| Line |
Branch |
Exec |
Source |
| 1 |
|
|
/* |
| 2 |
|
|
|
| 3 |
|
|
Copyright (c) 2023, Dominic Szablewski - https://phoboslab.org |
| 4 |
|
|
SPDX-License-Identifier: MIT |
| 5 |
|
|
|
| 6 |
|
|
QOA - The "Quite OK Audio" format for fast, lossy audio compression |
| 7 |
|
|
|
| 8 |
|
|
|
| 9 |
|
|
-- Data Format |
| 10 |
|
|
|
| 11 |
|
|
A QOA file has an 8 byte file header, followed by a number of frames. Each frame |
| 12 |
|
|
consists of an 8 byte frame header, the current 8 byte en-/decoder state per |
| 13 |
|
|
channel and 256 slices per channel. Each slice is 8 bytes wide and encodes 20 |
| 14 |
|
|
samples of audio data. |
| 15 |
|
|
|
| 16 |
|
|
Note that the last frame of a file may contain less than 256 slices per channel. |
| 17 |
|
|
The last slice (per channel) in the last frame may contain less 20 samples, but |
| 18 |
|
|
the slice will still be 8 bytes wide, with the unused samples zeroed out. |
| 19 |
|
|
|
| 20 |
|
|
The samplerate and number of channels is only stated in the frame headers, but |
| 21 |
|
|
not in the file header. A decoder may peek into the first frame of the file to |
| 22 |
|
|
find these values. |
| 23 |
|
|
|
| 24 |
|
|
In a valid QOA file all frames have the same number of channels and the same |
| 25 |
|
|
samplerate. These restrictions may be relaxed for streaming. This remains to |
| 26 |
|
|
be decided. |
| 27 |
|
|
|
| 28 |
|
|
All values in a QOA file are BIG ENDIAN. Luckily, EVERYTHING in a QOA file, |
| 29 |
|
|
including the headers, is 64 bit aligned, so it's possible to read files with |
| 30 |
|
|
just a read_u64() that does the byte swapping if necessary. |
| 31 |
|
|
|
| 32 |
|
|
In pseudocode, the file layout is as follows: |
| 33 |
|
|
|
| 34 |
|
|
struct { |
| 35 |
|
|
struct { |
| 36 |
|
|
char magic[4]; // magic bytes 'qoaf' |
| 37 |
|
|
uint32_t samples; // number of samples per channel in this file |
| 38 |
|
|
} file_header; // = 64 bits |
| 39 |
|
|
|
| 40 |
|
|
struct { |
| 41 |
|
|
struct { |
| 42 |
|
|
uint8_t num_channels; // number of channels |
| 43 |
|
|
uint24_t samplerate; // samplerate in hz |
| 44 |
|
|
uint16_t fsamples; // sample count per channel in this frame |
| 45 |
|
|
uint16_t fsize; // frame size (including the frame header) |
| 46 |
|
|
} frame_header; // = 64 bits |
| 47 |
|
|
|
| 48 |
|
|
struct { |
| 49 |
|
|
int16_t history[4]; // = 64 bits |
| 50 |
|
|
int16_t weights[4]; // = 64 bits |
| 51 |
|
|
} lms_state[num_channels]; |
| 52 |
|
|
|
| 53 |
|
|
qoa_slice_t slices[256][num_channels]; // = 64 bits each |
| 54 |
|
|
} frames[samples * channels / qoa_max_framesize()]; |
| 55 |
|
|
} qoa_file; |
| 56 |
|
|
|
| 57 |
|
|
Wheras the 64bit qoa_slice_t is defined as follows: |
| 58 |
|
|
|
| 59 |
|
|
.- QOA_SLICE -- 64 bits, 20 samples --------------------------/ /------------. |
| 60 |
|
|
| Byte[0] | Byte[1] | Byte[2] \ \ Byte[7] | |
| 61 |
|
|
| 7 6 5 4 3 2 1 0 | 7 6 5 4 3 2 1 0 | 7 6 5 / / 2 1 0 | |
| 62 |
|
|
|------------+--------+--------+--------+---------+---------+-\ \--+---------| |
| 63 |
|
|
| sf_index | r00 | r01 | r02 | r03 | r04 | / / | r19 | |
| 64 |
|
|
`-------------------------------------------------------------\ \------------` |
| 65 |
|
|
|
| 66 |
|
|
`sf_index` defines the scalefactor to use for this slice as an index into the |
| 67 |
|
|
qoa_scalefactor_tab[16] |
| 68 |
|
|
|
| 69 |
|
|
`r00`--`r19` are the residuals for the individual samples, divided by the |
| 70 |
|
|
scalefactor and quantized by the qoa_quant_tab[]. |
| 71 |
|
|
|
| 72 |
|
|
In the decoder, a prediction of the next sample is computed by multiplying the |
| 73 |
|
|
state (the last four output samples) with the predictor. The residual from the |
