1/* 2 * Copyright (c) 2012 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/clear_system_state.h" 22#include "test/register_state_check.h" 23#include "test/util.h" 24#include "vp9/common/vp9_entropy.h" 25#include "vpx/vpx_codec.h" 26#include "vpx/vpx_integer.h" 27#include "vpx_ports/mem.h" 28#include "vpx_ports/msvc.h" // for round() 29 30using libvpx_test::ACMRandom; 31 32namespace { 33 34const int kNumCoeffs = 1024; 35const double kPi = 3.141592653589793238462643383279502884; 36void reference_32x32_dct_1d(const double in[32], double out[32]) { 37 const double kInvSqrt2 = 0.707106781186547524400844362104; 38 for (int k = 0; k < 32; k++) { 39 out[k] = 0.0; 40 for (int n = 0; n < 32; n++) { 41 out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 64.0); 42 } 43 if (k == 0) out[k] = out[k] * kInvSqrt2; 44 } 45} 46 47void reference_32x32_dct_2d(const int16_t input[kNumCoeffs], 48 double output[kNumCoeffs]) { 49 // First transform columns 50 for (int i = 0; i < 32; ++i) { 51 double temp_in[32], temp_out[32]; 52 for (int j = 0; j < 32; ++j) temp_in[j] = input[j * 32 + i]; 53 reference_32x32_dct_1d(temp_in, temp_out); 54 for (int j = 0; j < 32; ++j) output[j * 32 + i] = temp_out[j]; 55 } 56 // Then transform rows 57 for (int i = 0; i < 32; ++i) { 58 double temp_in[32], temp_out[32]; 59 for (int j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32]; 60 reference_32x32_dct_1d(temp_in, temp_out); 61 // Scale by some magic number 62 for (int j = 0; j < 32; ++j) output[j + i * 32] = temp_out[j] / 4; 63 } 64} 65 66typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride); 67typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride); 68 69typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, int, vpx_bit_depth_t> 70 Trans32x32Param; 71 72#if CONFIG_VP9_HIGHBITDEPTH 73void idct32x32_10(const tran_low_t *in, uint8_t *out, int stride) { 74 vpx_highbd_idct32x32_1024_add_c(in, CAST_TO_SHORTPTR(out), stride, 10); 75} 76 77void idct32x32_12(const tran_low_t *in, uint8_t *out, int stride) { 78 vpx_highbd_idct32x32_1024_add_c(in, CAST_TO_SHORTPTR(out), stride, 12); 79} 80#endif // CONFIG_VP9_HIGHBITDEPTH 81 82class Trans32x32Test : public ::testing::TestWithParam<Trans32x32Param> { 83 public: 84 virtual ~Trans32x32Test() {} 85 virtual void SetUp() { 86 fwd_txfm_ = GET_PARAM(0); 87 inv_txfm_ = GET_PARAM(1); 88 version_ = GET_PARAM(2); // 0: high precision forward transform 89 // 1: low precision version for rd loop 90 bit_depth_ = GET_PARAM(3); 91 mask_ = (1 << bit_depth_) - 1; 92 } 93 94 virtual void TearDown() { libvpx_test::ClearSystemState(); } 95 96 protected: 97 int version_; 98 vpx_bit_depth_t bit_depth_; 99 int mask_; 100 FwdTxfmFunc fwd_txfm_; 101 InvTxfmFunc inv_txfm_; 102}; 103 104TEST_P(Trans32x32Test, AccuracyCheck) { 105 ACMRandom rnd(ACMRandom::DeterministicSeed()); 106 uint32_t max_error = 0; 107 int64_t total_error = 0; 108 const int count_test_block = 10000; 109 DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]); 110 DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]); 111 DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]); 112 DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]); 113#if CONFIG_VP9_HIGHBITDEPTH 114 DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]); 115 DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]); 116#endif 117 118 for (int i = 0; i < count_test_block; ++i) { 119 // Initialize a test block with input range [-mask_, mask_]. 