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: psychoacoustically optimised quantisation

Browse Source
This commit is contained in:
minjaesong
2025-11-07 03:40:56 +09:00
parent 00c882aa8d
commit d9d395c62c
4 changed files with 64 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
terranmon.txt
View File

@@ -955,7 +955,7 @@ transmission capability, and region-of-interest coding.
- 7 = Mirrored, Clockwise 270 deg - 7 = Mirrored, Clockwise 270 deg
uint8 File Role uint8 File Role
- 0 = generic - 0 = generic
- 1 = this file is header-only, and UCF payload will be followed (used by seekable movie file) - 1 = this file is header-only, and UCF payload will be followed (used by movie file with chapters)
When header-only file contain video packets, they should be presented as an Intro Movie When header-only file contain video packets, they should be presented as an Intro Movie
before the user-interactable selector (served by the UCF payoad) before the user-interactable selector (served by the UCF payoad)
@@ -1256,7 +1256,7 @@ TAV decoder requires new GraphicsJSR223Delegate functions:
- tavQuantise(): Multi-band quantisation - tavQuantise(): Multi-band quantisation
## Audio Support ## Audio Support
Reuses existing MP2 audio infrastructure from TEV/MOV formats for compatibility. MP2 frames, raw PCMu8, and TAD formats are supported.
## Subtitle Support ## Subtitle Support
Uses same Simple Subtitle Format (SSF) as TEV for text overlay functionality. Uses same Simple Subtitle Format (SSF) as TEV for text overlay functionality.

52
video_encoder/decoder_tad.c
View File

@@ -19,20 +19,32 @@
static const float TAD32_COEFF_SCALARS[] = {64.0f, 45.255f, 32.0f, 22.627f, 16.0f, 11.314f, 8.0f, 5.657f, 4.0f, 2.828f}; static const float TAD32_COEFF_SCALARS[] = {64.0f, 45.255f, 32.0f, 22.627f, 16.0f, 11.314f, 8.0f, 5.657f, 4.0f, 2.828f};
// Base quantiser weight table (10 subbands: LL + 9 H bands) // Base quantiser weight table (10 subbands: LL + 9 H bands)
// Linearly spaced from 1.0 (LL) to 2.0 (H9) // These weights are multiplied by quantiser_scale during quantization
// These weights are multiplied by quantiser_scale during dequantization static const float BASE_QUANTISER_WEIGHTS[2][10] = {
static const float BASE_QUANTISER_WEIGHTS[] = { { // mid channel
1.0f, // LL (L9) - finest preservation 4.0f, // LL (L9) DC
1.0f, // H (L9) 2.0f, // H (L9) 31.25 hz
1.0f, // H (L8) 1.8f, // H (L8) 62.5 hz
1.0f, // H (L7) 1.6f, // H (L7) 125 hz
1.0f, // H (L6) 1.4f, // H (L6) 250 hz
1.1f, // H (L5) 1.2f, // H (L5) 500 hz
1.2f, // H (L4) 1.0f, // H (L4) 1 khz
1.3f, // H (L3) 1.0f, // H (L3) 2 khz
1.4f, // H (L2) 1.3f, // H (L2) 4 khz
1.5f // H (L1) - coarsest quantization 1.8f // H (L1) 8 khz
}; },
{ // side channel
6.0f, // LL (L9) DC
5.0f, // H (L9) 31.25 hz
2.6f, // H (L8) 62.5 hz
2.4f, // H (L7) 125 hz
1.8f, // H (L6) 250 hz
1.3f, // H (L5) 500 hz
1.0f, // H (L4) 1 khz
1.0f, // H (L3) 2 khz
1.6f, // H (L2) 4 khz
3.2f // H (L1) 8 khz
}};
#define TAD_DEFAULT_CHUNK_SIZE 32768 #define TAD_DEFAULT_CHUNK_SIZE 32768
#define TAD_MIN_CHUNK_SIZE 1024 #define TAD_MIN_CHUNK_SIZE 1024
@@ -90,7 +102,7 @@ static void spectral_interpolate_band(float *c, size_t len, float Q, float lower
if (len < 4) return; if (len < 4) return;
uint32_t seed = 0x9E3779B9u ^ (uint32_t)len ^ (uint32_t)(Q * 65536.0f); uint32_t seed = 0x9E3779B9u ^ (uint32_t)len ^ (uint32_t)(Q * 65536.0f);
