curioustorvald/tsvm
1
0
Fork 0
mirror of https://github.com/curioustorvald/tsvm.git synced 2026-03-07 19:51:51 +09:00
Code Issues Packages Projects Releases Wiki Activity

TAD: now processing entirely in float

Browse Source
This commit is contained in:
minjaesong
2025-10-24 05:31:38 +09:00
parent a9319fd812
commit 9dc71095a0
5 changed files with 139 additions and 319 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
terranmon.txt
View File

@@ -1550,7 +1550,7 @@ is stored separately and quality index is shared with that of the video.
## Audio Properties
- **Sample Rate**: 32000 Hz (TSVM audio hardware native format)
- **Channels**: 2 (stereo)
- **Input Format**: PCM16LE (16-bit signed little-endian PCM)
- **Input Format**: PCM32fLE (32-bit float little-endian PCM)
- **Preprocessing**: 16 Hz highpass filter applied during extraction
- **Internal Representation**: Signed PCM8 with error-diffusion dithering
- **Chunk Size**: Variable (1024-32768+ samples per channel, must be power of 2)
@@ -1565,8 +1565,6 @@ Default is 32768 samples (65536 total samples, 1.024 seconds).
If the audio duration doesn't align to chunk boundaries, the final chunk can use
a smaller power-of-2 size or be zero-padded.
uint8 Significance Map Method: always 1 (2-bit twobitmap)
uint8 Compression Flag: 1=Zstd compressed, 0=uncompressed
uint16 Sample Count: number of samples per channel (must be power of 2, min 1024)
uint32 Chunk Payload Size: size of following payload in bytes
* Chunk Payload: encoded M/S stereo data (Zstd compressed if flag set)
@@ -1592,13 +1590,9 @@ as int16 in the order they appear.
## Encoding Pipeline
### Step 1: PCM16 to PCM8 Conversion with Error-Diffusion Dithering
Input stereo PCM16LE is converted to signed PCM8 using error-diffusion dithering
to minimize quantization noise:
dithered_value = pcm16_value / 256 + error
pcm8_value = clamp(round(dithered_value), -128, 127)
error = dithered_value - pcm8_value
### Step 1: PCM32f to PCM8 Conversion with Error-Diffusion Dithering
Input stereo PCM32fLE is converted to signed PCM8 using second-order noise-shaped
error-diffusion dithering to minimize quantization noise.
Error is propagated to the next sample (alternating between left/right channels).
@@ -1632,18 +1626,7 @@ For 32768 samples with 14 levels: boundaries at 0, 2, 4, 8, 16, 32, 64, 128, 256
For 1024 samples with 9 levels: boundaries at 0, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024
### Step 4: Frequency-Dependent Quantization
DWT coefficients are quantized using perceptually-tuned frequency-dependent weights:
Base Weights by Level:
Level 0 (16-8 KHz): 3.0
Level 1 (8-4 KHz): 2.0
Level 2 (4-2 KHz): 1.5
Level 3 (2-1 KHz): 1.0
Level 4 (1-0.5 KHz): 0.75
Level 5 (0.5-0.25 KHz): 0.5
Level 6-7 (DC-0.25 KHz): 0.25
Quality scaling factor: 1.0 + (5 - quality) * 0.3
DWT coefficients are quantized using perceptually-tuned frequency-dependent weights.
