curioustorvald/tsvm
1
0
Fork 0
mirror of https://github.com/curioustorvald/tsvm.git synced 2026-03-07 11:51:49 +09:00
Code Issues Packages Projects Releases Wiki Activity

TAV: first working psychovisual tuning

Browse Source
This commit is contained in:
minjaesong
2025-09-20 11:15:04 +09:00
parent d3a18c081a
commit 206e43a308
2 changed files with 186 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

95
tsvm_core/src/net/torvald/tsvm/GraphicsJSR223Delegate.kt
View File

@@ -3889,10 +3889,97 @@ class GraphicsJSR223Delegate(private val vm: VM) {
}
private fun getPerceptualWeight(level: Int, subbandType: Int, isChroma: Boolean, maxLevels: Int): Float {
return 1f
// Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity
// Data-driven model based on coefficient variance analysis - MUST match encoder exactly
if (!isChroma) {
// LUMA CHANNEL: Based on statistical analysis from real video content
when (subbandType) {
0 -> { // LL subband - contains most image energy, preserve carefully
return when {
level >= 6 -> 0.6f // LL6: High energy but can tolerate moderate quantization (range up to 22K)
level >= 5 -> 0.7f // LL5: Good preservation
else -> 0.8f // Lower LL levels: Fine preservation
}
}
1 -> { // LH subband - horizontal details (human eyes more sensitive)
return when {
level >= 6 -> 0.7f // LH6: Significant coefficients (max ~500), preserve well
level >= 5 -> 0.8f // LH5: Moderate coefficients (max ~600)
level >= 4 -> 1.0f // LH4: Small coefficients (max ~50)
level >= 3 -> 1.2f // LH3: Very small coefficients, can quantize more
level >= 2 -> 1.4f // LH2: Minimal impact
else -> 1.6f // LH1: Least important
}
}
2 -> { // HL subband - vertical details (less sensitive due to HVS characteristics)
return when {
level >= 6 -> 0.9f // HL6: Can quantize more aggressively than LH6
level >= 5 -> 1.0f // HL5: Standard quantization
level >= 4 -> 1.3f // HL4: Notable range but less critical
level >= 3 -> 1.5f // HL3: Can tolerate more quantization
level >= 2 -> 1.7f // HL2: Less important
else -> 2.0f // HL1: Most aggressive for vertical details
}
}
3 -> { // HH subband - diagonal details (least important for HVS)
return when {
level >= 6 -> 1.1f // HH6: Preserve some diagonal detail
level >= 5 -> 1.3f // HH5: Can quantize aggressively
level >= 4 -> 1.6f // HH4: Very aggressive
level >= 3 -> 2.0f // HH3: Minimal preservation
level >= 2 -> 2.2f // HH2: Maximum compression
else -> 2.5f // HH1: Most aggressive quantization
}
}
else -> 1.0f
}
} else {
// CHROMA CHANNELS: Less critical for human perception, more aggressive quantization
when (subbandType) {
0 -> { // LL chroma - still important but less than luma
return when {
level >= 6 -> 0.8f // Chroma LL6: Less critical than luma LL
level >= 5 -> 0.9f
else -> 1.0f
}
}
1 -> { // LH chroma - horizontal chroma details
return when {
level >= 6 -> 1.0f
level >= 5 -> 1.2f
level >= 4 -> 1.4f
level >= 3 -> 1.6f
level >= 2 -> 1.8f
else -> 2.0f
}
}
2 -> { // HL chroma - vertical chroma details (even less critical)
return when {
level >= 6 -> 1.2f
level >= 5 -> 1.4f
level >= 4 -> 1.6f
level >= 3 -> 1.8f
level >= 2 -> 2.0f
else -> 2.2f
}
}
3 -> { // HH chroma - diagonal chroma details (most aggressive)
return when {
level >= 6 -> 1.4f
level >= 5 -> 1.6f
level >= 4 -> 1.8f
level >= 3 -> 2.1f
level >= 2 -> 2.3f
else -> 2.5f
}
}
else -> 1.0f
}
}
return 1.0f
// Legacy data-driven model (kept for reference but not used)
/*if (!isChroma) {
// Luma strategy based on statistical variance analysis from real video data
return when (subbandType) {
0 -> { // LL
@@ -3939,7 +4026,7 @@ class GraphicsJSR223Delegate(private val vm: VM) {
// Chroma strategy - apply 0.85x reduction to luma weights for color preservation
val lumaWeight = getPerceptualWeight(level, subbandType, false, maxLevels)
return lumaWeight * 1.6f
}
}*/
}
// Helper function to calculate five-number summary for coefficient analysis
@@ -4027,7 +4114,7 @@ class GraphicsJSR223Delegate(private val vm: VM) {
}
}
private val tavDebugFrameTarget = 0 // use negative number to disable the debug print
private val tavDebugFrameTarget = -1 // use negative number to disable the debug print
private var tavDebugCurrentFrameNumber = 0
fun tavDecode(blockDataPtr: Long, currentRGBAddr: Long, prevRGBAddr: Long,

