vpx/test/dct32x32_test.cc

186 lines
5.6 KiB
C++
Raw Normal View History

32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
/*
* Copyright (c) 2012 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include "third_party/googletest/src/include/gtest/gtest.h"
extern "C" {
#include "vp9/common/vp9_entropy.h"
#include "./vp9_rtcd.h"
void vp9_short_fdct32x32_c(int16_t *input, int16_t *out, int pitch);
void vp9_short_idct32x32_c(short *input, short *output, int pitch);
}
#include "test/acm_random.h"
#include "vpx/vpx_integer.h"
using libvpx_test::ACMRandom;
namespace {
#ifdef _MSC_VER
static int round(double x) {
if (x < 0)
return (int)ceil(x - 0.5);
else
return (int)floor(x + 0.5);
}
#endif
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
static const double kPi = 3.141592653589793238462643383279502884;
static void reference2_32x32_idct_2d(double *input, double *output) {
double x;
for (int l = 0; l < 32; ++l) {
for (int k = 0; k < 32; ++k) {
double s = 0;
for (int i = 0; i < 32; ++i) {
for (int j = 0; j < 32; ++j) {
x = cos(kPi * j * (l + 0.5) / 32.0) *
cos(kPi * i * (k + 0.5) / 32.0) * input[i * 32 + j] / 1024;
if (i != 0)
x *= sqrt(2.0);
if (j != 0)
x *= sqrt(2.0);
s += x;
}
}
output[k * 32 + l] = s / 4;
}
}
}
static void reference_32x32_dct_1d(double in[32], double out[32], int stride) {
const double kInvSqrt2 = 0.707106781186547524400844362104;
for (int k = 0; k < 32; k++) {
out[k] = 0.0;
for (int n = 0; n < 32; n++)
out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 64.0);
if (k == 0)
out[k] = out[k] * kInvSqrt2;
}
}
static void reference_32x32_dct_2d(int16_t input[32*32], double output[32*32]) {
// First transform columns
for (int i = 0; i < 32; ++i) {
double temp_in[32], temp_out[32];
for (int j = 0; j < 32; ++j)
temp_in[j] = input[j*32 + i];
reference_32x32_dct_1d(temp_in, temp_out, 1);
for (int j = 0; j < 32; ++j)
output[j * 32 + i] = temp_out[j];
}
// Then transform rows
for (int i = 0; i < 32; ++i) {
double temp_in[32], temp_out[32];
for (int j = 0; j < 32; ++j)
temp_in[j] = output[j + i*32];
reference_32x32_dct_1d(temp_in, temp_out, 1);
// Scale by some magic number
for (int j = 0; j < 32; ++j)
output[j + i * 32] = temp_out[j] / 4;
}
}
TEST(VP9Idct32x32Test, AccuracyCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
const int count_test_block = 1000;
for (int i = 0; i < count_test_block; ++i) {
int16_t in[1024], coeff[1024];
int16_t out_c[1024];
double out_r[1024];
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 1024; ++j)
in[j] = rnd.Rand8() - rnd.Rand8();
reference_32x32_dct_2d(in, out_r);
for (int j = 0; j < 1024; j++)
coeff[j] = round(out_r[j]);
vp9_short_idct32x32_c(coeff, out_c, 64);
for (int j = 0; j < 1024; ++j) {
const int diff = out_c[j] - in[j];
const int error = diff * diff;
EXPECT_GE(1, error)
<< "Error: 3x32 IDCT has error " << error
<< " at index " << j;
}
}
}
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
TEST(VP9Fdct32x32Test, AccuracyCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
unsigned int max_error = 0;
int64_t total_error = 0;
const int count_test_block = 1000;
for (int i = 0; i < count_test_block; ++i) {
int16_t test_input_block[1024];
int16_t test_temp_block[1024];
int16_t test_output_block[1024];
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 1024; ++j)
test_input_block[j] = rnd.Rand8() - rnd.Rand8();
const int pitch = 64;
vp9_short_fdct32x32_c(test_input_block, test_temp_block, pitch);
vp9_short_idct32x32_c(test_temp_block, test_output_block, pitch);
for (int j = 0; j < 1024; ++j) {
const unsigned diff = test_input_block[j] - test_output_block[j];
const unsigned error = diff * diff;
if (max_error < error)
max_error = error;
total_error += error;
}
}
EXPECT_GE(1u, max_error)
<< "Error: 32x32 FDCT/IDCT has an individual roundtrip error > 1";
EXPECT_GE(count_test_block, total_error)
<< "Error: 32x32 FDCT/IDCT has average roundtrip error > 1 per block";
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
}
TEST(VP9Fdct32x32Test, CoeffSizeCheck) {
ACMRandom rnd(ACMRandom::DeterministicSeed());
const int count_test_block = 1000;
for (int i = 0; i < count_test_block; ++i) {
int16_t input_block[1024], input_extreme_block[1024];
int16_t output_block[1024], output_extreme_block[1024];
// Initialize a test block with input range [-255, 255].
for (int j = 0; j < 1024; ++j) {
input_block[j] = rnd.Rand8() - rnd.Rand8();
input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
}
if (i == 0)
for (int j = 0; j < 1024; ++j)
input_extreme_block[j] = 255;
const int pitch = 64;
32x32 transform for superblocks. This adds Debargha's DCT/DWT hybrid and a regular 32x32 DCT, and adds code all over the place to wrap that in the bitstream/encoder/decoder/RD. Some implementation notes (these probably need careful review): - token range is extended by 1 bit, since the value range out of this transform is [-16384,16383]. - the coefficients coming out of the FDCT are manually scaled back by 1 bit, or else they won't fit in int16_t (they are 17 bits). Because of this, the RD error scoring does not right-shift the MSE score by two (unlike for 4x4/8x8/16x16). - to compensate for this loss in precision, the quantizer is halved also. This is currently a little hacky. - FDCT and IDCT is double-only right now. Needs a fixed-point impl. - There are no default probabilities for the 32x32 transform yet; I'm simply using the 16x16 luma ones. A future commit will add newly generated probabilities for all transforms. - No ADST version. I don't think we'll add one for this level; if an ADST is desired, transform-size selection can scale back to 16x16 or lower, and use an ADST at that level. Additional notes specific to Debargha's DWT/DCT hybrid: - coefficient scale is different for the top/left 16x16 (DCT-over-DWT) block than for the rest (DWT pixel differences) of the block. Therefore, RD error scoring isn't easily scalable between coefficient and pixel domain. Thus, unfortunately, we need to compute the RD distortion in the pixel domain until we figure out how to scale these appropriately. Change-Id: I00386f20f35d7fabb19aba94c8162f8aee64ef2b
2012-12-07 23:45:05 +01:00
vp9_short_fdct32x32_c(input_block, output_block, pitch);
vp9_short_fdct32x32_c(input_extreme_block, output_extreme_block, pitch);
// The minimum quant value is 4.
for (int j = 0; j < 1024; ++j) {
EXPECT_GE(4*DCT_MAX_VALUE, abs(output_block[j]))
<< "Error: 32x32 FDCT has coefficient larger than 4*DCT_MAX_VALUE";
EXPECT_GE(4*DCT_MAX_VALUE, abs(output_extreme_block[j]))
<< "Error: 32x32 FDCT extreme has coefficient larger than "
"4*DCT_MAX_VALUE";
}
}
}
} // namespace