const float dither_amp = 0.05f * Q; // Very light dither (~-60 dBFS) const float dither_amp = 0.02f * Q; // Very light dither
// Just add ultra-light TPDF dither to reduce quantization grain // Just add ultra-light TPDF dither to reduce quantization grain
// No aggressive hole filling or AR prediction that might create artifacts // No aggressive hole filling or AR prediction that might create artifacts
@@ -468,7 +480,7 @@ static float lambda_decompanding(int8_t quant_val, int max_index) {
return sign * abs_val; return sign * abs_val;
} }
static void dequantize_dwt_coefficients(const int8_t *quantized, float *coeffs, size_t count, int chunk_size, int dwt_levels, int max_index, float quantiser_scale) { static void dequantize_dwt_coefficients(int channel, const int8_t *quantized, float *coeffs, size_t count, int chunk_size, int dwt_levels, int max_index, float quantiser_scale) {
// Calculate sideband boundaries dynamically // Calculate sideband boundaries dynamically
int first_band_size = chunk_size >> dwt_levels; int first_band_size = chunk_size >> dwt_levels;
@@ -494,7 +506,7 @@ static void dequantize_dwt_coefficients(const int8_t *quantized, float *coeffs,
float normalized_val = lambda_decompanding(quantized[i], max_index); float normalized_val = lambda_decompanding(quantized[i], max_index);
// Denormalize using the subband scalar and apply base weight + quantiser scaling // Denormalize using the subband scalar and apply base weight + quantiser scaling
float weight = BASE_QUANTISER_WEIGHTS[sideband] * quantiser_scale; float weight = BASE_QUANTISER_WEIGHTS[channel][sideband] * quantiser_scale;
coeffs[i] = normalized_val * TAD32_COEFF_SCALARS[sideband] * weight; coeffs[i] = normalized_val * TAD32_COEFF_SCALARS[sideband] * weight;
} }
@@ -509,7 +521,7 @@ static void dequantize_dwt_coefficients(const int8_t *quantized, float *coeffs,
size_t band_len = band_end - band_start; size_t band_len = band_end - band_start;
// Calculate quantization step Q for this band // Calculate quantization step Q for this band
float weight = BASE_QUANTISER_WEIGHTS[band] * quantiser_scale; float weight = BASE_QUANTISER_WEIGHTS[channel][band] * quantiser_scale;
float scalar = TAD32_COEFF_SCALARS[band] * weight; float scalar = TAD32_COEFF_SCALARS[band] * weight;
float Q = scalar / max_index; float Q = scalar / max_index;
@@ -585,8 +597,8 @@ static int decode_chunk(const uint8_t *input, size_t input_size, uint8_t *pcmu8_
// Dequantize with quantiser scaling and spectral interpolation // Dequantize with quantiser scaling and spectral interpolation
// Use quantiser_scale = 1.0f for baseline (must match encoder) // Use quantiser_scale = 1.0f for baseline (must match encoder)
float quantiser_scale = 1.0f; float quantiser_scale = 1.0f;
dequantize_dwt_coefficients(quant_mid, dwt_mid, sample_count, sample_count, dwt_levels, max_index, quantiser_scale); dequantize_dwt_coefficients(0, quant_mid, dwt_mid, sample_count, sample_count, dwt_levels, max_index, quantiser_scale);
dequantize_dwt_coefficients(quant_side, dwt_side, sample_count, sample_count, dwt_levels, max_index, quantiser_scale); dequantize_dwt_coefficients(1, quant_side, dwt_side, sample_count, sample_count, dwt_levels, max_index, quantiser_scale);
// Inverse DWT // Inverse DWT
dwt_inverse_multilevel(dwt_mid, sample_count, dwt_levels); dwt_inverse_multilevel(dwt_mid, sample_count, dwt_levels);