120 for (int j = 0; j < kNumCoeffs; ++j) { 121 if (bit_depth_ == VPX_BITS_8) { 122 src[j] = rnd.Rand8(); 123 dst[j] = rnd.Rand8(); 124 test_input_block[j] = src[j] - dst[j]; 125#if CONFIG_VP9_HIGHBITDEPTH 126 } else { 127 src16[j] = rnd.Rand16() & mask_; 128 dst16[j] = rnd.Rand16() & mask_; 129 test_input_block[j] = src16[j] - dst16[j]; 130#endif 131 } 132 } 133 134 ASM_REGISTER_STATE_CHECK(fwd_txfm_(test_input_block, test_temp_block, 32)); 135 if (bit_depth_ == VPX_BITS_8) { 136 ASM_REGISTER_STATE_CHECK(inv_txfm_(test_temp_block, dst, 32)); 137#if CONFIG_VP9_HIGHBITDEPTH 138 } else { 139 ASM_REGISTER_STATE_CHECK( 140 inv_txfm_(test_temp_block, CAST_TO_BYTEPTR(dst16), 32)); 141#endif 142 } 143 144 for (int j = 0; j < kNumCoeffs; ++j) { 145#if CONFIG_VP9_HIGHBITDEPTH 146 const int32_t diff = 147 bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j]; 148#else 149 const int32_t diff = dst[j] - src[j]; 150#endif 151 const uint32_t error = diff * diff; 152 if (max_error < error) max_error = error; 153 total_error += error; 154 } 155 } 156 157 if (version_ == 1) { 158 max_error /= 2; 159 total_error /= 45; 160 } 161 162 EXPECT_GE(1u << 2 * (bit_depth_ - 8), max_error) 163 << "Error: 32x32 FDCT/IDCT has an individual round-trip error > 1"; 164 165 EXPECT_GE(count_test_block << 2 * (bit_depth_ - 8), total_error) 166 << "Error: 32x32 FDCT/IDCT has average round-trip error > 1 per block"; 167} 168 169TEST_P(Trans32x32Test, CoeffCheck) { 170 ACMRandom rnd(ACMRandom::DeterministicSeed()); 171 const int count_test_block = 1000; 172 173 DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]); 174 DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]); 175 DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]); 176 177 for (int i = 0; i < count_test_block; ++i) { 178 for (int j = 0; j < kNumCoeffs; ++j) { 179 input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_); 180 } 181 182 const int stride = 32; 183 vpx_fdct32x32_c(input_block, output_ref_block, stride); 184 ASM_REGISTER_STATE_CHECK(fwd_txfm_(input_block, output_block, stride)); 185 186 if (version_ == 0) { 187 for (int j = 0; j < kNumCoeffs; ++j) 188 EXPECT_EQ(output_block[j], output_ref_block[j]) 189 << "Error: 32x32 FDCT versions have mismatched coefficients"; 190 } else { 191 for (int j = 0; j < kNumCoeffs; ++j) 192 EXPECT_GE(6, abs(output_block[j] - output_ref_block[j])) 193 << "Error: 32x32 FDCT rd has mismatched coefficients"; 194 } 195 } 196} 197 198TEST_P(Trans32x32Test, MemCheck) { 199 ACMRandom rnd(ACMRandom::DeterministicSeed()); 200 const int count_test_block = 2000; 201 202 DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]); 203 DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]); 204 DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]); 205 206 for (int i = 0; i < count_test_block; ++i) { 207 // Initialize a test block with input range [-mask_, mask_]. 