| 74 |
|
|
slice is then dequantized using the qoa_dequant_tab[] and added to the |
| 75 |
|
|
prediction. The result is clamped to int16 to form the final output sample. |
| 76 |
|
|
|
| 77 |
|
|
*/ |
| 78 |
|
|
|
| 79 |
|
|
|
| 80 |
|
|
|
| 81 |
|
|
/* ----------------------------------------------------------------------------- |
| 82 |
|
|
Header - Public functions */ |
| 83 |
|
|
|
| 84 |
|
|
#ifndef QOA_H |
| 85 |
|
|
#define QOA_H |
| 86 |
|
|
|
| 87 |
|
|
#ifdef __cplusplus |
| 88 |
|
|
extern "C" { |
| 89 |
|
|
#endif |
| 90 |
|
|
|
| 91 |
|
|
#define QOA_MIN_FILESIZE 16 |
| 92 |
|
|
#define QOA_MAX_CHANNELS 8 |
| 93 |
|
|
|
| 94 |
|
|
#define QOA_SLICE_LEN 20 |
| 95 |
|
|
#define QOA_SLICES_PER_FRAME 256 |
| 96 |
|
|
#define QOA_FRAME_LEN (QOA_SLICES_PER_FRAME * QOA_SLICE_LEN) |
| 97 |
|
|
#define QOA_LMS_LEN 4 |
| 98 |
|
|
#define QOA_MAGIC 0x716f6166 /* 'qoaf' */ |
| 99 |
|
|
|
| 100 |
|
|
#define QOA_FRAME_SIZE(channels, slices) \ |
| 101 |
|
|
(8 + QOA_LMS_LEN * 4 * channels + 8 * slices * channels) |
| 102 |
|
|
|
| 103 |
|
|
typedef struct { |
| 104 |
|
|
int history[QOA_LMS_LEN]; |
| 105 |
|
|
int weights[QOA_LMS_LEN]; |
| 106 |
|
|
} qoa_lms_t; |
| 107 |
|
|
|
| 108 |
|
|
typedef struct { |
| 109 |
|
|
unsigned int channels; |
| 110 |
|
|
unsigned int samplerate; |
| 111 |
|
|
unsigned int samples; |
| 112 |
|
|
qoa_lms_t lms[QOA_MAX_CHANNELS]; |
| 113 |
|
|
#ifdef QOA_RECORD_TOTAL_ERROR |
| 114 |
|
|
double error; |
| 115 |
|
|
#endif |
| 116 |
|
|
} qoa_desc; |
| 117 |
|
|
|
| 118 |
|
|
unsigned int qoa_encode_header(qoa_desc *qoa, unsigned char *bytes); |
| 119 |
|
|
unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned int frame_len, unsigned char *bytes); |
| 120 |
|
|
void *qoa_encode(const short *sample_data, qoa_desc *qoa, unsigned int *out_len); |
| 121 |
|
|
|
| 122 |
|
|
unsigned int qoa_max_frame_size(qoa_desc *qoa); |
| 123 |
|
|
unsigned int qoa_decode_header(const unsigned char *bytes, int size, qoa_desc *qoa); |
| 124 |
|
|
unsigned int qoa_decode_frame(const unsigned char *bytes, unsigned int size, qoa_desc *qoa, short *sample_data, unsigned int *frame_len); |
| 125 |
|
|
short *qoa_decode(const unsigned char *bytes, int size, qoa_desc *file); |
| 126 |
|
|
|
| 127 |
|
|
#ifndef QOA_NO_STDIO |
| 128 |
|
|
|
| 129 |
|
|
int qoa_write(const char *filename, const short *sample_data, qoa_desc *qoa); |
| 130 |
|
|
void *qoa_read(const char *filename, qoa_desc *qoa); |
| 131 |
|
|
|
| 132 |
|
|
#endif /* QOA_NO_STDIO */ |
| 133 |
|
|
|
| 134 |
|
|
|
| 135 |
|
|
#ifdef __cplusplus |
| 136 |
|
|
} |
| 137 |
|
|
#endif |
| 138 |
|
|
#endif /* QOA_H */ |
| 139 |
|
|
|
| 140 |
|
|
|
| 141 |
|
|
/* ----------------------------------------------------------------------------- |
| 142 |
|
|
Implementation */ |
| 143 |
|
|
|
| 144 |
|
|
#ifdef QOA_IMPLEMENTATION |
| 145 |
|
|
#include <stdlib.h> |
| 146 |
|
|
|
| 147 |
|
|
#ifndef QOA_MALLOC |
| 148 |
|
|
#define QOA_MALLOC(sz) malloc(sz) |
| 149 |
|
|
#define QOA_FREE(p) free(p) |
| 150 |
|
|
#endif |
| 151 |
|
|
|
| 152 |
|
|
typedef unsigned long long qoa_uint64_t; |
| 153 |
|
|
|
| 154 |
|
|
|
| 155 |
|
|
/* The quant_tab provides an index into the dequant_tab for residuals in the |
| 156 |
|
|
range of -8 .. 8. It maps this range to just 3bits and becomes less accurate at |
| 157 |
|
|
the higher end. Note that the residual zero is identical to the lowest positive |
| 158 |
|
|
value. This is mostly fine, since the qoa_div() function always rounds away |
| 159 |
|
|
from zero. */ |
| 160 |
|
|
|
| 161 |
|
|
static int qoa_quant_tab[17] = { |
| 162 |
|
|
7, 7, 7, 5, 5, 3, 3, 1, /* -8..-1 */ |
| 163 |
|
|
0, /* 0 */ |
| 164 |
|
|
0, 2, 2, 4, 4, 6, 6, 6 /* 1.. 8 */ |