138
video_encoder/encoder_tav.c
View File

@@ -799,51 +799,108 @@ static void quantise_dwt_coefficients(float *coeffs, int16_t *quantised, int siz
// Get perceptual weight for specific subband - Data-driven model based on coefficient variance analysis
static float get_perceptual_weight(int level, int subband_type, int is_chroma, int max_levels) {
// TEMPORARY: Test with uniform weights to verify linear layout works correctly
return 1.0f;
// Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity
if (!is_chroma) {
// Luma strategy based on statistical variance analysis from real video data
if (subband_type == 0) { // LL
// LL6 has extremely high variance (Range=8026.7) but contains most image energy
// Moderate quantization appropriate due to high variance tolerance
return 1.1f;
} else if (subband_type == 1) { // LH (horizontal detail)
// Data-driven weights based on observed coefficient patterns
if (level >= 6) return 0.7f; // LH6: significant coefficients (Range=243.1)
else if (level == 5) return 0.8f; // LH5: moderate coefficients (Range=264.3)
else if (level == 4) return 1.0f; // LH4: small coefficients (Range=50.8)
else if (level == 3) return 1.4f; // LH3: sparse but large outliers (Range=11909.1)
else if (level == 2) return 1.6f; // LH2: fewer coefficients (Range=6720.2)
else return 1.9f; // LH1: smallest detail (Range=1606.3)
} else if (subband_type == 2) { // HL (vertical detail)
// Similar pattern to LH but slightly different variance
if (level >= 6) return 0.8f; // HL6: moderate coefficients (Range=181.6)
else if (level == 5) return 0.9f; // HL5: small coefficients (Range=80.4)
else if (level == 4) return 1.2f; // HL4: surprising large outliers (Range=9737.9)
else if (level == 3) return 1.3f; // HL3: very large outliers (Range=13698.2)
else if (level == 2) return 1.5f; // HL2: moderate range (Range=2099.4)
else return 1.8f; // HL1: small coefficients (Range=851.1)
} else { // HH (diagonal detail)
// HH bands generally have lower energy but important for texture
if (level >= 6) return 1.0f; // HH6: some significant coefficients (Range=95.8)
else if (level == 5) return 1.1f; // HH5: small coefficients (Range=75.9)
else if (level == 4) return 1.3f; // HH4: moderate range (Range=89.8)
else if (level == 3) return 1.5f; // HH3: large outliers (Range=11611.2)
else if (level == 2) return 1.8f; // HH2: moderate range (Range=2499.2)
else return 2.1f; // HH1: smallest coefficients (Range=761.6)
// LUMA CHANNEL: Based on statistical analysis from real video content
if (subband_type == 0) { // LL subband - contains most image energy, preserve carefully
if (level >= 6) return 0.6f; // LL6: High energy but can tolerate moderate quantization (range up to 22K)
if (level >= 5) return 0.7f; // LL5: Good preservation
return 0.8f; // Lower LL levels: Fine preservation
} else if (subband_type == 1) { // LH subband - horizontal details (human eyes more sensitive)
if (level >= 6) return 0.7f; // LH6: Significant coefficients (max ~500), preserve well
if (level >= 5) return 0.8f; // LH5: Moderate coefficients (max ~600)
if (level >= 4) return 1.0f; // LH4: Small coefficients (max ~50)
if (level >= 3) return 1.2f; // LH3: Very small coefficients, can quantize more
if (level >= 2) return 1.4f; // LH2: Minimal impact
return 1.6f; // LH1: Least important
} else if (subband_type == 2) { // HL subband - vertical details (less sensitive due to HVS characteristics)
if (level >= 6) return 0.9f; // HL6: Can quantize more aggressively than LH6
if (level >= 5) return 1.0f; // HL5: Standard quantization
if (level >= 4) return 1.3f; // HL4: Notable range but less critical
if (level >= 3) return 1.5f; // HL3: Can tolerate more quantization
if (level >= 2) return 1.7f; // HL2: Less important
return 2.0f; // HL1: Most aggressive for vertical details
} else { // HH subband - diagonal details (least important for HVS)