Final quantization step: base_weight * quality_scale
@@ -1690,13 +1673,8 @@ Convert Mid/Side back to Left/Right stereo:
Left = Mid + Side
Right = Mid - Side
### Step 6: PCM8 to PCM16 Upsampling
Convert signed PCM8 back to PCM16LE by multiplying by 256:
pcm16_value = pcm8_value * 256
## Compression Performance
- **Target Ratio**: 2:1 against PCMu8 (4:1 against PCM16LE input)
- **Target Ratio**: 2:1 against PCMu8
- **Achieved Ratio**: 2.51:1 against PCMu8 at quality level 3
- **Quality**: Perceptually transparent at Q3+, preserves full 0-16 KHz bandwidth
- **Sparsity**: 86.9% zeros in Mid channel, 97.8% in Side channel (typical)
@@ -1721,10 +1699,10 @@ This allows TAV video files to embed TAD-compressed audio using packet type 0x24
TAD encoder uses two-pass FFmpeg extraction for optimal quality:
# Pass 1: Extract at original sample rate
ffmpeg -i input.mp4 -f s16le -ac 2 temp.pcm
ffmpeg -i input.mp4 -f f32le -ac 2 temp.pcm
# Pass 2: High-quality resample with SoXR and highpass filter
ffmpeg -f s16le -ar {original_rate} -ac 2 -i temp.pcm \
ffmpeg -f f32le -ar {original_rate} -ac 2 -i temp.pcm \
-ar 32000 -af "aresample=resampler=soxr:precision=28:cutoff=0.99,highpass=f=16" \
output.pcm

12
video_encoder/Makefile
View File

@@ -78,7 +78,7 @@ debug: $(TARGETS)
# Clean build artifacts
clean:
rm -f $(TARGETS) $(TAD_TARGETS) *.o
rm -f $(TARGETS) $(TAD_TARGETS) $(TAD16_TARGETS) $(TAD10_TARGETS) *.o
# Install (copy to PATH)
install: $(TARGETS) $(TAD_TARGETS)
@@ -106,6 +106,12 @@ help:
@echo " tad - Build all TAD audio tools (encoder, decoder)"
@echo " encoder_tad - Build TAD audio encoder"
@echo " decoder_tad - Build TAD audio decoder"
@echo " tad16 - Build TAD16 tools (PCM16 alternative for comparison)"
@echo " encoder_tad16- Build TAD16 audio encoder (PCM16 version)"
@echo " decoder_tad16- Build TAD16 audio decoder (PCM16 version)"
@echo " tad10 - Build TAD10 tools (PCM10 alternative for comparison)"
@echo " encoder_tad10- Build TAD10 audio encoder (PCM10 version)"
@echo " decoder_tad10- Build TAD10 audio decoder (PCM10 version)"
@echo " debug - Build with debug symbols"
@echo " clean - Remove build artifacts"
@echo " install - Install to /usr/local/bin"
@@ -117,6 +123,8 @@ help:
@echo " make tev # Build TEV encoder"
@echo " make tav # Build TAV encoder"
@echo " make tad # Build all TAD audio tools"
@echo " make tad16 # Build TAD16 tools (for comparison testing)"
@echo " make tad10 # Build TAD10 tools (for comparison testing)"
@echo " sudo make install # Install all encoders"
.PHONY: all clean install check-deps help debug tad
.PHONY: all clean install check-deps help debug tad tad16 tad10

160
video_encoder/decoder_tad.c
View File

@@ -12,6 +12,7 @@
#define DECODER_VENDOR_STRING "Decoder-TAD 20251023"
// TAD format constants (must match encoder)
#define TAD_COEFF_SCALAR 1024.0f
#define TAD_DEFAULT_CHUNK_SIZE 32768
#define TAD_MIN_CHUNK_SIZE 1024
#define TAD_SAMPLE_RATE 32000
@@ -148,22 +149,58 @@ static void dwt_haar_inverse_multilevel(float *data, int length, int levels) {
// M/S Stereo Correlation (inverse of decorrelation)
//=============================================================================
static void ms_correlate(const int8_t *mid, const int8_t *side, uint8_t *left, uint8_t *right, size_t count) {
// Uniform random in [0, 1)
static inline float frand01(void) {
return (float)rand() / ((float)RAND_MAX + 1.0f);
}
// TPDF noise in [-1, +1)
static inline float tpdf1(void) {
return (frand01() - frand01());
}
static void ms_correlate(const float *mid, const float *side, uint8_t *left, uint8_t *right, size_t count, float dither_error[2][2]) {
const float b1 = 1.5f; // 1st feedback coefficient
const float b2 = -0.75f; // 2nd feedback coefficient
const float scale = 127.5f;