| 165 |
|
|
}; |
| 166 |
|
|
|
| 167 |
|
|
|
| 168 |
|
|
/* We have 16 different scalefactors. Like the quantized residuals these become |
| 169 |
|
|
less accurate at the higher end. In theory, the highest scalefactor that we |
| 170 |
|
|
would need to encode the highest 16bit residual is (2**16)/8 = 8192. However we |
| 171 |
|
|
rely on the LMS filter to predict samples accurately enough that a maximum |
| 172 |
|
|
residual of one quarter of the 16 bit range is high sufficient. I.e. with the |
| 173 |
|
|
scalefactor 2048 times the quant range of 8 we can encode residuals up to 2**14. |
| 174 |
|
|
|
| 175 |
|
|
The scalefactor values are computed as: |
| 176 |
|
|
scalefactor_tab[s] <- round(pow(s + 1, 2.75)) */ |
| 177 |
|
|
|
| 178 |
|
|
static int qoa_scalefactor_tab[16] = { |
| 179 |
|
|
1, 7, 21, 45, 84, 138, 211, 304, 421, 562, 731, 928, 1157, 1419, 1715, 2048 |
| 180 |
|
|
}; |
| 181 |
|
|
|
| 182 |
|
|
|
| 183 |
|
|
/* The reciprocal_tab maps each of the 16 scalefactors to their rounded |
| 184 |
|
|
reciprocals 1/scalefactor. This allows us to calculate the scaled residuals in |
| 185 |
|
|
the encoder with just one multiplication instead of an expensive division. We |
| 186 |
|
|
do this in .16 fixed point with integers, instead of floats. |
| 187 |
|
|
|
| 188 |
|
|
The reciprocal_tab is computed as: |
| 189 |
|
|
reciprocal_tab[s] <- ((1<<16) + scalefactor_tab[s] - 1) / scalefactor_tab[s] */ |
| 190 |
|
|
|
| 191 |
|
|
static int qoa_reciprocal_tab[16] = { |
| 192 |
|
|
65536, 9363, 3121, 1457, 781, 475, 311, 216, 156, 117, 90, 71, 57, 47, 39, 32 |
| 193 |
|
|
}; |
| 194 |
|
|
|
| 195 |
|
|
|
| 196 |
|
|
/* The dequant_tab maps each of the scalefactors and quantized residuals to |
| 197 |
|
|
their unscaled & dequantized version. |
| 198 |
|
|
|
| 199 |
|
|
Since qoa_div rounds away from the zero, the smallest entries are mapped to 3/4 |
| 200 |
|
|
instead of 1. The dequant_tab assumes the following dequantized values for each |
| 201 |
|
|
of the quant_tab indices and is computed as: |
| 202 |
|
|
float dqt[8] = {0.75, -0.75, 2.5, -2.5, 4.5, -4.5, 7, -7}; |
| 203 |
|
|
dequant_tab[s][q] <- round(scalefactor_tab[s] * dqt[q]) */ |
| 204 |
|
|
|
| 205 |
|
|
static int qoa_dequant_tab[16][8] = { |
| 206 |
|
|
{ 1, -1, 3, -3, 5, -5, 7, -7}, |
| 207 |
|
|
{ 5, -5, 18, -18, 32, -32, 49, -49}, |
| 208 |
|
|
{ 16, -16, 53, -53, 95, -95, 147, -147}, |
| 209 |
|
|
{ 34, -34, 113, -113, 203, -203, 315, -315}, |
| 210 |
|
|
{ 63, -63, 210, -210, 378, -378, 588, -588}, |
| 211 |
|
|
{ 104, -104, 345, -345, 621, -621, 966, -966}, |
| 212 |
|
|
{ 158, -158, 528, -528, 950, -950, 1477, -1477}, |
| 213 |
|
|
{ 228, -228, 760, -760, 1368, -1368, 2128, -2128}, |
| 214 |
|
|
{ 316, -316, 1053, -1053, 1895, -1895, 2947, -2947}, |
| 215 |
|
|
{ 422, -422, 1405, -1405, 2529, -2529, 3934, -3934}, |
| 216 |
|
|
{ 548, -548, 1828, -1828, 3290, -3290, 5117, -5117}, |
| 217 |
|
|
{ 696, -696, 2320, -2320, 4176, -4176, 6496, -6496}, |
| 218 |
|
|
{ 868, -868, 2893, -2893, 5207, -5207, 8099, -8099}, |
| 219 |
|
|
{1064, -1064, 3548, -3548, 6386, -6386, 9933, -9933}, |
| 220 |
|
|
{1286, -1286, 4288, -4288, 7718, -7718, 12005, -12005}, |
| 221 |
|
|
{1536, -1536, 5120, -5120, 9216, -9216, 14336, -14336}, |
| 222 |
|
|
}; |
| 223 |
|
|
|
| 224 |
|
|
|
| 225 |
|
|
/* The Least Mean Squares Filter is the heart of QOA. It predicts the next |
| 226 |
|
|
sample based on the previous 4 reconstructed samples. It does so by continuously |
| 227 |
|
|
adjusting 4 weights based on the residual of the previous prediction. |
| 228 |
|
|
|
| 229 |
|
|
The next sample is predicted as the sum of (weight[i] * history[i]). |
| 230 |
|
|
|
| 231 |
|
|
