_); 332 int16_t *r_ptr = (is_fp_) ? round_fp_ptr_ : round_ptr_; 333 int16_t *q_ptr = (is_fp_) ? quant_fp_ptr_ : quant_ptr_; 334 ref_quantize_op_(coeff.TopLeftPixel(), count, skip_block, zbin_ptr_, r_ptr, 335 q_ptr, quant_shift_ptr_, ref_qcoeff.TopLeftPixel(), 336 ref_dqcoeff.TopLeftPixel(), dequant_ptr_, &ref_eob, 337 scan_order->scan, scan_order->iscan); 338 339 ASM_REGISTER_STATE_CHECK(quantize_op_( 340 coeff.TopLeftPixel(), count, skip_block, zbin_ptr_, r_ptr, q_ptr, 341 quant_shift_ptr_, qcoeff.TopLeftPixel(), dqcoeff.TopLeftPixel(), 342 dequant_ptr_, &eob, scan_order->scan, scan_order->iscan)); 343 344 EXPECT_TRUE(qcoeff.CheckValues(ref_qcoeff)); 345 EXPECT_TRUE(dqcoeff.CheckValues(ref_dqcoeff)); 346 347 EXPECT_EQ(eob, ref_eob); 348 349 if (HasFailure()) { 350 printf("Failure on iteration %d.\n", i); 351 qcoeff.PrintDifference(ref_qcoeff); 352 dqcoeff.PrintDifference(ref_dqcoeff); 353 return; 354 } 355 } 356 } 357 358 TEST_P(VP9QuantizeTest, DISABLED_Speed) { 359 ACMRandom rnd(ACMRandom::DeterministicSeed()); 360 Buffer<tran_low_t> coeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 16); 361 ASSERT_TRUE(coeff.Init()); 362 Buffer<tran_low_t> qcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 363 ASSERT_TRUE(qcoeff.Init()); 364 Buffer<tran_low_t> dqcoeff = Buffer<tran_low_t>(max_size_, max_size_, 0, 32); 365 ASSERT_TRUE(dqcoeff.Init()); 366 uint16_t eob; 367 TX_SIZE starting_sz, ending_sz; 368 369 if (max_size_ == 16) { 370 starting_sz = TX_4X4; 371 ending_sz = TX_16X16; 372 } else { 373 starting_sz = TX_32X32; 374 ending_sz = TX_32X32; 375 } 376 377 for (TX_SIZE sz = starting_sz; sz <= ending_sz; ++sz) { 378 // zbin > coeff, zbin < coeff. 379 for (int i = 0; i < 2; ++i) { 380 const int skip_block = 0; 381 // TX_TYPE defines the scan order. That is not relevant to the speed test. 382 // Pick the first one. 383 const TX_TYPE tx_type = DCT_DCT; 384 const scan_order *scan_order = &vp9_scan_orders[sz][tx_type]; 385 const int count = (4 << sz) * (4 << sz); 386 387 GenerateHelperArrays(&rnd, zbin_ptr_, round_ptr_, quant_ptr_, 388 quant_shift_ptr_, dequant_ptr_, round_fp_ptr_, 389 quant_fp_ptr_); 390 int16_t *r_ptr = (is_fp_) ? round_fp_ptr_ : round_ptr_; 391 int16_t *q_ptr = (is_fp_) ? quant_fp_ptr_ : quant_ptr_; 392 393 if (i == 0) { 394 // When |coeff values| are less than zbin the results are 0. 395 int threshold = 100; 396 if (max_size_ == 32) { 397 // For 32x32, the threshold is halved. Double it to keep the values 398 // from clearing it. 399 threshold = 200; 400 } 401 for (int j = 0; j < 8; ++j) zbin_ptr_[j] = threshold; 402 coeff.Set(&rnd, -99, 99); 403 } else if (i == 1) { 404 for (int j = 0; j < 8; ++j) zbin_ptr_[j] = 50; 405 coeff.Set(&rnd, -500, 500); 406 } 407 408 vpx_usec_timer timer; 409 vpx_usec_timer_start(&timer); 410 for (int j = 0; j < 100000000 / count; ++j) { 411 quantize_op_(coeff.TopLeftPixel(), count, skip_block, zbin_ptr_, r_ptr, 412 q_ptr, quant_shift_ptr_, qcoeff.TopLeftPixel(), 413 dqcoeff.TopLeftPixel(), dequant_ptr_, &eob, 414 scan_order->scan, scan_order->iscan); 415 } 416 vpx_usec_timer_mark(&timer); 417 const int elapsed_time = static_cast<int>(vpx_usec_timer_elapsed(&timer)); 418 if (i == 0) printf("Bypass calculations.\n"); 419 if (i == 1) printf("Full calculations.\n"); 420 printf("Quantize %dx%d time: %5d ms\n", 4 << sz, 4 << sz, 421 elapsed_time / 1000); 422 } 423 printf("\n"); 424 } 425 } 426 427 using std::tr1::make_tuple; 428 429 #if HAVE_SSE2 430 #if CONFIG_VP9_HIGHBITDEPTH 431 // TODO(johannkoenig): Fix vpx_quantize_b_sse2 in highbitdepth builds. 