51
video_encoder/encoder_tad.c
View File

@@ -19,25 +19,32 @@
static const float TAD32_COEFF_SCALARS[] = {64.0f, 45.255f, 32.0f, 22.627f, 16.0f, 11.314f, 8.0f, 5.657f, 4.0f, 2.828f}; static const float TAD32_COEFF_SCALARS[] = {64.0f, 45.255f, 32.0f, 22.627f, 16.0f, 11.314f, 8.0f, 5.657f, 4.0f, 2.828f};
// Base quantiser weight table (10 subbands: LL + 9 H bands) // Base quantiser weight table (10 subbands: LL + 9 H bands)
// Linearly spaced from 1.0 (LL) to 2.0 (H9)
// These weights are multiplied by quantiser_scale during quantization // These weights are multiplied by quantiser_scale during quantization
static const float BASE_QUANTISER_WEIGHTS[] = { static const float BASE_QUANTISER_WEIGHTS[2][10] = {
1.0f, // LL (L9) - finest preservation { // mid channel
1.0f, // H (L9) 4.0f, // LL (L9) DC
1.0f, // H (L8) 2.0f, // H (L9) 31.25 hz
1.0f, // H (L7) 1.8f, // H (L8) 62.5 hz
1.0f, // H (L6) 1.6f, // H (L7) 125 hz
1.1f, // H (L5) 1.4f, // H (L6) 250 hz
1.2f, // H (L4) 1.2f, // H (L5) 500 hz
1.3f, // H (L3) 1.0f, // H (L4) 1 khz
1.4f, // H (L2) 1.0f, // H (L3) 2 khz
1.5f // H (L1) - coarsest quantization 1.3f, // H (L2) 4 khz
}; 1.8f // H (L1) 8 khz
},
// Forward declarations for internal functions { // side channel
static void dwt_dd4_forward_1d(float *data, int length); 6.0f, // LL (L9) DC
static void dwt_forward_multilevel(float *data, int length, int levels); 5.0f, // H (L9) 31.25 hz
static void quantize_dwt_coefficients(const float *coeffs, int8_t *quantized, size_t count, int apply_deadzone, int chunk_size, int dwt_levels, int quant_bits, int *current_subband_index, float quantiser_scale); 2.6f, // H (L8) 62.5 hz
2.4f, // H (L7) 125 hz
1.8f, // H (L6) 250 hz
1.3f, // H (L5) 500 hz
1.0f, // H (L4) 1 khz
1.0f, // H (L3) 2 khz
1.6f, // H (L2) 4 khz
3.2f // H (L1) 8 khz
}};
static inline float FCLAMP(float x, float min, float max) { static inline float FCLAMP(float x, float min, float max) {
return x < min ? min : (x > max ? max : x); return x < min ? min : (x > max ? max : x);
@@ -279,7 +286,7 @@ static int8_t lambda_companding(float val, int max_index) {
return (int8_t)(sign * index); return (int8_t)(sign * index);
} }
static void quantize_dwt_coefficients(const float *coeffs, int8_t *quantized, size_t count, int apply_deadzone, int chunk_size, int dwt_levels, int max_index, int *current_subband_index, float quantiser_scale) { static void quantize_dwt_coefficients(int channel, const float *coeffs, int8_t *quantized, size_t count, int apply_deadzone, int chunk_size, int dwt_levels, int max_index, int *current_subband_index, float quantiser_scale) {
int first_band_size = chunk_size >> dwt_levels; int first_band_size = chunk_size >> dwt_levels;
int *sideband_starts = malloc((dwt_levels + 2) * sizeof(int)); int *sideband_starts = malloc((dwt_levels + 2) * sizeof(int));
@@ -304,7 +311,7 @@ static void quantize_dwt_coefficients(const float *coeffs, int8_t *quantized, si
} }
// Apply base weight and quantiser scaling // Apply base weight and quantiser scaling
float weight = BASE_QUANTISER_WEIGHTS[sideband] * quantiser_scale; float weight = BASE_QUANTISER_WEIGHTS[channel][sideband] * quantiser_scale;
float val = (coeffs[i] / (TAD32_COEFF_SCALARS[sideband] * weight)); // val is normalised to [-1,1] float val = (coeffs[i] / (TAD32_COEFF_SCALARS[sideband] * weight)); // val is normalised to [-1,1]
int8_t quant_val = lambda_companding(val, max_index); int8_t quant_val = lambda_companding(val, max_index);
@@ -779,8 +786,8 @@ size_t tad32_encode_chunk(const float *pcm32_stereo, size_t num_samples,
} }
// Step 4: Quantize with frequency-dependent weights and quantiser scaling // Step 4: Quantize with frequency-dependent weights and quantiser scaling
quantize_dwt_coefficients(dwt_mid, quant_mid, num_samples, 1, num_samples, dwt_levels, max_index, NULL, quantiser_scale); quantize_dwt_coefficients(0, dwt_mid, quant_mid, num_samples, 1, num_samples, dwt_levels, max_index, NULL, quantiser_scale);
quantize_dwt_coefficients(dwt_side, quant_side, num_samples, 1, num_samples, dwt_levels, max_index, NULL, quantiser_scale); quantize_dwt_coefficients(1, dwt_side, quant_side, num_samples, 1, num_samples, dwt_levels, max_index, NULL, quantiser_scale);
// Step 4.5: Accumulate quantized coefficient statistics if enabled // Step 4.5: Accumulate quantized coefficient statistics if enabled
if (stats_enabled) { if (stats_enabled) {

5
video_encoder/encoder_tad_standalone.c
View File

@@ -12,7 +12,7 @@
#include <time.h> #include <time.h>
#include "encoder_tad.h" #include "encoder_tad.h"
#define ENCODER_VENDOR_STRING "Encoder-TAD32 (PCM32f version) 20251026" #define ENCODER_VENDOR_STRING "Encoder-TAD32 (PCM32f version) 20251107"
// TAD32 format constants // TAD32 format constants
#define TAD32_DEFAULT_CHUNK_SIZE 32768 // Default: power of 2 for optimal performance (2^15) #define TAD32_DEFAULT_CHUNK_SIZE 32768 // Default: power of 2 for optimal performance (2^15)
@@ -55,9 +55,6 @@ static void print_usage(const char *prog_name) {
printf(" 3 = good quality (max_index=47) [DEFAULT]\n"); printf(" 3 = good quality (max_index=47) [DEFAULT]\n");
printf(" 4 = high quality (max_index=56)\n"); printf(" 4 = high quality (max_index=56)\n");
printf(" 5 = very high quality/largest (max_index=89)\n"); printf(" 5 = very high quality/largest (max_index=89)\n");
printf(" -s <scale> Quantiser scaling factor (default: 1.0, range: 0.5-4.0)\n");
printf(" Higher = more aggressive quantization, smaller files\n");
printf(" 2.0 = quantize 2x coarser than baseline\n");
printf(" -v Verbose output\n"); printf(" -v Verbose output\n");
printf(" -h, --help Show this help\n"); printf(" -h, --help Show this help\n");
printf("\nVersion: %s\n", ENCODER_VENDOR_STRING); printf("\nVersion: %s\n", ENCODER_VENDOR_STRING);