The adjustment of the weights is done with a "Sign-Sign-LMS" that adds or |
| 232 |
|
|
subtracts the residual to each weight, based on the corresponding sample from |
| 233 |
|
|
the history. This, surprisingly, is sufficient to get worthwhile predictions. |
| 234 |
|
|
|
| 235 |
|
|
This is all done with fixed point integers. Hence the right-shifts when updating |
| 236 |
|
|
the weights and calculating the prediction. */ |
| 237 |
|
|
|
| 238 |
|
|
static int qoa_lms_predict(qoa_lms_t *lms) { |
| 239 |
|
|
int prediction = 0; |
| 240 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN; i++) { |
| 241 |
|
✗ |
prediction += lms->weights[i] * lms->history[i]; |
| 242 |
|
|
} |
| 243 |
|
✗ |
return prediction >> 13; |
| 244 |
|
|
} |
| 245 |
|
|
|
| 246 |
|
✗ |
static void qoa_lms_update(qoa_lms_t *lms, int sample, int residual) { |
| 247 |
|
✗ |
int delta = residual >> 4; |
| 248 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN; i++) { |
| 249 |
|
✗ |
lms->weights[i] += lms->history[i] < 0 ? -delta : delta; |
| 250 |
|
|
} |
| 251 |
|
|
|
| 252 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN-1; i++) { |
| 253 |
|
✗ |
lms->history[i] = lms->history[i+1]; |
| 254 |
|
|
} |
| 255 |
|
✗ |
lms->history[QOA_LMS_LEN-1] = sample; |
| 256 |
|
|
} |
| 257 |
|
|
|
| 258 |
|
|
|
| 259 |
|
|
/* qoa_div() implements a rounding division, but avoids rounding to zero for |
| 260 |
|
|
small numbers. E.g. 0.1 will be rounded to 1. Note that 0 itself still |
| 261 |
|
|
returns as 0, which is handled in the qoa_quant_tab[]. |
| 262 |
|
|
qoa_div() takes an index into the .16 fixed point qoa_reciprocal_tab as an |
| 263 |
|
|
argument, so it can do the division with a cheaper integer multiplication. */ |
| 264 |
|
|
|
| 265 |
|
|
static inline int qoa_div(int v, int scalefactor) { |
| 266 |
|
✗ |
int reciprocal = qoa_reciprocal_tab[scalefactor]; |
| 267 |
|
✗ |
int n = (v * reciprocal + (1 << 15)) >> 16; |
| 268 |
|
✗ |
n = n + ((v > 0) - (v < 0)) - ((n > 0) - (n < 0)); /* round away from 0 */ |
| 269 |
|
|
return n; |
| 270 |
|
|
} |
| 271 |
|
|
|
| 272 |
|
|
static inline int qoa_clamp(int v, int min, int max) { |
| 273 |
|
✗ |
return (v < min) ? min : (v > max) ? max : v; |
| 274 |
|
|
} |
| 275 |
|
|
|
| 276 |
|
✗ |
static inline qoa_uint64_t qoa_read_u64(const unsigned char *bytes, unsigned int *p) { |
| 277 |
|
✗ |
bytes += *p; |
| 278 |
|
✗ |
*p += 8; |
| 279 |
|
|
return |
| 280 |
|
✗ |
((qoa_uint64_t)(bytes[0]) << 56) | ((qoa_uint64_t)(bytes[1]) << 48) | |
| 281 |
|
✗ |
((qoa_uint64_t)(bytes[2]) << 40) | ((qoa_uint64_t)(bytes[3]) << 32) | |
| 282 |
|
✗ |
((qoa_uint64_t)(bytes[4]) << 24) | ((qoa_uint64_t)(bytes[5]) << 16) | |
| 283 |
|
✗ |
((qoa_uint64_t)(bytes[6]) << 8) | ((qoa_uint64_t)(bytes[7]) << 0); |
| 284 |
|
|
} |
| 285 |
|
|
|
| 286 |
|
|
static inline void qoa_write_u64(qoa_uint64_t v, unsigned char *bytes, unsigned int *p) { |
| 287 |
|
✗ |
bytes += *p; |
| 288 |
|
✗ |
*p += 8; |
| 289 |
|
✗ |
bytes[0] = (v >> 56) & 0xff; |
| 290 |
|
✗ |
bytes[1] = (v >> 48) & 0xff; |
| 291 |
|
✗ |
bytes[2] = (v >> 40) & 0xff; |
| 292 |
|
✗ |
bytes[3] = (v >> 32) & 0xff; |
| 293 |
|
✗ |
bytes[4] = (v >> 24) & 0xff; |
| 294 |
|
✗ |
bytes[5] = (v >> 16) & 0xff; |
| 295 |
|
✗ |
bytes[6] = (v >> 8) & 0xff; |
| 296 |
|
✗ |
bytes[7] = (v >> 0) & 0xff; |
| 297 |
|
|
} |
| 298 |
|
|
|
| 299 |
|
|
|
| 300 |
|
|
/* ----------------------------------------------------------------------------- |
| 301 |
|
|
Encoder */ |
| 302 |
|
|
|
| 303 |
|
✗ |
unsigned int qoa_encode_header(qoa_desc *qoa, unsigned char *bytes) { |
| 304 |
|
|
unsigned int p = 0; |
| 305 |
|
✗ |
qoa_write_u64(((qoa_uint64_t)QOA_MAGIC << 32) | qoa->samples, bytes, &p); |
| 306 |