const float bias = 128.0f;
for (size_t i = 0; i < count; i++) {
// L = M + S, R = M - S
int32_t m = mid[i];
int32_t s = side[i];
int32_t l = m + s;
int32_t r = m - s;
// Decode M/S → L/R
float m = mid[i];
float s = side[i];
float l = FCLAMP(m + s, -1.0f, 1.0f);
float r = FCLAMP(m - s, -1.0f, 1.0f);
// Clamp to [-128, 127] then convert to unsigned [0, 255]
if (l < -128) l = -128;
if (l > 127) l = 127;
if (r < -128) r = -128;
if (r > 127) r = 127;
// --- LEFT channel ---
float feedbackL = b1 * dither_error[0][0] + b2 * dither_error[0][1];
float ditherL = 0.5f * tpdf1(); // ±0.5 LSB TPDF
float shapedL = l + feedbackL + ditherL / scale;
shapedL = FCLAMP(shapedL, -1.0f, 1.0f);
left[i] = (uint8_t)(l + 128);
right[i] = (uint8_t)(r + 128);
int qL = (int)lrintf(shapedL * scale);
if (qL < -128) qL = -128;
else if (qL > 127) qL = 127;
left[i] = (uint8_t)(qL + bias);
float qerrL = shapedL - (float)qL / scale;
dither_error[0][1] = dither_error[0][0]; // shift history
dither_error[0][0] = qerrL;
// --- RIGHT channel ---
float feedbackR = b1 * dither_error[1][0] + b2 * dither_error[1][1];
float ditherR = 0.5f * tpdf1();
float shapedR = r + feedbackR + ditherR / scale;
shapedR = FCLAMP(shapedR, -1.0f, 1.0f);
int qR = (int)lrintf(shapedR * scale);
if (qR < -128) qR = -128;
else if (qR > 127) qR = 127;
right[i] = (uint8_t)(qR + bias);
float qerrR = shapedR - (float)qR / scale;
dither_error[1][1] = dither_error[1][0];
dither_error[1][0] = qerrR;
}
}
@@ -188,11 +225,10 @@ static void get_quantization_weights(int quality, int dwt_levels, float *weights
/*12*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f, 1.5f, 1.5f, 1.5f},
/*13*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f, 1.5f, 1.5f},
/*14*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f, 1.5f},
/*15*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f},
/*16*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f}
/*15*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f}
};
float quality_scale = 1.0f + FCLAMP((3 - quality) * 0.5f, 0.0f, 1000.0f);
float quality_scale = 4.0f + FCLAMP((3 - quality) * 0.5f, 0.0f, 1000.0f);
for (int i = 0; i < dwt_levels; i++) {
weights[i] = FCLAMP(base_weights[dwt_levels][i] * quality_scale, 1.0f, 1000.0f);
@@ -227,7 +263,7 @@ static void dequantize_dwt_coefficients(const int16_t *quantized, float *coeffs,
if (weight_idx >= dwt_levels) weight_idx = dwt_levels - 1;
float weight = weights[weight_idx];
coeffs[i] = (float)quantized[i] * weight;
coeffs[i] = (float)quantized[i] * weight / TAD_COEFF_SCALAR;
}
free(sideband_starts);
@@ -237,29 +273,6 @@ static void dequantize_dwt_coefficients(const int16_t *quantized, float *coeffs,
// Significance Map Decoding
//=============================================================================
static size_t decode_sigmap_1bit(const uint8_t *input, int16_t *values, size_t count) {
size_t map_bytes = (count + 7) / 8;
const uint8_t *map = input;
const uint8_t *read_ptr = input + map_bytes;
uint32_t nonzero_count = *((const uint32_t*)read_ptr);
read_ptr += sizeof(uint32_t);
const int16_t *value_ptr = (const int16_t*)read_ptr;
uint32_t value_idx = 0;
// Reconstruct values
for (size_t i = 0; i < count; i++) {
if (map[i / 8] & (1 << (i % 8))) {
values[i] = value_ptr[value_idx++];
} else {
values[i] = 0;
}
}
return map_bytes + sizeof(uint32_t) + nonzero_count * sizeof(int16_t);
}
static size_t decode_sigmap_2bit(const uint8_t *input, int16_t *values, size_t count) {
size_t map_bytes = (count * 2 + 7) / 8;
const uint8_t *map = input;
@@ -291,48 +304,6 @@ static size_t decode_sigmap_2bit(const uint8_t *input, int16_t *values, size_t c
return map_bytes + other_idx * sizeof(int16_t);
}
static size_t decode_sigmap_rle(const uint8_t *input, int16_t *values, size_t count) {