if (level >= 6) return 1.1f; // HH6: Preserve some diagonal detail
if (level >= 5) return 1.3f; // HH5: Can quantize aggressively
if (level >= 4) return 1.6f; // HH4: Very aggressive
if (level >= 3) return 2.0f; // HH3: Minimal preservation
if (level >= 2) return 2.2f; // HH2: Maximum compression
return 2.5f; // HH1: Most aggressive quantization
}
} else {
// Chroma strategy - apply 0.85x reduction to luma weights for color preservation
float luma_weight = get_perceptual_weight(level, subband_type, 0, max_levels);
return luma_weight * 0.85f;
// CHROMA CHANNELS: Less critical for human perception, more aggressive quantization
if (subband_type == 0) { // LL chroma - still important but less than luma
if (level >= 6) return 0.8f; // Chroma LL6: Less critical than luma LL
if (level >= 5) return 0.9f;
return 1.0f;
} else if (subband_type == 1) { // LH chroma - horizontal chroma details
if (level >= 6) return 1.0f;
if (level >= 5) return 1.2f;
if (level >= 4) return 1.4f;
if (level >= 3) return 1.6f;
if (level >= 2) return 1.8f;
return 2.0f;
} else if (subband_type == 2) { // HL chroma - vertical chroma details (even less critical)
if (level >= 6) return 1.2f;
if (level >= 5) return 1.4f;
if (level >= 4) return 1.6f;
if (level >= 3) return 1.8f;
if (level >= 2) return 2.0f;
return 2.2f;
} else { // HH chroma - diagonal chroma details (most aggressive)
if (level >= 6) return 1.4f;
if (level >= 5) return 1.6f;
if (level >= 4) return 1.8f;
if (level >= 3) return 2.1f;
if (level >= 2) return 2.3f;
return 2.5f;
}
}
}
// Determine perceptual weight for coefficient at linear position (matches actual DWT layout)
static float get_perceptual_weight_for_position(int linear_idx, int width, int height, int decomp_levels, int is_chroma) {
// For now, return uniform weight while we figure out the actual DWT layout
// TODO: Map linear_idx to correct DWT subband and return appropriate weight
// Map linear coefficient index to DWT subband using same layout as decoder
int offset = 0;
// First: LL subband at maximum decomposition level
int ll_width = width >> decomp_levels;
int ll_height = height >> decomp_levels;
int ll_size = ll_width * ll_height;
if (linear_idx < offset + ll_size) {
// LL subband at maximum level - use get_perceptual_weight for consistency
return get_perceptual_weight(decomp_levels, 0, is_chroma, decomp_levels);
}
offset += ll_size;
// Then: LH, HL, HH subbands for each level from max down to 1
for (int level = decomp_levels; level >= 1; level--) {
int level_width = width >> (decomp_levels - level + 1);
int level_height = height >> (decomp_levels - level + 1);
int subband_size = level_width * level_height;
// LH subband (horizontal details)
if (linear_idx < offset + subband_size) {
return get_perceptual_weight(level, 1, is_chroma, decomp_levels);
}
offset += subband_size;
// HL subband (vertical details)
if (linear_idx < offset + subband_size) {
return get_perceptual_weight(level, 2, is_chroma, decomp_levels);
}
offset += subband_size;
// HH subband (diagonal details)
if (linear_idx < offset + subband_size) {
return get_perceptual_weight(level, 3, is_chroma, decomp_levels);
}
offset += subband_size;
}
// Fallback for out-of-bounds indices
return 1.0f;
}
@@ -2668,12 +2725,7 @@ int main(int argc, char *argv[]) {
printf("Base quantiser: Y=%d, Co=%d, Cg=%d\n", enc->quantiser_y, enc->quantiser_co, enc->quantiser_cg);
}
if (enc->perceptual_tuning) {
printf("Perceptual weights: LL=%.1fx, LH/HL=%.1f-%.1fx, HH=%.1f-%.1fx (varies by level)\n",
get_perceptual_weight(enc->decomp_levels, 0, 0, enc->decomp_levels),
get_perceptual_weight(enc->decomp_levels, 1, 0, enc->decomp_levels),
get_perceptual_weight(1, 1, 0, enc->decomp_levels),
get_perceptual_weight(enc->decomp_levels, 3, 0, enc->decomp_levels),
get_perceptual_weight(1, 3, 0, enc->decomp_levels));
printf("Perceptual tuning enabled\n");
}
// Open output file