208 for (int j = 0; j < kNumCoeffs; ++j) { 209 input_extreme_block[j] = rnd.Rand8() & 1 ? mask_ : -mask_; 210 } 211 if (i == 0) { 212 for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = mask_; 213 } else if (i == 1) { 214 for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = -mask_; 215 } 216 217 const int stride = 32; 218 vpx_fdct32x32_c(input_extreme_block, output_ref_block, stride); 219 ASM_REGISTER_STATE_CHECK( 220 fwd_txfm_(input_extreme_block, output_block, stride)); 221 222 // The minimum quant value is 4. 223 for (int j = 0; j < kNumCoeffs; ++j) { 224 if (version_ == 0) { 225 EXPECT_EQ(output_block[j], output_ref_block[j]) 226 << "Error: 32x32 FDCT versions have mismatched coefficients"; 227 } else { 228 EXPECT_GE(6, abs(output_block[j] - output_ref_block[j])) 229 << "Error: 32x32 FDCT rd has mismatched coefficients"; 230 } 231 EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_ref_block[j])) 232 << "Error: 32x32 FDCT C has coefficient larger than 4*DCT_MAX_VALUE"; 233 EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j])) 234 << "Error: 32x32 FDCT has coefficient larger than " 235 << "4*DCT_MAX_VALUE"; 236 } 237 } 238} 239 240TEST_P(Trans32x32Test, InverseAccuracy) { 241 ACMRandom rnd(ACMRandom::DeterministicSeed()); 242 const int count_test_block = 1000; 243 DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]); 244 DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]); 245 DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]); 246 DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]); 247#if CONFIG_VP9_HIGHBITDEPTH 248 DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]); 249 DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]); 250#endif 251 252 for (int i = 0; i < count_test_block; ++i) { 253 double out_r[kNumCoeffs]; 254 255 // Initialize a test block with input range [-255, 255] 256 for (int j = 0; j < kNumCoeffs; ++j) { 257 if (bit_depth_ == VPX_BITS_8) { 258 src[j] = rnd.Rand8(); 259 dst[j] = rnd.Rand8(); 260 in[j] = src[j] - dst[j]; 261#if CONFIG_VP9_HIGHBITDEPTH 262 } else { 263 src16[j] = rnd.Rand16() & mask_; 264 dst16[j] = rnd.Rand16() & mask_; 265 in[j] = src16[j] - dst16[j]; 266#endif 267 } 268 } 269 270 reference_32x32_dct_2d(in, out_r); 271 for (int j = 0; j < kNumCoeffs; ++j) { 272 coeff[j] = static_cast<tran_low_t>(round(out_r[j])); 273 } 274 if (bit_depth_ == VPX_BITS_8) { 275 ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, dst, 32)); 276#if CONFIG_VP9_HIGHBITDEPTH 277 } else { 278 ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, CAST_TO_BYTEPTR(dst16), 32)); 279#endif 280 } 281 for (int j = 0; j < kNumCoeffs; ++j) { 282#if CONFIG_VP9_HIGHBITDEPTH 283 const int diff = 284 bit_depth_ == VPX_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j]; 285#else 286 const int diff = dst[j] - src[j]; 287#endif 288 const int error = diff * diff; 289 EXPECT_GE(1, error) << "Error: 32x32 IDCT has error " << error 290 << " at index " << j; 291 } 292 } 293} 294 295using std::tr1::make_tuple; 296 297#if CONFIG_VP9_HIGHBITDEPTH 298INSTANTIATE_TEST_CASE_P( 299 C, Trans32x32Test, 300 ::testing::Values( 301 make_tuple(&vpx_highbd_fdct32x32_c, &idct32x32_10, 0, VPX_BITS_10), 302 make_tuple(&vpx_highbd_fdct32x32_rd_c, &idct32x32_10, 1, VPX_BITS_10), 303 