const uint8_t *read_ptr = input;
uint32_t run_count = *((const uint32_t*)read_ptr);
read_ptr += sizeof(uint32_t);
size_t value_idx = 0;
for (uint32_t run = 0; run < run_count; run++) {
// Decode zero run length (varint)
uint32_t zero_run = 0;
int shift = 0;
uint8_t byte;
do {
byte = *read_ptr++;
zero_run |= ((uint32_t)(byte & 0x7F) << shift);
shift += 7;
} while (byte & 0x80);
// Fill zeros
for (uint32_t i = 0; i < zero_run && value_idx < count; i++) {
values[value_idx++] = 0;
}
// Read non-zero value
int16_t val = *((const int16_t*)read_ptr);
read_ptr += sizeof(int16_t);
if (value_idx < count && val != 0) {
values[value_idx++] = val;
}
}
// Fill remaining with zeros
while (value_idx < count) {
values[value_idx++] = 0;
}
return read_ptr - input;
}
//=============================================================================
// Chunk Decoding
//=============================================================================
@@ -381,8 +352,6 @@ static int decode_chunk(const uint8_t *input, size_t input_size, uint8_t *pcmu8_
int16_t *quant_side = malloc(sample_count * sizeof(int16_t));
float *dwt_mid = malloc(sample_count * sizeof(float));
float *dwt_side = malloc(sample_count * sizeof(float));
int8_t *pcm8_mid = malloc(sample_count * sizeof(int8_t));
int8_t *pcm8_side = malloc(sample_count * sizeof(int8_t));
uint8_t *pcm8_left = malloc(sample_count * sizeof(uint8_t));
uint8_t *pcm8_right = malloc(sample_count * sizeof(uint8_t));
@@ -401,23 +370,10 @@ static int decode_chunk(const uint8_t *input, size_t input_size, uint8_t *pcmu8_
dwt_haar_inverse_multilevel(dwt_mid, sample_count, dwt_levels);
dwt_haar_inverse_multilevel(dwt_side, sample_count, dwt_levels);
// Convert to signed PCM8
for (size_t i = 0; i < sample_count; i++) {
float m = dwt_mid[i];
float s = dwt_side[i];
// Clamp and round
if (m < -128.0f) m = -128.0f;
if (m > 127.0f) m = 127.0f;
if (s < -128.0f) s = -128.0f;
if (s > 127.0f) s = 127.0f;
pcm8_mid[i] = (int8_t)roundf(m);
pcm8_side[i] = (int8_t)roundf(s);
}
float err[2][2] = {{0,0},{0,0}};
// M/S to L/R correlation
ms_correlate(pcm8_mid, pcm8_side, pcm8_left, pcm8_right, sample_count);
ms_correlate(dwt_mid, dwt_side, pcm8_left, pcm8_right, sample_count, err);
// Interleave stereo output (PCMu8)
for (size_t i = 0; i < sample_count; i++) {
@@ -427,7 +383,7 @@ static int decode_chunk(const uint8_t *input, size_t input_size, uint8_t *pcmu8_
// Cleanup
free(quant_mid); free(quant_side); free(dwt_mid); free(dwt_side);
free(pcm8_mid); free(pcm8_side); free(pcm8_left); free(pcm8_right);
free(pcm8_left); free(pcm8_right);
if (decompressed) free(decompressed);
return 0;
@@ -442,7 +398,7 @@ static void print_usage(const char *prog_name) {
printf("Options:\n");
printf(" -i <file> Input TAD file\n");
printf(" -o <file> Output PCMu8 file (raw 8-bit unsigned stereo @ 32kHz)\n");
printf(" -q <0-5> Quality level used during encoding (default: 2)\n");
printf(" -q <0-5> Quality level used during encoding (default: 3)\n");
printf(" -v Verbose output\n");
printf(" -h, --help Show this help\n");
printf("\nVersion: %s\n", DECODER_VENDOR_STRING);
@@ -453,7 +409,7 @@ static void print_usage(const char *prog_name) {
int main(int argc, char *argv[]) {
char *input_file = NULL;
char *output_file = NULL;
int quality = 2; // Must match encoder quality
int quality = 3; // Must match encoder quality
int verbose = 0;
int opt;

208
video_encoder/encoder_tad.c
View File

@@ -1,6 +1,7 @@
// Created by CuriousTorvald and Claude on 2025-10-23.
// TAD (Terrarum Advanced Audio) Encoder Library - DWT-based audio compression
// This file contains only the encoding functions for use by encoder_tad.c and encoder_tav.c
// Created by CuriousTorvald and Claude on 2025-10-24.
