test/dct16x16_test.cc

   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 #include "vpx_ports/mem.h"
  17
  18 extern "C" {
  19 #include "vp9/common/vp9_entropy.h"
  20 #include "vp9_rtcd.h"
  21 void vp9_short_idct16x16_add_c(short *input, uint8_t *output, int pitch);
  22 }
  23
  24 #include "acm_random.h"
  25 #include "vpx/vpx_integer.h"
  26
  27 using libvpx_test::ACMRandom;
  28
  29 namespace {
  30
  31 #ifdef _MSC_VER
  32 static int round(double x) {
  33   if (x < 0)
  34     return (int)ceil(x - 0.5);
  35   else
  36     return (int)floor(x + 0.5);
  37 }
  38 #endif
  39
  40 const double PI = 3.1415926535898;
  41 void reference2_16x16_idct_2d(double *input, double *output) {
  42   double x;
  43   for (int l = 0; l < 16; ++l) {
  44     for (int k = 0; k < 16; ++k) {
  45       double s = 0;
  46       for (int i = 0; i < 16; ++i) {
  47         for (int j = 0; j < 16; ++j) {
  48           x=cos(PI*j*(l+0.5)/16.0)*cos(PI*i*(k+0.5)/16.0)*input[i*16+j]/256;
  49           if (i != 0)
  50             x *= sqrt(2.0);
  51           if (j != 0)
  52             x *= sqrt(2.0);
  53           s += x;
  54         }
  55       }
  56       output[k*16+l] = s;
  57     }
  58   }
  59 }
  60
  61
  62 static const double C1 = 0.995184726672197;
  63 static const double C2 = 0.98078528040323;
  64 static const double C3 = 0.956940335732209;
  65 static const double C4 = 0.923879532511287;
  66 static const double C5 = 0.881921264348355;
  67 static const double C6 = 0.831469612302545;
  68 static const double C7 = 0.773010453362737;
  69 static const double C8 = 0.707106781186548;
  70 static const double C9 = 0.634393284163646;
  71 static const double C10 = 0.555570233019602;
  72 static const double C11 = 0.471396736825998;
  73 static const double C12 = 0.38268343236509;
  74 static const double C13 = 0.290284677254462;
  75 static const double C14 = 0.195090322016128;
  76 static const double C15 = 0.098017140329561;
  77
  78 static void butterfly_16x16_dct_1d(double input[16], double output[16]) {
  79   double step[16];
  80   double intermediate[16];
  81   double temp1, temp2;
  82
  83   // step 1
  84   step[ 0] = input[0] + input[15];
  85   step[ 1] = input[1] + input[14];
  86   step[ 2] = input[2] + input[13];
  87   step[ 3] = input[3] + input[12];
  88   step[ 4] = input[4] + input[11];
  89   step[ 5] = input[5] + input[10];
  90   step[ 6] = input[6] + input[ 9];
  91   step[ 7] = input[7] + input[ 8];
  92   step[ 8] = input[7] - input[ 8];
  93   step[ 9] = input[6] - input[ 9];
  94   step[10] = input[5] - input[10];
  95   step[11] = input[4] - input[11];
  96   step[12] = input[3] - input[12];
  97   step[13] = input[2] - input[13];
  98   step[14] = input[1] - input[14];
  99   step[15] = input[0] - input[15];
 100
 101   // step 2
 102   output[0] = step[0] + step[7];
 103   output[1] = step[1] + step[6];
 104   output[2] = step[2] + step[5];
 105   output[3] = step[3] + step[4];
 106   output[4] = step[3] - step[4];
 107   output[5] = step[2] - step[5];
 108   output[6] = step[1] - step[6];
 109   output[7] = step[0] - step[7];
 110
 111   temp1 = step[ 8]*C7;
 112   temp2 = step[15]*C9;
 113   output[ 8] = temp1 + temp2;
 114
 115   temp1 = step[ 9]*C11;
 116   temp2 = step[14]*C5;
 117   output[ 9] = temp1 - temp2;
 118
 119   temp1 = step[10]*C3;
 120   temp2 = step[13]*C13;
 121   output[10] = temp1 + temp2;
 122
 123   temp1 = step[11]*C15;
