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

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};
// 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 dequantization
static const float BASE_QUANTISER_WEIGHTS[] = {
1.0f, // LL (L9) - finest preservation
1.0f, // H (L9)
1.0f, // H (L8)
1.0f, // H (L7)
1.0f, // H (L6)
1.1f, // H (L5)
1.2f, // H (L4)
1.3f, // H (L3)
1.4f, // H (L2)
1.5f // H (L1) - coarsest quantization
};
// These weights are multiplied by quantiser_scale during quantization
static const float BASE_QUANTISER_WEIGHTS[2][10] = {
{ // mid channel
4.0f, // LL (L9) DC
2.0f, // H (L9) 31.25 hz
1.8f, // H (L8) 62.5 hz
1.6f, // H (L7) 125 hz
1.4f, // H (L6) 250 hz
1.2f, // H (L5) 500 hz
1.0f, // H (L4) 1 khz
1.0f, // H (L3) 2 khz
1.3f, // H (L2) 4 khz
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_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;
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
// 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;
}
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
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);
// 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;
}
@@ -509,7 +521,7 @@ static void dequantize_dwt_coefficients(const int8_t *quantized, float *coeffs,
size_t band_len = band_end - band_start;
// 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 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
// Use quantiser_scale = 1.0f for baseline (must match encoder)
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(quant_side, dwt_side, 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(1, quant_side, dwt_side, sample_count, sample_count, dwt_levels, max_index, quantiser_scale);
// Inverse DWT
dwt_inverse_multilevel(dwt_mid, sample_count, dwt_levels);