|
✗ |
return p; |
| 307 |
|
|
} |
| 308 |
|
|
|
| 309 |
|
✗ |
unsigned int qoa_encode_frame(const short *sample_data, qoa_desc *qoa, unsigned int frame_len, unsigned char *bytes) { |
| 310 |
|
✗ |
unsigned int channels = qoa->channels; |
| 311 |
|
|
|
| 312 |
|
|
unsigned int p = 0; |
| 313 |
|
✗ |
unsigned int slices = (frame_len + QOA_SLICE_LEN - 1) / QOA_SLICE_LEN; |
| 314 |
|
✗ |
unsigned int frame_size = QOA_FRAME_SIZE(channels, slices); |
| 315 |
|
|
|
| 316 |
|
|
/* Write the frame header */ |
| 317 |
|
✗ |
qoa_write_u64(( |
| 318 |
|
✗ |
(qoa_uint64_t)qoa->channels << 56 | |
| 319 |
|
✗ |
(qoa_uint64_t)qoa->samplerate << 32 | |
| 320 |
|
✗ |
(qoa_uint64_t)frame_len << 16 | |
| 321 |
|
✗ |
(qoa_uint64_t)frame_size |
| 322 |
|
|
), bytes, &p); |
| 323 |
|
|
|
| 324 |
|
|
/* Write the current LMS state */ |
| 325 |
|
✗ |
for (int c = 0; c < channels; c++) { |
| 326 |
|
|
qoa_uint64_t weights = 0; |
| 327 |
|
|
qoa_uint64_t history = 0; |
| 328 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN; i++) { |
| 329 |
|
✗ |
history = (history << 16) | (qoa->lms[c].history[i] & 0xffff); |
| 330 |
|
✗ |
weights = (weights << 16) | (qoa->lms[c].weights[i] & 0xffff); |
| 331 |
|
|
} |
| 332 |
|
|
qoa_write_u64(history, bytes, &p); |
| 333 |
|
|
qoa_write_u64(weights, bytes, &p); |
| 334 |
|
|
} |
| 335 |
|
|
|
| 336 |
|
|
/* We encode all samples with the channels interleaved on a slice level. |
| 337 |
|
|
E.g. for stereo: (ch-0, slice 0), (ch 1, slice 0), (ch 0, slice 1), ...*/ |
| 338 |
|
✗ |
for (int sample_index = 0; sample_index < frame_len; sample_index += QOA_SLICE_LEN) { |
| 339 |
|
|
|
| 340 |
|
✗ |
for (int c = 0; c < channels; c++) { |
| 341 |
|
✗ |
int slice_len = qoa_clamp(QOA_SLICE_LEN, 0, frame_len - sample_index); |
| 342 |
|
✗ |
int slice_start = sample_index * channels + c; |
| 343 |
|
✗ |
int slice_end = (sample_index + slice_len) * channels + c; |
| 344 |
|
|
|
| 345 |
|
|
/* Brute for search for the best scalefactor. Just go through all |
| 346 |
|
|
16 scalefactors, encode all samples for the current slice and |
| 347 |
|
|
meassure the total squared error. */ |
| 348 |
|
|
qoa_uint64_t best_error = -1; |
| 349 |
|
|
qoa_uint64_t best_slice; |
| 350 |
|
|
qoa_lms_t best_lms; |
| 351 |
|
|
|
| 352 |
|
✗ |
for (int scalefactor = 0; scalefactor < 16; scalefactor++) { |
| 353 |
|
|
|
| 354 |
|
|
/* We have to reset the LMS state to the last known good one |
| 355 |
|
|
before trying each scalefactor, as each pass updates the LMS |
| 356 |
|
|
state when encoding. */ |
| 357 |
|
✗ |
qoa_lms_t lms = qoa->lms[c]; |
| 358 |
|
✗ |
qoa_uint64_t slice = scalefactor; |
| 359 |
|
|
qoa_uint64_t current_error = 0; |
| 360 |
|
|
|
| 361 |
|
✗ |
for (int si = slice_start; si < slice_end; si += channels) { |
| 362 |
|
✗ |
int sample = sample_data[si]; |
| 363 |
|
|
int predicted = qoa_lms_predict(&lms); |
| 364 |
|
|
|
| 365 |
|
✗ |
int residual = sample - predicted; |
| 366 |
|
|
int scaled = qoa_div(residual, scalefactor); |
| 367 |
|
|
int clamped = qoa_clamp(scaled, -8, 8); |
| 368 |
|
✗ |
int quantized = qoa_quant_tab[clamped + 8]; |
| 369 |
|
✗ |
int dequantized = qoa_dequant_tab[scalefactor][quantized]; |
| 370 |
|
✗ |
int reconstructed = qoa_clamp(predicted + dequantized, -32768, 32767); |
| 371 |
|
|
|
| 372 |
|
✗ |
long long error = (sample - reconstructed); |
| 373 |
|
✗ |
current_error += error * error; |
| 374 |
|
✗ |
if (current_error > best_error) { |
| 375 |
|
|
break; |
| 376 |
|
|
} |
| 377 |
|
|
|
| 378 |
|
✗ |
qoa_lms_update(&lms, reconstructed, dequantized); |
| 379 |
|
✗ |
slice = (slice << 3) | quantized; |
| 380 |
|
|
} |
| 381 |
|
|
|
| 382 |
|
✗ |
if (current_error < best_error) { |
| 383 |
|
|