 124   temp2 = step[12]*C1;
 125   output[11] = temp1 - temp2;
 126
 127   temp1 = step[11]*C1;
 128   temp2 = step[12]*C15;
 129   output[12] = temp2 + temp1;
 130
 131   temp1 = step[10]*C13;
 132   temp2 = step[13]*C3;
 133   output[13] = temp2 - temp1;
 134
 135   temp1 = step[ 9]*C5;
 136   temp2 = step[14]*C11;
 137   output[14] = temp2 + temp1;
 138
 139   temp1 = step[ 8]*C9;
 140   temp2 = step[15]*C7;
 141   output[15] = temp2 - temp1;
 142
 143   // step 3
 144   step[ 0] = output[0] + output[3];
 145   step[ 1] = output[1] + output[2];
 146   step[ 2] = output[1] - output[2];
 147   step[ 3] = output[0] - output[3];
 148
 149   temp1 = output[4]*C14;
 150   temp2 = output[7]*C2;
 151   step[ 4] = temp1 + temp2;
 152
 153   temp1 = output[5]*C10;
 154   temp2 = output[6]*C6;
 155   step[ 5] = temp1 + temp2;
 156
 157   temp1 = output[5]*C6;
 158   temp2 = output[6]*C10;
 159   step[ 6] = temp2 - temp1;
 160
 161   temp1 = output[4]*C2;
 162   temp2 = output[7]*C14;
 163   step[ 7] = temp2 - temp1;
 164
 165   step[ 8] = output[ 8] + output[11];
 166   step[ 9] = output[ 9] + output[10];
 167   step[10] = output[ 9] - output[10];
 168   step[11] = output[ 8] - output[11];
 169
 170   step[12] = output[12] + output[15];
 171   step[13] = output[13] + output[14];
 172   step[14] = output[13] - output[14];
 173   step[15] = output[12] - output[15];
 174
 175   // step 4
 176   output[ 0] = (step[ 0] + step[ 1]);
 177   output[ 8] = (step[ 0] - step[ 1]);
 178
 179   temp1 = step[2]*C12;
 180   temp2 = step[3]*C4;
 181   temp1 = temp1 + temp2;
 182   output[ 4] = 2*(temp1*C8);
 183
 184   temp1 = step[2]*C4;
 185   temp2 = step[3]*C12;
 186   temp1 = temp2 - temp1;
 187   output[12] = 2*(temp1*C8);
 188
 189   output[ 2] = 2*((step[4] + step[ 5])*C8);
 190   output[14] = 2*((step[7] - step[ 6])*C8);
 191
 192   temp1 = step[4] - step[5];
 193   temp2 = step[6] + step[7];
 194   output[ 6] = (temp1 + temp2);
 195   output[10] = (temp1 - temp2);
 196
 197   intermediate[8] = step[8] + step[14];
 198   intermediate[9] = step[9] + step[15];
 199
 200   temp1 = intermediate[8]*C12;
 201   temp2 = intermediate[9]*C4;
 202   temp1 = temp1 - temp2;
 203   output[3] = 2*(temp1*C8);
 204
 205   temp1 = intermediate[8]*C4;
 206   temp2 = intermediate[9]*C12;
 207   temp1 = temp2 + temp1;
 208   output[13] = 2*(temp1*C8);
 209
 210   output[ 9] = 2*((step[10] + step[11])*C8);
 211
 212   intermediate[11] = step[10] - step[11];
 213   intermediate[12] = step[12] + step[13];
 214   intermediate[13] = step[12] - step[13];
 215   intermediate[14] = step[ 8] - step[14];
 216   intermediate[15] = step[ 9] - step[15];
 217
 218   output[15] = (intermediate[11] + intermediate[12]);
 219   output[ 1] = -(intermediate[11] - intermediate[12]);
 220
 221   output[ 7] = 2*(intermediate[13]*C8);
 222
 223   temp1 = intermediate[14]*C12;
 224   temp2 = intermediate[15]*C4;
 225   temp1 = temp1 - temp2;
 226   output[11] = -2*(temp1*C8);
 227
 228   temp1 = intermediate[14]*C4;
 229   temp2 = intermediate[15]*C12;
 230   temp1 = temp2 + temp1;
 231   output[ 5] = 2*(temp1*C8);
 232 }
 233
 234 static void reference_16x16_dct_1d(double in[16], double out[16]) {
 235   const double kPi = 3.141592653589793238462643383279502884;
 236   const double kInvSqrt2 = 0.707106781186547524400844362104;
 237   for (int k = 0; k < 16; k++) {
 238     out[k] = 0.0;
 239     for (int n = 0; n < 16; n++)
