| 1 |
#include "SPECK3D.h" |
| 2 |
#include "CDF97.h" |
| 3 |
|
| 4 |
#include <algorithm> |
| 5 |
#include <cstdlib> |
| 6 |
#include <cstring> |
| 7 |
#include <fstream> |
| 8 |
#include <iostream> |
| 9 |
#include <chrono> |
| 10 |
|
| 11 |
|
| 12 |
// #define FIRST_STEP |
| 13 |
|
| 14 |
|
| 15 |
extern "C" // C Function calls, and don't include the C header! |
| 16 |
{ |
| 17 |
int sam_read_n_bytes( const char*, size_t, void* ); |
| 18 |
int sam_write_n_doubles( const char*, size_t, const double* ); |
| 19 |
int sam_get_statsd( const double* arr1, const double* arr2, size_t len, |
| 20 |
double* rmse, double* lmax, double* psnr, |
| 21 |
double* arr1min, double* arr1max ); |
| 22 |
} |
| 23 |
|
| 24 |
|
| 25 |
int main( int argc, char* argv[] ) |
| 26 |
{ |
| 27 |
|
| 28 |
if( argc != 6 ) |
| 29 |
{ |
| 30 |
|
| 31 |
#ifdef QZ_TERM |
| 32 |
std::cerr << "Usage: ./a.out input_filename dim_x dim_y dim_z qz_levels" << std::endl; |
| 33 |
#else |
| 34 |
std::cerr << "Usage: ./a.out input_filename dim_x dim_y dim_z cratio" << std::endl; |
| 35 |
#endif |
| 36 |
|
| 37 |
return 1; |
| 38 |
} |
| 39 |
|
| 40 |
char output[256]; |
| 41 |
std::strcpy( output, argv[0] ); |
| 42 |
std::strcat( output, ".tmp" ); |
| 43 |
|
| 44 |
const char* input = argv[1]; |
| 45 |
const size_t dim_x = std::atol( argv[2] ); |
| 46 |
const size_t dim_y = std::atol( argv[3] ); |
| 47 |
const size_t dim_z = std::atol( argv[4] ); |
| 48 |
const size_t total_vals = dim_x * dim_y * dim_z; |
| 49 |
|
| 50 |
#ifdef QZ_TERM |
| 51 |
const int qz_levels = std::atoi( argv[5] ); |
| 52 |
#else |
| 53 |
const float cratio = std::atof( argv[5] ); |
| 54 |
#endif |
| 55 |
|
| 56 |
#ifdef NO_CPP14 |
| 57 |
speck::buffer_type_f in_buf( new float[total_vals] ); |
| 58 |
#else |
| 59 |
speck::buffer_type_f in_buf = std::make_unique<float[]>( total_vals ); |
| 60 |
#endif |
| 61 |
|
| 62 |
// Let's read in binaries as 4-byte floats |
| 63 |
if( sam_read_n_bytes( input, sizeof(float) * total_vals, in_buf.get() ) ) |
| 64 |
{ |
| 65 |
std::cerr << "Input read error!" << std::endl; |
| 66 |
return 1; |
| 67 |
} |
| 68 |
|
| 69 |
// Take input to go through DWT. |
| 70 |
speck::CDF97 cdf; |
| 71 |
cdf.set_dims( dim_x, dim_y, dim_z ); |
| 72 |
cdf.copy_data( in_buf, total_vals ); |
| 73 |
const auto startT = std::chrono::high_resolution_clock::now(); |
| 74 |
cdf.dwt3d(); |
| 75 |
|
| 76 |
// Do a speck encoding |
| 77 |
speck::SPECK3D encoder; |
| 78 |
encoder.set_dims( dim_x, dim_y, dim_z ); |
| 79 |
encoder.set_image_mean( cdf.get_mean() ); |
| 80 |
encoder.take_coeffs( cdf.release_data(), total_vals ); |
| 81 |
|
| 82 |
#ifdef QZ_TERM |
| 83 |
#ifdef FIRST_STEP |
| 84 |
encoder.set_quantization_iterations( qz_levels ); |
| 85 |
#else |
| 86 |
encoder.set_quantization_term_level( qz_levels ); |
| 87 |
#endif |
| 88 |
#else |
| 89 |
const size_t total_bits = size_t(32.0f * total_vals / cratio); |
| 90 |
encoder.set_bit_budget( total_bits ); |
| 91 |
#endif |
| 92 |
|
| 93 |
encoder.encode(); |
| 94 |
encoder.write_to_disk( output ); |
| 95 |
|
| 96 |
// Do a speck decoding |
| 97 |
speck::SPECK3D decoder; |
| 98 |
decoder.read_from_disk( output ); |
| 99 |
|
| 100 |
#ifdef QZ_TERM |
| 101 |
decoder.set_bit_budget( 0 ); |
| 102 |
#else |