make_tuple(&vpx_highbd_fdct32x32_c, &idct32x32_12, 0, VPX_BITS_12), 304 make_tuple(&vpx_highbd_fdct32x32_rd_c, &idct32x32_12, 1, VPX_BITS_12), 305 make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c, 0, VPX_BITS_8), 306 make_tuple(&vpx_fdct32x32_rd_c, &vpx_idct32x32_1024_add_c, 1, 307 VPX_BITS_8))); 308#else 309INSTANTIATE_TEST_CASE_P( 310 C, Trans32x32Test, 311 ::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_c, 0, 312 VPX_BITS_8), 313 make_tuple(&vpx_fdct32x32_rd_c, &vpx_idct32x32_1024_add_c, 314 1, VPX_BITS_8))); 315#endif // CONFIG_VP9_HIGHBITDEPTH 316 317#if HAVE_NEON && !CONFIG_EMULATE_HARDWARE 318INSTANTIATE_TEST_CASE_P( 319 NEON, Trans32x32Test, 320 ::testing::Values(make_tuple(&vpx_fdct32x32_neon, 321 &vpx_idct32x32_1024_add_neon, 0, VPX_BITS_8), 322 make_tuple(&vpx_fdct32x32_rd_neon, 323 &vpx_idct32x32_1024_add_neon, 1, VPX_BITS_8))); 324#endif // HAVE_NEON && !CONFIG_EMULATE_HARDWARE 325 326#if HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 327INSTANTIATE_TEST_CASE_P( 328 SSE2, Trans32x32Test, 329 ::testing::Values(make_tuple(&vpx_fdct32x32_sse2, 330 &vpx_idct32x32_1024_add_sse2, 0, VPX_BITS_8), 331 make_tuple(&vpx_fdct32x32_rd_sse2, 332 &vpx_idct32x32_1024_add_sse2, 1, VPX_BITS_8))); 333#endif // HAVE_SSE2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 334 335#if HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 336INSTANTIATE_TEST_CASE_P( 337 SSE2, Trans32x32Test, 338 ::testing::Values( 339 make_tuple(&vpx_highbd_fdct32x32_sse2, &idct32x32_10, 0, VPX_BITS_10), 340 make_tuple(&vpx_highbd_fdct32x32_rd_sse2, &idct32x32_10, 1, 341 VPX_BITS_10), 342 make_tuple(&vpx_highbd_fdct32x32_sse2, &idct32x32_12, 0, VPX_BITS_12), 343 make_tuple(&vpx_highbd_fdct32x32_rd_sse2, &idct32x32_12, 1, 344 VPX_BITS_12), 345 make_tuple(&vpx_fdct32x32_sse2, &vpx_idct32x32_1024_add_c, 0, 346 VPX_BITS_8), 347 make_tuple(&vpx_fdct32x32_rd_sse2, &vpx_idct32x32_1024_add_c, 1, 348 VPX_BITS_8))); 349#endif // HAVE_SSE2 && CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 350 351#if HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 352INSTANTIATE_TEST_CASE_P( 353 AVX2, Trans32x32Test, 354 ::testing::Values(make_tuple(&vpx_fdct32x32_avx2, 355 &vpx_idct32x32_1024_add_sse2, 0, VPX_BITS_8), 356 make_tuple(&vpx_fdct32x32_rd_avx2, 357 &vpx_idct32x32_1024_add_sse2, 1, VPX_BITS_8))); 358#endif // HAVE_AVX2 && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 359 360#if HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 361INSTANTIATE_TEST_CASE_P( 362 MSA, Trans32x32Test, 363 ::testing::Values(make_tuple(&vpx_fdct32x32_msa, 364 &vpx_idct32x32_1024_add_msa, 0, VPX_BITS_8), 365 make_tuple(&vpx_fdct32x32_rd_msa, 366 &vpx_idct32x32_1024_add_msa, 1, VPX_BITS_8))); 367#endif // HAVE_MSA && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 368 369#if HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 370INSTANTIATE_TEST_CASE_P( 371 VSX, Trans32x32Test, 372 ::testing::Values(make_tuple(&vpx_fdct32x32_c, &vpx_idct32x32_1024_add_vsx, 373 0, VPX_BITS_8), 374 make_tuple(&vpx_fdct32x32_rd_c, 375 &vpx_idct32x32_1024_add_vsx, 1, VPX_BITS_8))); 376#endif // HAVE_VSX && !CONFIG_VP9_HIGHBITDEPTH && !CONFIG_EMULATE_HARDWARE 377} // namespace 378