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

TAV: even more preset changes

Browse Source
This commit is contained in:
minjaesong
2025-10-07 02:27:15 +09:00
parent 21e3fe4c1e
commit cdec0fe020
3 changed files with 120 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
tsvm_core/src/net/torvald/tsvm/GraphicsJSR223Delegate.kt
View File

@@ -4235,14 +4235,14 @@ class GraphicsJSR223Delegate(private val vm: VM) {
return subbands
}
var ANISOTROPY_MULT = floatArrayOf(2.0f, 1.8f, 1.6f, 1.4f, 1.2f, 1.0f)
var ANISOTROPY_BIAS = floatArrayOf(0.4f, 0.2f, 0.1f, 0.0f, 0.0f, 0.0f)
var ANISOTROPY_MULT_CHROMA = floatArrayOf(6.6f, 5.5f, 4.4f, 3.3f, 2.2f, 1.1f)
var ANISOTROPY_BIAS_CHROMA = floatArrayOf(1.0f, 0.8f, 0.6f, 0.4f, 0.2f, 0.0f)
var ANISOTROPY_MULT = floatArrayOf(2.0f, 1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 1.0f)
var ANISOTROPY_BIAS = floatArrayOf(0.4f, 0.2f, 0.1f, 0.0f, 0.0f, 0.0f, 0.0f)
var ANISOTROPY_MULT_CHROMA = floatArrayOf(6.6f, 5.5f, 4.4f, 3.3f, 2.2f, 1.1f, 1.0f)
var ANISOTROPY_BIAS_CHROMA = floatArrayOf(1.0f, 0.8f, 0.6f, 0.4f, 0.2f, 0.0f, 0.0f)
private fun perceptual_model3_LH(quality: Int, level: Float): Float {
private fun perceptual_model3_LH(level: Float): Float {
val H4 = 1.2f
val Q = 2f // using fixed value for fixed curve; quantiser will scale it up anyway
val Q12 = Q * 12f
@@ -4268,9 +4268,9 @@ class GraphicsJSR223Delegate(private val vm: VM) {
return lerp(LH, HL, Kx)
}
fun perceptual_model3_LL(quality: Int, level: Float): Float {
val n = perceptual_model3_LH(quality, level)
val m = perceptual_model3_LH(quality, level - 1) / n
fun perceptual_model3_LL(level: Float): Float {
val n = perceptual_model3_LH(level)
val m = perceptual_model3_LH(level - 1) / n
return n / m
}
@@ -4284,13 +4284,13 @@ class GraphicsJSR223Delegate(private val vm: VM) {
private fun tavDeriveEncoderQindex(qIndex: Int, qYGlobal: Int): Int {
if (qIndex > 0) return qIndex - 1
return if (qYGlobal >= 60) 0
else if (qYGlobal >= 42) 1
else if (qYGlobal >= 25) 2
else if (qYGlobal >= 12) 3
else if (qYGlobal >= 6) 4
return if (qYGlobal >= 79) 0
else if (qYGlobal >= 47) 1
else if (qYGlobal >= 23) 2
else if (qYGlobal >= 11) 3
else if (qYGlobal >= 5) 4
else if (qYGlobal >= 2) 5
else 5
else 6
}
// level is one-based index
@@ -4306,10 +4306,10 @@ class GraphicsJSR223Delegate(private val vm: VM) {
// LUMA CHANNEL: Based on statistical analysis from real video content
// LL subband - contains most image energy, preserve carefully
if (subbandType == 0) return perceptual_model3_LL(qualityLevel, level)
if (subbandType == 0) return perceptual_model3_LL(level)
// LH subband - horizontal details (human eyes more sensitive)
val LH: Float = perceptual_model3_LH(qualityLevel, level)
val LH: Float = perceptual_model3_LH(level)
if (subbandType == 1) return LH
// HL subband - vertical details

6
video_encoder/encoder_tav.c
View File

@@ -181,14 +181,14 @@ static const int QLUT[] = {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21
// Quality level to quantisation mapping for different channels
// the values are indices to the QLUT
static const int QUALITY_Y[] = {79, 47, 23, 11, 5, 2, 0}; // 96, 48, 24, 12, 6, 3, 1
static const int QUALITY_Y[] = {79, 47, 23, 11, 5, 2, 1}; // 96, 48, 24, 12, 6, 3, 2
static const int QUALITY_CO[] = {123, 108, 91, 76, 59, 29, 4}; // 240, 180, 120, 90, 60, 30, 5
static const int QUALITY_CG[] = {148, 133, 113, 99, 76, 39, 7}; // 424, 304, 200, 144, 90, 40, 8
static const int QUALITY_ALPHA[] = {79, 47, 23, 11, 5, 2, 0}; // 96, 48, 24, 12, 6, 3, 1
static const int QUALITY_ALPHA[] = {79, 47, 23, 11, 5, 2, 1}; // 96, 48, 24, 12, 6, 3, 2
// Dead-zone quantisation thresholds per quality level
// Higher values = more aggressive (more coefficients set to zero)
static const float DEAD_ZONE_THRESHOLD[] = {2.0f, 1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 0.0f};
static const float DEAD_ZONE_THRESHOLD[] = {1.5f, 1.5f, 1.2f, 1.0f, 0.8f, 0.6f, 0.0f};
// Dead-zone scaling factors for different subband levels
#define DEAD_ZONE_FINEST_SCALE 1.0f // Full dead-zone for finest level (level 6)