best_error = current_error; |
| 384 |
|
|
best_slice = slice; |
| 385 |
|
✗ |
best_lms = lms; |
| 386 |
|
|
} |
| 387 |
|
|
} |
| 388 |
|
|
|
| 389 |
|
✗ |
qoa->lms[c] = best_lms; |
| 390 |
|
|
#ifdef QOA_RECORD_TOTAL_ERROR |
| 391 |
|
|
qoa->error += best_error; |
| 392 |
|
|
#endif |
| 393 |
|
|
|
| 394 |
|
|
/* If this slice was shorter than QOA_SLICE_LEN, we have to left- |
| 395 |
|
|
shift all encoded data, to ensure the rightmost bits are the empty |
| 396 |
|
|
ones. This should only happen in the last frame of a file as all |
| 397 |
|
|
slices are completely filled otherwise. */ |
| 398 |
|
✗ |
best_slice <<= (QOA_SLICE_LEN - slice_len) * 3; |
| 399 |
|
|
qoa_write_u64(best_slice, bytes, &p); |
| 400 |
|
|
} |
| 401 |
|
|
} |
| 402 |
|
|
|
| 403 |
|
✗ |
return p; |
| 404 |
|
|
} |
| 405 |
|
|
|
| 406 |
|
✗ |
void *qoa_encode(const short *sample_data, qoa_desc *qoa, unsigned int *out_len) { |
| 407 |
|
✗ |
if ( |
| 408 |
|
✗ |
qoa->samples == 0 || |
| 409 |
|
✗ |
qoa->samplerate == 0 || qoa->samplerate > 0xffffff || |
| 410 |
|
✗ |
qoa->channels == 0 || qoa->channels > QOA_MAX_CHANNELS |
| 411 |
|
|
) { |
| 412 |
|
|
return NULL; |
| 413 |
|
|
} |
| 414 |
|
|
|
| 415 |
|
|
/* Calculate the encoded size and allocate */ |
| 416 |
|
✗ |
unsigned int num_frames = (qoa->samples + QOA_FRAME_LEN-1) / QOA_FRAME_LEN; |
| 417 |
|
✗ |
unsigned int num_slices = (qoa->samples + QOA_SLICE_LEN-1) / QOA_SLICE_LEN; |
| 418 |
|
✗ |
unsigned int encoded_size = 8 + /* 8 byte file header */ |
| 419 |
|
✗ |
num_frames * 8 + /* 8 byte frame headers */ |
| 420 |
|
✗ |
num_frames * QOA_LMS_LEN * 4 * qoa->channels + /* 4 * 4 bytes lms state per channel */ |
| 421 |
|
✗ |
num_slices * 8 * qoa->channels; /* 8 byte slices */ |
| 422 |
|
|
|
| 423 |
|
✗ |
unsigned char *bytes = QOA_MALLOC(encoded_size); |
| 424 |
|
|
|
| 425 |
|
✗ |
for (int c = 0; c < qoa->channels; c++) { |
| 426 |
|
|
/* Set the initial LMS weights to {0, 0, -1, 2}. This helps with the |
| 427 |
|
|
prediction of the first few ms of a file. */ |
| 428 |
|
✗ |
qoa->lms[c].weights[0] = 0; |
| 429 |
|
✗ |
qoa->lms[c].weights[1] = 0; |
| 430 |
|
✗ |
qoa->lms[c].weights[2] = -(1<<13); |
| 431 |
|
✗ |
qoa->lms[c].weights[3] = (1<<14); |
| 432 |
|
|
|
| 433 |
|
|
/* Explicitly set the history samples to 0, as we might have some |
| 434 |
|
|
garbage in there. */ |
| 435 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN; i++) { |
| 436 |
|
✗ |
qoa->lms[c].history[i] = 0; |
| 437 |
|
|
} |
| 438 |
|
|
} |
| 439 |
|
|
|
| 440 |
|
|
|
| 441 |
|
|
/* Encode the header and go through all frames */ |
| 442 |
|
✗ |
unsigned int p = qoa_encode_header(qoa, bytes); |
| 443 |
|
|
#ifdef QOA_RECORD_TOTAL_ERROR |
| 444 |
|
|
qoa->error = 0; |
| 445 |
|
|
#endif |
| 446 |
|
|
|
| 447 |
|
|
int frame_len = QOA_FRAME_LEN; |
| 448 |
|
✗ |
for (int sample_index = 0; sample_index < qoa->samples; sample_index += frame_len) { |
| 449 |
|
✗ |
frame_len = qoa_clamp(QOA_FRAME_LEN, 0, qoa->samples - sample_index); |
| 450 |
|
✗ |
const short *frame_samples = sample_data + sample_index * qoa->channels; |
| 451 |
|
✗ |
unsigned int frame_size = qoa_encode_frame(frame_samples, qoa, frame_len, bytes + p); |
| 452 |
|
✗ |
p += frame_size; |
| 453 |
|
|
} |
| 454 |
|
|
|
| 455 |
|
✗ |
*out_len = p; |
| 456 |
|
✗ |
return bytes; |
| 457 |
|
|
} |
| 458 |
|
|
|
| 459 |
|
|
|
| 460 |
|
|
|
| 461 |
|
|
/* ----------------------------------------------------------------------------- |
| 462 |
|
|
Decoder */ |
| 463 |
|
|
|
| 464 |
|
✗ |
unsigned int qoa_max_frame_size(qoa_desc *qoa) { |
| 465 |
|
✗ |
return QOA_FRAME_SIZE(qoa->channels, QOA_SLICES_PER_FRAME); |
| 466 |
|
|
} |
| 467 |
|
|
|
| 468 |
|
✗ |