 240       out[k] += in[n]*cos(kPi*(2*n+1)*k/32.0);
 241     if (k == 0)
 242       out[k] = out[k]*kInvSqrt2;
 243   }
 244 }
 245
 246 void reference_16x16_dct_2d(int16_t input[16*16], double output[16*16]) {
 247   // First transform columns
 248   for (int i = 0; i < 16; ++i) {
 249     double temp_in[16], temp_out[16];
 250     for (int j = 0; j < 16; ++j)
 251       temp_in[j] = input[j*16 + i];
 252     butterfly_16x16_dct_1d(temp_in, temp_out);
 253     for (int j = 0; j < 16; ++j)
 254       output[j*16 + i] = temp_out[j];
 255   }
 256   // Then transform rows
 257   for (int i = 0; i < 16; ++i) {
 258     double temp_in[16], temp_out[16];
 259     for (int j = 0; j < 16; ++j)
 260       temp_in[j] = output[j + i*16];
 261     butterfly_16x16_dct_1d(temp_in, temp_out);
 262     // Scale by some magic number
 263     for (int j = 0; j < 16; ++j)
 264       output[j + i*16] = temp_out[j]/2;
 265   }
 266 }
 267
 268 void fdct16x16(int16_t *in, int16_t *out, uint8_t* /*dst*/,
 269                int stride, int /*tx_type*/) {
 270   vp9_short_fdct16x16_c(in, out, stride);
 271 }
 272 void idct16x16_add(int16_t* /*in*/, int16_t *out, uint8_t *dst,
 273                    int stride, int /*tx_type*/) {
 274   vp9_short_idct16x16_add_c(out, dst, stride >> 1);
 275 }
 276 void fht16x16(int16_t *in, int16_t *out, uint8_t* /*dst*/,
 277               int stride, int tx_type) {
 278   // FIXME(jingning): need to test both SSE2 and c
 279 #if HAVE_SSE2
 280   vp9_short_fht16x16_sse2(in, out, stride >> 1, tx_type);
 281 #else
 282   vp9_short_fht16x16_c(in, out, stride >> 1, tx_type);
 283 #endif
 284 }
 285 void iht16x16_add(int16_t* /*in*/, int16_t *out, uint8_t *dst,
 286                   int stride, int tx_type) {
 287   vp9_short_iht16x16_add_c(out, dst, stride >> 1, tx_type);
 288 }
 289
 290 class FwdTrans16x16Test : public ::testing::TestWithParam<int> {
 291  public:
 292   virtual ~FwdTrans16x16Test() {}
 293
 294   virtual void SetUp() {
 295     tx_type_ = GetParam();
 296     if (tx_type_ == 0) {
 297       fwd_txfm = fdct16x16;
 298       inv_txfm = idct16x16_add;
 299     } else {
 300       fwd_txfm = fht16x16;
 301       inv_txfm = iht16x16_add;
 302     }
 303   }
 304
 305  protected:
 306   void RunFwdTxfm(int16_t *in, int16_t *out, uint8_t *dst,
 307                   int stride, int tx_type) {
 308     (*fwd_txfm)(in, out, dst, stride, tx_type);
 309   }
 310   void RunInvTxfm(int16_t *in, int16_t *out, uint8_t *dst,
 311                   int stride, int tx_type) {
 312     (*inv_txfm)(in, out, dst, stride, tx_type);
 313   }
 314
 315   int tx_type_;
 316   void (*fwd_txfm)(int16_t*, int16_t*, uint8_t*, int, int);
 317   void (*inv_txfm)(int16_t*, int16_t*, uint8_t*, int, int);
 318 };
 319
 320 TEST_P(FwdTrans16x16Test, AccuracyCheck) {
 321   ACMRandom rnd(ACMRandom::DeterministicSeed());
 322   int max_error = 0;
 323   double total_error = 0;
 324   const int count_test_block = 10000;
 325   for (int i = 0; i < count_test_block; ++i) {
 326     DECLARE_ALIGNED_ARRAY(16, int16_t, test_input_block, 256);
 327     DECLARE_ALIGNED_ARRAY(16, int16_t, test_temp_block, 256);
 328     DECLARE_ALIGNED_ARRAY(16, uint8_t, dst, 256);
 329     DECLARE_ALIGNED_ARRAY(16, uint8_t, src, 256);
 330
 331     for (int j = 0; j < 256; ++j) {
 332       src[j] = rnd.Rand8();
 333       dst[j] = rnd.Rand8();
 334       // Initialize a test block with input range [-255, 255].