104
video_encoder/visualise_coefficients.c
View File

@@ -104,10 +104,11 @@ int main(int argc, char *argv[]) {
return 1;
}
// Analyse coefficient distribution
// Analyse coefficient distribution - Overall and per-subband
size_t zeros = 0, positives = 0, negatives = 0;
int16_t min_val = INT16_MAX, max_val = INT16_MIN;
// Calculate overall statistics
for (size_t i = 0; i < expected_count; i++) {
if (coeffs[i] == 0) zeros++;
else if (coeffs[i] > 0) positives++;
@@ -117,12 +118,109 @@ int main(int argc, char *argv[]) {
if (coeffs[i] > max_val) max_val = coeffs[i];
}
printf("Coefficient statistics:\n");
printf("Overall coefficient statistics:\n");
printf(" Total: %zu\n", expected_count);
printf(" Zeros: %zu (%.1f%%)\n", zeros, 100.0 * zeros / expected_count);
printf(" Positives: %zu (%.1f%%)\n", positives, 100.0 * positives / expected_count);
printf(" Negatives: %zu (%.1f%%)\n", negatives, 100.0 * negatives / expected_count);
printf(" Range: [%d, %d]\n", min_val, max_val);
printf(" Range: [%d, %d]\n\n", min_val, max_val);
// Per-subband statistics
// Linear layout: [LL1, LH1, HL1, HH1, LH2, HL2, HH2, ..., LH6, HL6, HH6]
size_t offset = 0;
// Determine number of DWT levels (assuming standard 6-level for 560x448)
int num_levels = 6;
int w = width, h = height;
// LL subband (deepest level, smallest)
int ll_divisor = 1 << num_levels; // 2^6 = 64
int ll_w = w / ll_divisor;
int ll_h = h / ll_divisor;
size_t ll_size = ll_w * ll_h;
if (offset + ll_size <= expected_count) {
size_t ll_zeros = 0, ll_pos = 0, ll_neg = 0;
int16_t ll_min = INT16_MAX, ll_max = INT16_MIN;
for (size_t i = 0; i < ll_size; i++) {
int16_t val = coeffs[offset + i];
if (val == 0) ll_zeros++;
else if (val > 0) ll_pos++;
else ll_neg++;
if (val < ll_min) ll_min = val;
if (val > ll_max) ll_max = val;
}
printf("LL%d subband:\n", num_levels);
printf(" Total: %zu\n", ll_size);
printf(" Zeros: %zu (%.1f%%)\n", ll_zeros, 100.0 * ll_zeros / ll_size);
printf(" Positives: %zu (%.1f%%)\n", ll_pos, 100.0 * ll_pos / ll_size);
printf(" Negatives: %zu (%.1f%%)\n", ll_neg, 100.0 * ll_neg / ll_size);
printf(" Range: [%d, %d]\n\n", ll_min, ll_max);
offset += ll_size;
}
// LH, HL, HH subbands for each level (from deepest to finest)
for (int level = num_levels; level >= 1; level--) {
int divisor = 1 << level; // 2^level
int sub_w = w / divisor;
int sub_h = h / divisor;
size_t sub_size = sub_w * sub_h;
if (offset + 3 * sub_size > expected_count) break;
// LH subband
size_t lh_zeros = 0, lh_pos = 0, lh_neg = 0;
int16_t lh_min = INT16_MAX, lh_max = INT16_MIN;
for (size_t i = 0; i < sub_size; i++) {
int16_t val = coeffs[offset + i];
if (val == 0) lh_zeros++;
else if (val > 0) lh_pos++;
else lh_neg++;
if (val < lh_min) lh_min = val;
if (val > lh_max) lh_max = val;
}
offset += sub_size;
// HL subband
size_t hl_zeros = 0, hl_pos = 0, hl_neg = 0;
int16_t hl_min = INT16_MAX, hl_max = INT16_MIN;
for (size_t i = 0; i < sub_size; i++) {
int16_t val = coeffs[offset + i];
if (val == 0) hl_zeros++;
else if (val > 0) hl_pos++;
else hl_neg++;
if (val < hl_min) hl_min = val;
if (val > hl_max) hl_max = val;
}
offset += sub_size;
// HH subband
size_t hh_zeros = 0, hh_pos = 0, hh_neg = 0;
int16_t hh_min = INT16_MAX, hh_max = INT16_MIN;
for (size_t i = 0; i < sub_size; i++) {
int16_t val = coeffs[offset + i];
if (val == 0) hh_zeros++;
else if (val > 0) hh_pos++;
else hh_neg++;
if (val < hh_min) hh_min = val;
if (val > hh_max) hh_max = val;
}
offset += sub_size;
printf("Level %d subbands (%dx%d each):\n", level, sub_w, sub_h);
printf(" LH%d: Total=%zu, Zeros=%zu (%.1f%%), Pos=%zu (%.1f%%), Neg=%zu (%.1f%%), Range=[%d,%d]\n",
level, sub_size, lh_zeros, 100.0*lh_zeros/sub_size,
lh_pos, 100.0*lh_pos/sub_size, lh_neg, 100.0*lh_neg/sub_size, lh_min, lh_max);
printf(" HL%d: Total=%zu, Zeros=%zu (%.1f%%), Pos=%zu (%.1f%%), Neg=%zu (%.1f%%), Range=[%d,%d]\n",
level, sub_size, hl_zeros, 100.0*hl_zeros/sub_size,
hl_pos, 100.0*hl_pos/sub_size, hl_neg, 100.0*hl_neg/sub_size, hl_min, hl_max);
printf(" HH%d: Total=%zu, Zeros=%zu (%.1f%%), Pos=%zu (%.1f%%), Neg=%zu (%.1f%%), Range=[%d,%d]\n\n",
level, sub_size, hh_zeros, 100.0*hh_zeros/sub_size,
hh_pos, 100.0*hh_pos/sub_size, hh_neg, 100.0*hh_neg/sub_size, hh_min, hh_max);
}
// Write PPM image
FILE *fp_out = fopen(output_file, "wb");