unsigned int qoa_decode_header(const unsigned char *bytes, int size, qoa_desc *qoa) { |
| 469 |
|
✗ |
unsigned int p = 0; |
| 470 |
|
✗ |
if (size < QOA_MIN_FILESIZE) { |
| 471 |
|
|
return 0; |
| 472 |
|
|
} |
| 473 |
|
|
|
| 474 |
|
|
|
| 475 |
|
|
/* Read the file header, verify the magic number ('qoaf') and read the |
| 476 |
|
|
total number of samples. */ |
| 477 |
|
✗ |
qoa_uint64_t file_header = qoa_read_u64(bytes, &p); |
| 478 |
|
|
|
| 479 |
|
✗ |
if ((file_header >> 32) != QOA_MAGIC) { |
| 480 |
|
|
return 0; |
| 481 |
|
|
} |
| 482 |
|
|
|
| 483 |
|
✗ |
qoa->samples = file_header & 0xffffffff; |
| 484 |
|
✗ |
if (!qoa->samples) { |
| 485 |
|
|
return 0; |
| 486 |
|
|
} |
| 487 |
|
|
|
| 488 |
|
|
/* Peek into the first frame header to get the number of channels and |
| 489 |
|
|
the samplerate. */ |
| 490 |
|
✗ |
qoa_uint64_t frame_header = qoa_read_u64(bytes, &p); |
| 491 |
|
✗ |
qoa->channels = (frame_header >> 56) & 0x0000ff; |
| 492 |
|
✗ |
qoa->samplerate = (frame_header >> 32) & 0xffffff; |
| 493 |
|
|
|
| 494 |
|
✗ |
if (qoa->channels == 0 || qoa->samples == 0 || qoa->samplerate == 0) { |
| 495 |
|
✗ |
return 0; |
| 496 |
|
|
} |
| 497 |
|
|
|
| 498 |
|
|
return 8; |
| 499 |
|
|
} |
| 500 |
|
|
|
| 501 |
|
✗ |
unsigned int qoa_decode_frame(const unsigned char *bytes, unsigned int size, qoa_desc *qoa, short *sample_data, unsigned int *frame_len) { |
| 502 |
|
✗ |
unsigned int p = 0; |
| 503 |
|
✗ |
*frame_len = 0; |
| 504 |
|
|
|
| 505 |
|
✗ |
if (size < 8 + QOA_LMS_LEN * 4 * qoa->channels) { |
| 506 |
|
|
return 0; |
| 507 |
|
|
} |
| 508 |
|
|
|
| 509 |
|
|
/* Read and verify the frame header */ |
| 510 |
|
✗ |
qoa_uint64_t frame_header = qoa_read_u64(bytes, &p); |
| 511 |
|
✗ |
int channels = (frame_header >> 56) & 0x0000ff; |
| 512 |
|
✗ |
int samplerate = (frame_header >> 32) & 0xffffff; |
| 513 |
|
✗ |
int samples = (frame_header >> 16) & 0x00ffff; |
| 514 |
|
✗ |
int frame_size = (frame_header ) & 0x00ffff; |
| 515 |
|
|
|
| 516 |
|
✗ |
int data_size = frame_size - 8 - QOA_LMS_LEN * 4 * channels; |
| 517 |
|
✗ |
int num_slices = data_size / 8; |
| 518 |
|
✗ |
int max_total_samples = num_slices * QOA_SLICE_LEN; |
| 519 |
|
|
|
| 520 |
|
✗ |
if ( |
| 521 |
|
✗ |
channels != qoa->channels || |
| 522 |
|
✗ |
samplerate != qoa->samplerate || |
| 523 |
|
✗ |
frame_size > size || |
| 524 |
|
✗ |
samples * channels > max_total_samples |
| 525 |
|
|
) { |
| 526 |
|
|
return 0; |
| 527 |
|
|
} |
| 528 |
|
|
|
| 529 |
|
|
|
| 530 |
|
|
/* Read the LMS state: 4 x 2 bytes history, 4 x 2 bytes weights per channel */ |
| 531 |
|
✗ |
for (int c = 0; c < channels; c++) { |
| 532 |
|
✗ |
qoa_uint64_t history = qoa_read_u64(bytes, &p); |
| 533 |
|
✗ |
qoa_uint64_t weights = qoa_read_u64(bytes, &p); |
| 534 |
|
✗ |
for (int i = 0; i < QOA_LMS_LEN; i++) { |
| 535 |
|
✗ |
qoa->lms[c].history[i] = ((signed short)(history >> 48)); |
| 536 |
|
✗ |
history <<= 16; |
| 537 |
|
✗ |
qoa->lms[c].weights[i] = ((signed short)(weights >> 48)); |
| 538 |
|
✗ |
weights <<= 16; |
| 539 |
|
|
} |
| 540 |
|
|
} |
| 541 |
|
|
|
| 542 |
|
|
|
| 543 |
|
|
/* Decode all slices for all channels in this frame */ |
| 544 |
|
✗ |
for (int sample_index = 0; sample_index < samples; sample_index += QOA_SLICE_LEN) { |
| 545 |
|
✗ |
for (int c = 0; c < channels; c++) { |
| 546 |
|
✗ |
qoa_uint64_t slice = qoa_read_u64(bytes, &p); |
| 547 |
|
|
|
| 548 |
|
✗ |
int scalefactor = (slice >> 60) & 0xf; |
| 549 |
|
✗ |
int slice_start = sample_index * channels + c; |
| 550 |
|
✗ |
int slice_end = qoa_clamp(sample_index + QOA_SLICE_LEN, 0, samples) * channels + c; |
| 551 |
|
|
|
| 552 |
|
✗ |
for (int si = slice_start; si < slice_end; si += channels) { |
| 553 |
|
|