 335       test_input_block[j] = src[j] - dst[j];
 336     }
 337
 338     const int pitch = 32;
 339     RunFwdTxfm(test_input_block, test_temp_block, dst, pitch, tx_type_);
 340     RunInvTxfm(test_input_block, test_temp_block, dst, pitch, tx_type_);
 341
 342     for (int j = 0; j < 256; ++j) {
 343       const int diff = dst[j] - src[j];
 344       const int error = diff * diff;
 345       if (max_error < error)
 346         max_error = error;
 347       total_error += error;
 348     }
 349   }
 350
 351   EXPECT_GE(1, max_error)
 352       << "Error: 16x16 FHT/IHT has an individual round trip error > 1";
 353
 354   EXPECT_GE(count_test_block , total_error)
 355       << "Error: 16x16 FHT/IHT has average round trip error > 1 per block";
 356 }
 357
 358 TEST_P(FwdTrans16x16Test, CoeffSizeCheck) {
 359   ACMRandom rnd(ACMRandom::DeterministicSeed());
 360   const int count_test_block = 1000;
 361   for (int i = 0; i < count_test_block; ++i) {
 362     DECLARE_ALIGNED_ARRAY(16, int16_t, input_block, 256);
 363     DECLARE_ALIGNED_ARRAY(16, int16_t, input_extreme_block, 256);
 364     DECLARE_ALIGNED_ARRAY(16, int16_t, output_block, 256);
 365     DECLARE_ALIGNED_ARRAY(16, int16_t, output_extreme_block, 256);
 366     DECLARE_ALIGNED_ARRAY(16, uint8_t, dst, 256);
 367
 368     // Initialize a test block with input range [-255, 255].
 369     for (int j = 0; j < 256; ++j) {
 370       input_block[j] = rnd.Rand8() - rnd.Rand8();
 371       input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
 372     }
 373     if (i == 0)
 374       for (int j = 0; j < 256; ++j)
 375         input_extreme_block[j] = 255;
 376
 377     const int pitch = 32;
 378     RunFwdTxfm(input_block, output_block, dst, pitch, tx_type_);
 379     RunFwdTxfm(input_extreme_block, output_extreme_block, dst, pitch, tx_type_);
 380
 381     // The minimum quant value is 4.
 382     for (int j = 0; j < 256; ++j) {
 383       EXPECT_GE(4 * DCT_MAX_VALUE, abs(output_block[j]))
 384           << "Error: 16x16 FDCT has coefficient larger than 4*DCT_MAX_VALUE";
 385       EXPECT_GE(4 * DCT_MAX_VALUE, abs(output_extreme_block[j]))
 386           << "Error: 16x16 FDCT extreme has coefficient larger "
 387           << "than 4*DCT_MAX_VALUE";
 388     }
 389   }
 390 }
 391
 392 INSTANTIATE_TEST_CASE_P(VP9, FwdTrans16x16Test, ::testing::Range(0, 4));
 393
 394 TEST(VP9Idct16x16Test, AccuracyCheck) {
 395   ACMRandom rnd(ACMRandom::DeterministicSeed());
 396   const int count_test_block = 1000;
 397   for (int i = 0; i < count_test_block; ++i) {
 398     int16_t in[256], coeff[256];
 399     uint8_t dst[256], src[256];
 400     double out_r[256];
 401
 402     for (int j = 0; j < 256; ++j) {
 403       src[j] = rnd.Rand8();
 404       dst[j] = rnd.Rand8();
 405     }
 406     // Initialize a test block with input range [-255, 255].
 407     for (int j = 0; j < 256; ++j)
 408       in[j] = src[j] - dst[j];
 409
 410     reference_16x16_dct_2d(in, out_r);
 411     for (int j = 0; j < 256; j++)
 412       coeff[j] = round(out_r[j]);
 413     vp9_short_idct16x16_add_c(coeff, dst, 16);
 414     for (int j = 0; j < 256; ++j) {
 415       const int diff = dst[j] - src[j];
 416       const int error = diff * diff;
 417       EXPECT_GE(1, error)
 418           << "Error: 16x16 IDCT has error " << error
 419           << " at index " << j;
 420     }
 421   }
 422 }
 423
 424 }  // namespace