// TAD32 (Terrarum Advanced Audio - PCM32f version) Encoder Library
// Alternative version: PCM32f throughout encoding, PCM8 conversion only at decoder
// This file contains only the encoding functions for comparison testing
#include <stdio.h>
#include <stdlib.h>
@@ -11,12 +12,9 @@
#include "encoder_tad.h"
// Forward declarations for internal functions
static void dwt_haar_forward_1d(float *data, int length);
static void dwt_dd4_forward_1d(float *data, int length);
static void dwt_97_forward_1d(float *data, int length);
static void dwt_haar_forward_multilevel(float *data, int length, int levels);
static void ms_decorrelate(const int8_t *left, const int8_t *right, int8_t *mid, int8_t *side, size_t count);
static void convert_pcm16_to_pcm8_dithered(const int16_t *pcm16, int8_t *pcm8, int num_samples, int16_t *dither_error);
static void dwt_dd4_forward_multilevel(float *data, int length, int levels);
static void ms_decorrelate_16(const float *left, const float *right, float *mid, float *side, size_t count);
static void get_quantization_weights(int quality, int dwt_levels, float *weights);
static int get_deadzone_threshold(int quality);
static void quantize_dwt_coefficients(const float *coeffs, int16_t *quantized, size_t count, int quality, int apply_deadzone, int chunk_size, int dwt_levels);
@@ -26,15 +24,13 @@ static inline float FCLAMP(float x, float min, float max) {
return x < min ? min : (x > max ? max : x);
}
// Calculate DWT levels from chunk size (non-power-of-2 supported, >= 1024)
// Calculate DWT levels from chunk size
static int calculate_dwt_levels(int chunk_size) {
if (chunk_size < TAD_MIN_CHUNK_SIZE) {
fprintf(stderr, "Error: Chunk size %d is below minimum %d\n", chunk_size, TAD_MIN_CHUNK_SIZE);
if (chunk_size < TAD32_MIN_CHUNK_SIZE) {
fprintf(stderr, "Error: Chunk size %d is below minimum %d\n", chunk_size, TAD32_MIN_CHUNK_SIZE);
return -1;
}
// For non-power-of-2, find next power of 2 and calculate levels
// Then subtract 2 for maximum decomposition
int levels = 0;
int size = chunk_size;
while (size > 1) {
@@ -48,39 +44,13 @@ static int calculate_dwt_levels(int chunk_size) {
levels++;
}
return levels - 2; // Maximum decomposition leaves 2-sample approximation
return levels - 2; // Maximum decomposition
}
//=============================================================================
// Haar DWT Implementation
// DD-4 DWT Implementation
//=============================================================================
static void dwt_haar_forward_1d(float *data, int length) {
if (length < 2) return;
float *temp = malloc(length * sizeof(float));
int half = (length + 1) / 2;
// Haar transform: compute averages (low-pass) and differences (high-pass)
for (int i = 0; i < half; i++) {
if (2 * i + 1 < length) {
// Average of adjacent pairs (low-pass)
temp[i] = (data[2 * i] + data[2 * i + 1]) / 2.0f;
// Difference of adjacent pairs (high-pass)
temp[half + i] = (data[2 * i] - data[2 * i + 1]) / 2.0f;
} else {
// Handle odd length: last sample goes to low-pass
temp[i] = data[2 * i];
if (half + i < length) {
temp[half + i] = 0.0f;
}
}
}
memcpy(data, temp, length * sizeof(float));
free(temp);
}
// Four-point interpolating Deslauriers-Dubuc (DD-4) wavelet forward 1D transform
static void dwt_dd4_forward_1d(float *data, int length) {
if (length < 2) return;
@@ -129,76 +99,8 @@ static void dwt_dd4_forward_1d(float *data, int length) {
free(temp);
}
// 1D DWT using lifting scheme for 9/7 irreversible filter
static void dwt_97_forward_1d(float *data, int length) {
if (length < 2) return;
float *temp = malloc(length * sizeof(float));
int half = (length + 1) / 2;
// Split into even/odd samples
for (int i = 0; i < half; i++) {
temp[i] = data[2 * i]; // Even (low)