int predicted = qoa_lms_predict(&qoa->lms[c]); |
| 554 |
|
✗ |
int quantized = (slice >> 57) & 0x7; |
| 555 |
|
✗ |
int dequantized = qoa_dequant_tab[scalefactor][quantized]; |
| 556 |
|
✗ |
int reconstructed = qoa_clamp(predicted + dequantized, -32768, 32767); |
| 557 |
|
|
|
| 558 |
|
✗ |
sample_data[si] = reconstructed; |
| 559 |
|
✗ |
slice <<= 3; |
| 560 |
|
|
|
| 561 |
|
✗ |
qoa_lms_update(&qoa->lms[c], reconstructed, dequantized); |
| 562 |
|
|
} |
| 563 |
|
|
} |
| 564 |
|
|
} |
| 565 |
|
|
|
| 566 |
|
✗ |
*frame_len = samples; |
| 567 |
|
✗ |
return p; |
| 568 |
|
|
} |
| 569 |
|
|
|
| 570 |
|
✗ |
short *qoa_decode(const unsigned char *bytes, int size, qoa_desc *qoa) { |
| 571 |
|
✗ |
unsigned int p = qoa_decode_header(bytes, size, qoa); |
| 572 |
|
✗ |
if (!p) { |
| 573 |
|
|
return NULL; |
| 574 |
|
|
} |
| 575 |
|
|
|
| 576 |
|
|
/* Calculate the required size of the sample buffer and allocate */ |
| 577 |
|
✗ |
int total_samples = qoa->samples * qoa->channels; |
| 578 |
|
✗ |
short *sample_data = QOA_MALLOC(total_samples * sizeof(short)); |
| 579 |
|
|
|
| 580 |
|
|
unsigned int sample_index = 0; |
| 581 |
|
|
unsigned int frame_len; |
| 582 |
|
|
unsigned int frame_size; |
| 583 |
|
|
|
| 584 |
|
|
/* Decode all frames */ |
| 585 |
|
|
do { |
| 586 |
|
✗ |
short *sample_ptr = sample_data + sample_index * qoa->channels; |
| 587 |
|
✗ |
frame_size = qoa_decode_frame(bytes + p, size - p, qoa, sample_ptr, &frame_len); |
| 588 |
|
|
|
| 589 |
|
✗ |
p += frame_size; |
| 590 |
|
✗ |
sample_index += frame_len; |
| 591 |
|
✗ |
} while (frame_size && sample_index < qoa->samples); |
| 592 |
|
|
|
| 593 |
|
✗ |
qoa->samples = sample_index; |
| 594 |
|
✗ |
return sample_data; |
| 595 |
|
|
} |
| 596 |
|
|
|
| 597 |
|
|
|
| 598 |
|
|
|
| 599 |
|
|
/* ----------------------------------------------------------------------------- |
| 600 |
|
|
File read/write convenience functions */ |
| 601 |
|
|
|
| 602 |
|
|
#ifndef QOA_NO_STDIO |
| 603 |
|
|
#include <stdio.h> |
| 604 |
|
|
|
| 605 |
|
✗ |
int qoa_write(const char *filename, const short *sample_data, qoa_desc *qoa) { |
| 606 |
|
✗ |
FILE *f = fopen(filename, "wb"); |
| 607 |
|
|
unsigned int size; |
| 608 |
|
|
void *encoded; |
| 609 |
|
|
|
| 610 |
|
✗ |
if (!f) { |
| 611 |
|
|
return 0; |
| 612 |
|
|
} |
| 613 |
|
|
|
| 614 |
|
✗ |
encoded = qoa_encode(sample_data, qoa, &size); |
| 615 |
|
✗ |
if (!encoded) { |
| 616 |
|
✗ |
fclose(f); |
| 617 |
|
✗ |
return 0; |
| 618 |
|
|
} |
| 619 |
|
|
|
| 620 |
|
✗ |
fwrite(encoded, 1, size, f); |
| 621 |
|
✗ |
fclose(f); |
| 622 |
|
|
|
| 623 |
|
✗ |
QOA_FREE(encoded); |
| 624 |
|
✗ |
return size; |
| 625 |
|
|
} |
| 626 |
|
|
|
| 627 |
|
✗ |
void *qoa_read(const char *filename, qoa_desc *qoa) { |
| 628 |
|
✗ |
FILE *f = fopen(filename, "rb"); |
| 629 |
|
|
int size, bytes_read; |
| 630 |
|
|
void *data; |
| 631 |
|
|
short *sample_data; |
| 632 |
|
|
|
| 633 |
|
✗ |
if (!f) { |
| 634 |
|
|
return NULL; |
| 635 |
|
|
} |
| 636 |
|
|
|
| 637 |
|
✗ |
fseek(f, 0, SEEK_END); |
| 638 |
|
✗ |
size = ftell(f); |
| 639 |
|
✗ |
if (size <= 0) { |
| 640 |
|
✗ |
fclose(f); |
| 641 |
|
✗ |
return NULL; |
| 642 |
|
|
} |
| 643 |
|
✗ |
fseek(f, 0, SEEK_SET); |
| 644 |
|
|
|
| 645 |
|
✗ |
data = QOA_MALLOC(size); |
| 646 |
|
✗ |
if (!data) { |
| 647 |
|
✗ |
fclose(f); |
| 648 |
|
✗ |
return NULL; |
| 649 |
|
|
} |
| 650 |
|
|
|
| 651 |
|
✗ |
bytes_read = fread(data, 1, size, f); |
| 652 |
|
✗ |
fclose(f); |
| 653 |
|
|
|
| 654 |
|
✗ |
sample_data = qoa_decode(data, bytes_read, qoa); |
| 655 |
|
✗ |
QOA_FREE(data); |
| 656 |
|
✗ |
return sample_data; |
| 657 |
|
|
} |
| 658 |
|
|
|
| 659 |
|
|
#endif /* QOA_NO_STDIO */ |
| 660 |
|
|
#endif /* QOA_IMPLEMENTATION */ |
| 661 |
|
|
|