432 // make_tuple(&vpx_quantize_b_sse2, &vpx_highbd_quantize_b_c, VPX_BITS_8), 433 INSTANTIATE_TEST_CASE_P( 434 SSE2, VP9QuantizeTest, 435 ::testing::Values( 436 make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c, 437 VPX_BITS_8, 16, false), 438 make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c, 439 VPX_BITS_10, 16, false), 440 make_tuple(&vpx_highbd_quantize_b_sse2, &vpx_highbd_quantize_b_c, 441 VPX_BITS_12, 16, false), 442 make_tuple(&vpx_highbd_quantize_b_32x32_sse2, 443 &vpx_highbd_quantize_b_32x32_c, VPX_BITS_8, 32, false), 444 make_tuple(&vpx_highbd_quantize_b_32x32_sse2, 445 &vpx_highbd_quantize_b_32x32_c, VPX_BITS_10, 32, false), 446 make_tuple(&vpx_highbd_quantize_b_32x32_sse2, 447 &vpx_highbd_quantize_b_32x32_c, VPX_BITS_12, 32, false))); 448 449 #else 450 INSTANTIATE_TEST_CASE_P( 451 SSE2, VP9QuantizeTest, 452 ::testing::Values(make_tuple(&vpx_quantize_b_sse2, &vpx_quantize_b_c, 453 VPX_BITS_8, 16, false), 454 make_tuple(&QuantFPWrapper<vp9_quantize_fp_sse2>, 455 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 456 16, true))); 457 #endif // CONFIG_VP9_HIGHBITDEPTH 458 #endif // HAVE_SSE2 459 460 #if HAVE_SSSE3 && !CONFIG_VP9_HIGHBITDEPTH 461 #if ARCH_X86_64 462 INSTANTIATE_TEST_CASE_P( 463 SSSE3, VP9QuantizeTest, 464 ::testing::Values(make_tuple(&vpx_quantize_b_ssse3, &vpx_quantize_b_c, 465 VPX_BITS_8, 16, false), 466 make_tuple(&QuantFPWrapper<vp9_quantize_fp_ssse3>, 467 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 468 16, true))); 469 #else 470 INSTANTIATE_TEST_CASE_P(SSSE3, VP9QuantizeTest, 471 ::testing::Values(make_tuple(&vpx_quantize_b_ssse3, 472 &vpx_quantize_b_c, 473 VPX_BITS_8, 16, false))); 474 #endif 475 476 #if ARCH_X86_64 477 // TODO(johannkoenig): SSSE3 optimizations do not yet pass this test. 478 INSTANTIATE_TEST_CASE_P( 479 DISABLED_SSSE3, VP9QuantizeTest, 480 ::testing::Values(make_tuple(&vpx_quantize_b_32x32_ssse3, 481 &vpx_quantize_b_32x32_c, VPX_BITS_8, 32, 482 false), 483 make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_ssse3>, 484 &QuantFPWrapper<vp9_quantize_fp_32x32_c>, 485 VPX_BITS_8, 32, true))); 486 #endif // ARCH_X86_64 487 #endif // HAVE_SSSE3 && !CONFIG_VP9_HIGHBITDEPTH 488 489 // TODO(johannkoenig): AVX optimizations do not yet pass the 32x32 test or 490 // highbitdepth configurations. 491 #if HAVE_AVX && !CONFIG_VP9_HIGHBITDEPTH 492 INSTANTIATE_TEST_CASE_P( 493 AVX, VP9QuantizeTest, 494 ::testing::Values(make_tuple(&vpx_quantize_b_avx, &vpx_quantize_b_c, 495 VPX_BITS_8, 16, false), 496 // Even though SSSE3 and AVX do not match the reference 497 // code, we can keep them in sync with each other. 498 make_tuple(&vpx_quantize_b_32x32_avx, 499 &vpx_quantize_b_32x32_ssse3, VPX_BITS_8, 32, 500 false))); 501 #endif // HAVE_AVX && !CONFIG_VP9_HIGHBITDEPTH 502 503 // TODO(webm:1448): dqcoeff is not handled correctly in HBD builds. 504 #if HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH 505 INSTANTIATE_TEST_CASE_P( 506 NEON, VP9QuantizeTest, 507 ::testing::Values(make_tuple(&vpx_quantize_b_neon, &vpx_quantize_b_c, 508 VPX_BITS_8, 16, false), 509 make_tuple(&vpx_quantize_b_32x32_neon, 510 &vpx_quantize_b_32x32_c, VPX_BITS_8, 32, 511 false), 512 make_tuple(&QuantFPWrapper<vp9_quantize_fp_neon>, 513 &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8, 514 16, true), 515 make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_neon>, 516 &QuantFPWrapper<vp9_quantize_fp_32x32_c>, 517 VPX_BITS_8, 32, true))); 518 #endif // HAVE_NEON && !CONFIG_VP9_HIGHBITDEPTH 519 520 // Only useful to compare "Speed" test results. 521 INSTANTIATE_TEST_CASE_P( 522 DISABLED_C, VP9QuantizeTest, 523 ::testing::Values( 524 make_tuple(&vpx_quantize_b_c, &vpx_quantize_b_c, VPX_BITS_8, 16, false), 525 make_tuple(&vpx_quantize_b_32x32_c, &vpx_quantize_b_32x32_c, VPX_BITS_8, 526 32, false), 527 make_tuple(&QuantFPWrapper<vp9_quantize_fp_c>, 528 &QuantFPWrapper<vp9_quantize_fp_c>, VPX_BITS_8, 16, true), 529 make_tuple(&QuantFPWrapper<quantize_fp_nz_c>, 530 &QuantFPWrapper<quantize_fp_nz_c>, VPX_BITS_8, 16, true), 531 make_tuple(&QuantFPWrapper<vp9_quantize_fp_32x32_c>, 532 &QuantFPWrapper<vp9_quantize_fp_32x32_c>, VPX_BITS_8, 32, 533 true))); 534 } // namespace 535