}
for (int i = 0; i < length / 2; i++) {
temp[half + i] = data[2 * i + 1]; // Odd (high)
}
// JPEG2000 9/7 forward lifting steps
const float alpha = -1.586134342f;
const float beta = -0.052980118f;
const float gamma = 0.882911076f;
const float delta = 0.443506852f;
const float K = 1.230174105f;
// Step 1: Predict α
for (int i = 0; i < length / 2; i++) {
if (half + i < length) {
float s_curr = temp[i];
float s_next = (i + 1 < half) ? temp[i + 1] : s_curr;
temp[half + i] += alpha * (s_curr + s_next);
}
}
// Step 2: Update β
for (int i = 0; i < half; i++) {
float d_curr = (half + i < length) ? temp[half + i] : 0.0f;
float d_prev = (i > 0 && half + i - 1 < length) ? temp[half + i - 1] : d_curr;
temp[i] += beta * (d_prev + d_curr);
}
// Step 3: Predict γ
for (int i = 0; i < length / 2; i++) {
if (half + i < length) {
float s_curr = temp[i];
float s_next = (i + 1 < half) ? temp[i + 1] : s_curr;
temp[half + i] += gamma * (s_curr + s_next);
}
}
// Step 4: Update δ
for (int i = 0; i < half; i++) {
float d_curr = (half + i < length) ? temp[half + i] : 0.0f;
float d_prev = (i > 0 && half + i - 1 < length) ? temp[half + i - 1] : d_curr;
temp[i] += delta * (d_prev + d_curr);
}
// Step 5: Scaling
for (int i = 0; i < half; i++) {
temp[i] *= K;
}
for (int i = 0; i < length / 2; i++) {
if (half + i < length) {
temp[half + i] /= K;
}
}
memcpy(data, temp, length * sizeof(float));
free(temp);
}
// Apply multi-level DWT (using DD-4 wavelet)
static void dwt_haar_forward_multilevel(float *data, int length, int levels) {
static void dwt_dd4_forward_multilevel(float *data, int length, int levels) {
int current_length = length;
for (int level = 0; level < levels; level++) {
dwt_dd4_forward_1d(data, current_length);
@@ -207,35 +109,16 @@ static void dwt_haar_forward_multilevel(float *data, int length, int levels) {
}
//=============================================================================
// M/S Stereo Decorrelation
// M/S Stereo Decorrelation (PCM32f version)
//=============================================================================
static void ms_decorrelate(const int8_t *left, const int8_t *right, int8_t *mid, int8_t *side, size_t count) {
static void ms_decorrelate_16(const float *left, const float *right, float *mid, float *side, size_t count) {
for (size_t i = 0; i < count; i++) {
// Mid = (L + R) / 2, Side = (L - R) / 2
int32_t l = left[i];
int32_t r = right[i];
mid[i] = (int8_t)((l + r) / 2);
side[i] = (int8_t)((l - r) / 2);
}
}
//=============================================================================
// PCM16 to Signed PCM8 Conversion with Dithering
//=============================================================================
static void convert_pcm16_to_pcm8_dithered(const int16_t *pcm16, int8_t *pcm8, int num_samples, int16_t *dither_error) {
for (int i = 0; i < num_samples; i++) {
for (int ch = 0; ch < 2; ch++) { // Stereo: L and R
int idx = i * 2 + ch;
int32_t sample = (int32_t)pcm16[idx];
sample += dither_error[ch];
int32_t quantized = sample >> 8;
if (quantized < -128) quantized = -128;
if (quantized > 127) quantized = 127;
pcm8[idx] = (int8_t)quantized;
dither_error[ch] = sample - (quantized << 8);
}
float l = left[i];
float r = right[i];
mid[i] = (l + r) / 2.0f;
side[i] = (l - r) / 2.0f;
}
}
@@ -263,15 +146,15 @@ static void get_quantization_weights(int quality, int dwt_levels, float *weights
/*15*/{0.2f, 0.2f, 0.8f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.25f, 1.5f, 1.5f}
};
float quality_scale = 1.0f + FCLAMP((3 - quality) * 0.5f, 0.0f, 1000.0f);
float quality_scale = 4.0f * (1.0f + FCLAMP((3 - quality) * 0.5f, 0.0f, 1000.0f));
for (int i = 0; i < dwt_levels; i++) {