| 103 |
decoder.set_bit_budget( total_bits ); |
| 104 |
#endif |
| 105 |
|
| 106 |
decoder.decode(); |
| 107 |
|
| 108 |
// Do an inverse wavelet transform |
| 109 |
speck::CDF97 idwt; |
| 110 |
size_t dim_x_r, dim_y_r, dim_z_r; |
| 111 |
decoder.get_dims( dim_x_r, dim_y_r, dim_z_r ); |
| 112 |
idwt.set_dims( dim_x_r, dim_y_r, dim_z_r ); |
| 113 |
idwt.set_mean( decoder.get_image_mean() ); |
| 114 |
idwt.take_data( decoder.release_coeffs_double(), dim_x_r * dim_y_r * dim_z_r ); |
| 115 |
idwt.idwt3d(); |
| 116 |
|
| 117 |
// Finish timer and print timing |
| 118 |
const auto endT = std::chrono::high_resolution_clock::now(); |
| 119 |
const std::chrono::duration<double> diffT = endT - startT; |
| 120 |
//std::cout << "Time for SPECK in milliseconds: " << diffT.count() * 1000.0f << std::endl; |
| 121 |
|
| 122 |
// Compare the result with the original input in double precision |
| 123 |
|
| 124 |
#ifdef QZ_TERM |
| 125 |
float bpp = float(encoder.get_num_of_bits()) / float(total_vals); |
| 126 |
#endif |
| 127 |
|
| 128 |
|
| 129 |
#ifdef FIRST_STEP |
| 130 |
speck::buffer_type_d in_bufd = std::make_unique<double[]>( total_vals ); |
| 131 |
for( size_t i = 0; i < total_vals; i++ ) |
| 132 |
in_bufd[i] = in_buf[i]; |
| 133 |
|
| 134 |
double rmse, lmax, psnr, arr1min, arr1max; |
| 135 |
sam_get_statsd( in_bufd.get(), idwt.get_read_only_data().get(), |
| 136 |
total_vals, &rmse, &lmax, &psnr, &arr1min, &arr1max ); |
| 137 |
printf("Sam stats: rmse = %f, lmax = %f, psnr = %fdB, orig_min = %f, orig_max = %f\n", |
| 138 |
rmse, lmax, psnr, arr1min, arr1max ); |
| 139 |
in_bufd.reset( nullptr ); |
| 140 |
|
| 141 |
printf("With %d levels of quantization, average BPP = %f, and qz terminates at level %d\n", |
| 142 |
qz_levels, bpp, encoder.get_quantization_term_level() ); |
| 143 |
|
| 144 |
// For the first step, print out the error values at the following positions: |
| 145 |
// 0, 50, 100, 200, 300, 400, 500, 1000 |
| 146 |
|
| 147 |
std::vector<float> diff; |
| 148 |
diff.reserve( total_vals ); |
| 149 |
for( size_t i = 0; i < total_vals; i++ ) |
| 150 |
diff.push_back( std::abs(in_buf[i] - float(idwt.get_read_only_data()[i]) )); |
| 151 |
std::partial_sort( diff.begin(), diff.begin() + 1001, diff.end(), std::greater<float>() ); |
| 152 |
std::array<size_t, 8> idx { 0, 50, 100, 200, 300, 400, 500, 1000 }; |
| 153 |
for( auto i : idx ) |
| 154 |
std::cout << diff[i] << ", "; |
| 155 |
std::cout << std::endl; |
| 156 |
|
| 157 |
#else |
| 158 |
|
| 159 |
// Count how many data points end up having errors greater than the thresholds. |
| 160 |
// |
| 161 |
const size_t num_of_th = 8; |
| 162 |
// Variable dbz, qz_levels = 17 |
| 163 |
//std::array<float, num_of_th> threshold{0.734512, 0.596821, 0.56641, 0.535465, |
| 164 |
// 0.519047, 0.504593, 0.491455, 0.459518 }; |
| 165 |
// Variable hvort, qz_levels = 14 -9 |
| 166 |
//std::array<float, num_of_th> threshold{0.00615794, 0.00417704, 0.00400225, 0.00382341, |
| 167 |
// 0.00369853, 0.00361842, 0.00356053, 0.00338412 }; |
| 168 |
// Variable prespert, qz_levels = 20 -7 |
| 169 |
//std::array<float, num_of_th> threshold{0.0246198, 0.0170196, 0.01632, 0.0154797, |
| 170 |
// 0.0149632, 0.0146113, 0.0143417, 0.0134916}; |
| 171 |
// Variable qc, qz_levels = 20 |
| 172 |
//std::array<float, num_of_th> threshold{0.010118, 0.00804029, 0.00763616, 0.00728243, |
| 173 |
// 0.00708556, 0.00691819, 0.00680445, 0.00638294}; |
| 174 |
// Variable qg, qz_levels = 23 -11 |
| 175 |
//std::array<float, num_of_th> threshold{0.00146534, 0.0010819, 0.00103881, 0.000984068, |
| 176 |
// 0.000951942, 0.000932336, 0.000913382, 0.000861526}; |
| 177 |
// Variable qr, qz_levels = 20 -9 |
| 178 |
//std::array<float, num_of_th> threshold{0.00518513, 0.00401469, 0.00384271, 0.00364681, |
| 179 |
// 0.00351448, 0.00343102, 0.00336915, 0.0031701}; |
| 180 |
// Variable streamvort, qz_levels = 14 -9 |
| 181 |
//std::array<float, num_of_th> threshold{0.00578914, 0.00446857, 0.00430652, 0.00411505, |
| 182 |
// 0.00397266, 0.00388597, 0.00380701, 0.00359325}; |
| 183 |
// Variable thrhopert, qz_levels = 18 -6 |
| 184 |
std::array<float, num_of_th> threshold{0.0559266, 0.0360634, 0.0344715, 0.03269, |
| 185 |
0.0315752, 0.0308189, 0.0302553, 0.0283475 }; |
| 186 |
// Variable uinterp, qz_levels = 19 -5 |
| 187 |
//std::array<float, num_of_th> threshold{0.0894642, 0.0709276, 0.0674381, 0.0647449, |
| 188 |
// 0.0627775, 0.0612926, 0.0602803, 0.0566506}; |
| 189 |
// Variable vortmag, qz_levels = 15 -9 |
| 190 |
//std::array<float, num_of_th> threshold{0.00588624, 0.00424516, 0.00401397, 0.00382658, |
| 191 |
// 0.0037123, 0.00362666, 0.00357628, 0.00338255}; |
| 192 |
|
| 193 |
std::vector<float> diff; |
| 194 |
diff.reserve( total_vals ); |
| 195 |
for( size_t i = 0; i < total_vals; i++ ) |
| 196 |
diff.push_back( std::abs(in_buf[i] - float(idwt.get_read_only_data()[i]) )); |
| 197 |
|
| 198 |
std::array<size_t, num_of_th> count; |
| 199 |
count.fill( 0 ); |
| 200 |
for( auto& e : diff ) { |
| 201 |
for( size_t i = 0; i < num_of_th; i++ ) { |
| 202 |
if( e > threshold[i] ) { |
| 203 |
while( i < num_of_th ) { |
| 204 |
count[i]++; |
| 205 |
i++; |
| 206 |
} |
| 207 |
} |
| 208 |
} |
| 209 |
} |
| 210 |
|
| 211 |
// First print out the effective bpp, then print out the number of outliers |
| 212 |
std::cout << bpp << ", "; |
| 213 |
for( auto& e : count ) |
| 214 |
std::cout << e << ", "; |
| 215 |
std::cout << std::endl; |
| 216 |
#endif |
| 217 |
|
| 218 |
|
| 219 |
|
| 220 |
#ifdef EXPERIMENT |
| 221 |
// Experiment 1: |
| 222 |
// Sort the differences and then write a tenth of it to disk. |
| 223 |
std::vector<speck::Outlier> LOS( total_vals, speck::Outlier{} ); |
| 224 |
for( size_t i = 0; i < total_vals; i++ ) { |
| 225 |
LOS[i].location = i; |
| 226 |
LOS[i].error = in_buf[i] - float(idwt.get_read_only_data()[i]); |
| 227 |
} |
| 228 |
|
| 229 |
const size_t num_of_outliers = total_vals / 10; |
| 230 |
std::partial_sort( LOS.begin(), LOS.begin() + num_of_outliers, LOS.end(), |
| 231 |
[](auto& a, auto& b) { return (std::abs(a.error) > std::abs(b.error)); } ); |
| 232 |
|
| 233 |
for( size_t i = 0; i < 10; i++ ) |
| 234 |
printf("outliers: (%ld, %f)\n", LOS[i].location, LOS[i].error ); |
| 235 |
|
| 236 |
std::ofstream file( "top_outliers", std::ios::binary ); |
| 237 |
if( file.is_open() ) { |
| 238 |
file.write( reinterpret_cast<char*>(LOS.data()), sizeof(speck::Outlier) * num_of_outliers ); |
| 239 |
file.close(); |
| 240 |
} |
| 241 |
#endif |
| 242 |
|
| 243 |
} |