weights[i] = FCLAMP(base_weights[dwt_levels][i] * quality_scale, 1.0f, 1000.0f);
weights[i] = base_weights[dwt_levels][i] * quality_scale;
}
}
static int get_deadzone_threshold(int quality) {
const int thresholds[] = {1,1,0,0,0,0}; // Q0 to Q5
const int thresholds[] = {1,1,1,1,1,1}; // Q0 to Q5
return thresholds[quality];
}
@@ -302,7 +185,7 @@ static void quantize_dwt_coefficients(const float *coeffs, int16_t *quantized, s
if (weight_idx >= dwt_levels) weight_idx = dwt_levels - 1;
float weight = weights[weight_idx];
float val = coeffs[i] / weight;
float val = coeffs[i] / weight * TAD32_COEFF_SCALAR;
int16_t quant_val = (int16_t)roundf(val);
if (apply_deadzone && sideband >= dwt_levels - 1) {
@@ -359,7 +242,7 @@ static size_t encode_sigmap_2bit(const int16_t *values, size_t count, uint8_t *o
// Public API: Chunk Encoding
//=============================================================================
size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int quality,
size_t tad32_encode_chunk(const float *pcm32_stereo, size_t num_samples, int quality,
int use_zstd, uint8_t *output) {
// Calculate DWT levels from chunk size
int dwt_levels = calculate_dwt_levels(num_samples);
@@ -368,12 +251,11 @@ size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int qua
return 0;
}
// Allocate working buffers
int8_t *pcm8_stereo = malloc(num_samples * 2 * sizeof(int8_t));
int8_t *pcm8_left = malloc(num_samples * sizeof(int8_t));
int8_t *pcm8_right = malloc(num_samples * sizeof(int8_t));
int8_t *pcm8_mid = malloc(num_samples * sizeof(int8_t));
int8_t *pcm8_side = malloc(num_samples * sizeof(int8_t));
// Allocate working buffers (PCM32f throughout, int32 coefficients)
float *pcm32_left = malloc(num_samples * sizeof(float));
float *pcm32_right = malloc(num_samples * sizeof(float));
float *pcm32_mid = malloc(num_samples * sizeof(float));
float *pcm32_side = malloc(num_samples * sizeof(float));
float *dwt_mid = malloc(num_samples * sizeof(float));
float *dwt_side = malloc(num_samples * sizeof(float));
@@ -381,34 +263,30 @@ size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int qua
int16_t *quant_mid = malloc(num_samples * sizeof(int16_t));
int16_t *quant_side = malloc(num_samples * sizeof(int16_t));
// Step 1: Convert PCM16 to signed PCM8 with dithering
int16_t dither_error[2] = {0, 0};
convert_pcm16_to_pcm8_dithered(pcm16_stereo, pcm8_stereo, num_samples, dither_error);
// Deinterleave stereo
// Step 1: Deinterleave stereo
for (size_t i = 0; i < num_samples; i++) {
pcm8_left[i] = pcm8_stereo[i * 2];
pcm8_right[i] = pcm8_stereo[i * 2 + 1];
pcm32_left[i] = pcm32_stereo[i * 2];
pcm32_right[i] = pcm32_stereo[i * 2 + 1];
}
// Step 2: M/S decorrelation
ms_decorrelate(pcm8_left, pcm8_right, pcm8_mid, pcm8_side, num_samples);
ms_decorrelate_16(pcm32_left, pcm32_right, pcm32_mid, pcm32_side, num_samples);
// Step 3: Convert to float and apply DWT
for (size_t i = 0; i < num_samples; i++) {
dwt_mid[i] = (float)pcm8_mid[i];
dwt_side[i] = (float)pcm8_side[i];
dwt_mid[i] = pcm32_mid[i];
dwt_side[i] = pcm32_side[i];
}
dwt_haar_forward_multilevel(dwt_mid, num_samples, dwt_levels);
dwt_haar_forward_multilevel(dwt_side, num_samples, dwt_levels);
dwt_dd4_forward_multilevel(dwt_mid, num_samples, dwt_levels);
dwt_dd4_forward_multilevel(dwt_side, num_samples, dwt_levels);
// Step 4: Quantize with frequency-dependent weights and dead zone
quantize_dwt_coefficients(dwt_mid, quant_mid, num_samples, quality, 1, num_samples, dwt_levels);
quantize_dwt_coefficients(dwt_side, quant_side, num_samples, quality, 1, num_samples, dwt_levels);
// Step 5: Encode with 2-bit significance map
uint8_t *temp_buffer = malloc(num_samples * 4 * sizeof(int16_t));
// Step 5: Encode with 2-bit significance map (32-bit version)
uint8_t *temp_buffer = malloc(num_samples * 4 * sizeof(int32_t));
size_t mid_size = encode_sigmap_2bit(quant_mid, num_samples, temp_buffer);
size_t side_size = encode_sigmap_2bit(quant_side, num_samples, temp_buffer + mid_size);
@@ -429,13 +307,13 @@ size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int qua
size_t zstd_bound = ZSTD_compressBound(uncompressed_size);
uint8_t *zstd_buffer = malloc(zstd_bound);
payload_size = ZSTD_compress(zstd_buffer, zstd_bound, temp_buffer, uncompressed_size, TAD_ZSTD_LEVEL);
payload_size = ZSTD_compress(zstd_buffer, zstd_bound, temp_buffer, uncompressed_size, TAD32_ZSTD_LEVEL);
if (ZSTD_isError(payload_size)) {
fprintf(stderr, "Error: Zstd compression failed: %s\n", ZSTD_getErrorName(payload_size));
free(zstd_buffer);
free(pcm8_stereo); free(pcm8_left); free(pcm8_right);
free(pcm8_mid); free(pcm8_side); free(dwt_mid); free(dwt_side);
free(pcm32_left); free(pcm32_right);
free(pcm32_mid); free(pcm32_side); free(dwt_mid); free(dwt_side);
free(quant_mid); free(quant_side); free(temp_buffer);
return 0;
}
@@ -451,8 +329,8 @@ size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int qua
write_ptr += payload_size;
// Cleanup
free(pcm8_stereo); free(pcm8_left); free(pcm8_right);
free(pcm8_mid); free(pcm8_side); free(dwt_mid); free(dwt_side);
free(pcm32_left); free(pcm32_right);
free(pcm32_mid); free(pcm32_side); free(dwt_mid); free(dwt_side);
free(quant_mid); free(quant_side); free(temp_buffer);
return write_ptr - output;

36
video_encoder/encoder_tad.h
View File

@@ -1,40 +1,40 @@
#ifndef TAD_ENCODER_H
#define TAD_ENCODER_H
#ifndef TAD32_ENCODER_H
#define TAD32_ENCODER_H
#include <stdint.h>
#include <stddef.h>
// TAD (Terrarum Advanced Audio) Encoder
// TAD32 (Terrarum Advanced Audio - PCM32f version) Encoder
// DWT-based perceptual audio codec for TSVM
// Alternative version: PCM32f throughout encoding, PCM8 conversion only at decoder
// Constants
#define TAD_MIN_CHUNK_SIZE 1024 // Minimum: 1024 samples (supports non-power-of-2)
#define TAD_SAMPLE_RATE 32000
#define TAD_CHANNELS 2 // Stereo
#define TAD_SIGMAP_2BIT 1 // 2-bit: 00=0, 01=+1, 10=-1, 11=other
#define TAD_QUALITY_MIN 0
#define TAD_QUALITY_MAX 5
#define TAD_QUALITY_DEFAULT 3
#define TAD_ZSTD_LEVEL 7
#define TAD32_COEFF_SCALAR 1024.0f
#define TAD32_MIN_CHUNK_SIZE 1024 // Minimum: 1024 samples
#define TAD32_SAMPLE_RATE 32000
#define TAD32_CHANNELS 2 // Stereo
#define TAD32_SIGMAP_2BIT 1 // 2-bit: 00=0, 01=+1, 10=-1, 11=other
#define TAD32_QUALITY_MIN 0
#define TAD32_QUALITY_MAX 5
#define TAD32_QUALITY_DEFAULT 3
#define TAD32_ZSTD_LEVEL 7
/**
* Encode audio chunk with TAD codec
* Encode audio chunk with TAD32 codec (PCM32f version)
*
* @param pcm16_stereo Input PCM16LE stereo samples (interleaved L,R)
* @param num_samples Number of samples per channel (supports non-power-of-2, min 1024)
* @param pcm32_stereo Input PCM32fLE stereo samples (interleaved L,R)
* @param num_samples Number of samples per channel (min 1024)
* @param quality Quality level 0-5 (0=lowest, 5=highest)
* @param use_zstd 1=enable Zstd compression, 0=disable
* @param output Output buffer (must be large enough)
* @return Number of bytes written to output, or 0 on error
*
* Output format:
* uint8 sigmap_method (always 1 = 2-bit twobitmap)
* uint8 compressed_flag (1=Zstd, 0=raw)
* uint16 sample_count (samples per channel)
* uint32 payload_size (bytes in payload)
* * payload (encoded M/S data, optionally Zstd-compressed)
*/
size_t tad_encode_chunk(const int16_t *pcm16_stereo, size_t num_samples, int quality,
size_t tad32_encode_chunk(const float *pcm32_stereo, size_t num_samples, int quality,
int use_zstd, uint8_t *output);
#endif // TAD_ENCODER_H
#endif // TAD32_ENCODER_H