curioustorvald/tsvm
1
0
Fork 0
mirror of https://github.com/curioustorvald/tsvm.git synced 2026-06-14 00:14:05 +09:00
Code Issues Packages Projects Releases Wiki Activity

more psychovisual model

Browse Source
This commit is contained in:
minjaesong
2025-09-22 00:03:45 +09:00
parent 613b8a97dd
commit 3584520ff9
2 changed files with 54 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
tsvm_core/src/net/torvald/tsvm/GraphicsJSR223Delegate.kt
View File

@@ -4091,6 +4091,7 @@ class GraphicsJSR223Delegate(private val vm: VM) {
var ANISOTROPY_MULT = floatArrayOf(1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 1.0f) var ANISOTROPY_MULT = floatArrayOf(1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 1.0f)
var ANISOTROPY_BIAS = floatArrayOf(0.2f, 0.1f, 0.0f, 0.0f, 0.0f, 0.0f) var ANISOTROPY_BIAS = floatArrayOf(0.2f, 0.1f, 0.0f, 0.0f, 0.0f, 0.0f)
var ANISOTROPY_BIAS_CHROMA = floatArrayOf(0.4f, 0.3f, 0.2f, 0.1f, 0.0f, 0.0f)
@@ -4119,8 +4120,14 @@ class GraphicsJSR223Delegate(private val vm: VM) {
return n / m return n / m
} }
private val FOUR_PIXEL_DETAILER = 0.88f fun perceptual_model3_chroma_basecurve(quality: Int, level: Int): Float {
return 1.0f - (1.0f / (0.5f * quality * quality + 1.0f)) * (level - 4f) // just a line that passes (4,1)
}
private val FOUR_PIXEL_DETAILER = 0.88f
private val TWO_PIXEL_DETAILER = 0.92f
// level is one-based index
private fun getPerceptualWeight(qYGlobal: Int, level: Int, subbandType: Int, isChroma: Boolean, maxLevels: Int): Float { private fun getPerceptualWeight(qYGlobal: Int, level: Int, subbandType: Int, isChroma: Boolean, maxLevels: Int): Float {
// Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity // Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity
@@ -4144,58 +4151,28 @@ class GraphicsJSR223Delegate(private val vm: VM) {
// HL subband - vertical details // HL subband - vertical details
val HL: Float = perceptual_model3_HL(qualityLevel, LH) val HL: Float = perceptual_model3_HL(qualityLevel, LH)
if (subbandType == 2) return HL * (if (level == 3) FOUR_PIXEL_DETAILER else 1f) if (subbandType == 2) return HL * (if (level == 2) TWO_PIXEL_DETAILER else if (level == 3) FOUR_PIXEL_DETAILER else 1f)
// HH subband - diagonal details // HH subband - diagonal details
else return perceptual_model3_HH(LH, HL) * (if (level == 3) FOUR_PIXEL_DETAILER else 1f) else return perceptual_model3_HH(LH, HL) * (if (level == 2) TWO_PIXEL_DETAILER else if (level == 3) FOUR_PIXEL_DETAILER else 1f)
} else { } else {
// CHROMA CHANNELS: Less critical for human perception, more aggressive quantization // CHROMA CHANNELS: Less critical for human perception, more aggressive quantization
when (subbandType) { val base = perceptual_model3_chroma_basecurve(qualityLevel, level)
0 -> { // LL chroma - still important but less than luma
return 1f if (subbandType == 0) { // LL chroma - still important but less than luma
return when { return 1.0f
level >= 6 -> 0.8f // Chroma LL6: Less critical than luma LL }
level >= 5 -> 0.9f else if (subbandType == 1) { // LH chroma - horizontal chroma details
else -> 1.0f return base.coerceAtLeast(1.0f)
} }
} else if (subbandType == 2) { // HL chroma - vertical chroma details (even less critical)
1 -> { // LH chroma - horizontal chroma details return (base * ANISOTROPY_MULT[qualityLevel]).coerceAtLeast(1.0f)
return 1.8f }
return when { else { // HH chroma - diagonal chroma details (most aggressive)
level >= 6 -> 1.0f return (base * ANISOTROPY_MULT[qualityLevel] + ANISOTROPY_BIAS_CHROMA[qualityLevel]).coerceAtLeast(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 1.3f;
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 2.5f
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
}
}
} }
} }
return 1.0f
} }
@@ -4218,7 +4195,9 @@ class GraphicsJSR223Delegate(private val vm: VM) {
subbands: List<DWTSubbandInfo>, baseQuantizer: Float, isChroma: Boolean, decompLevels: Int) { subbands: List<DWTSubbandInfo>, baseQuantizer: Float, isChroma: Boolean, decompLevels: Int) {
// Initialize output array to zero (critical for detecting missing coefficients) // Initialize output array to zero (critical for detecting missing coefficients)
if (tavDebugFrameTarget >= 0) {
Arrays.fill(dequantised, 0.0f) Arrays.fill(dequantised, 0.0f)
}
// Track coefficient coverage for debugging // Track coefficient coverage for debugging
var totalProcessed = 0 var totalProcessed = 0

83
video_encoder/encoder_tav.c
View File

@@ -151,6 +151,7 @@ static const int QUALITY_CG[] = {240, 180, 120, 60, 30, 5};
// psychovisual tuning parameters // psychovisual tuning parameters
static const float ANISOTROPY_MULT[] = {1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 1.0f}; static const float ANISOTROPY_MULT[] = {1.8f, 1.6f, 1.4f, 1.2f, 1.0f, 1.0f};
static const float ANISOTROPY_BIAS[] = {0.2f, 0.1f, 0.0f, 0.0f, 0.0f, 0.0f}; static const float ANISOTROPY_BIAS[] = {0.2f, 0.1f, 0.0f, 0.0f, 0.0f, 0.0f};
static const float ANISOTROPY_BIAS_CHROMA[] = {0.4f, 0.3f, 0.2f, 0.1f, 0.0f, 0.0f};
// DWT coefficient structure for each subband // DWT coefficient structure for each subband
typedef struct { typedef struct {
@@ -828,6 +829,10 @@ static float perceptual_model3_LL(int quality, int level) {
return n / m; return n / m;
} }
static float perceptual_model3_chroma_basecurve(int quality, int level) {
return 1.0f - (1.0f / (0.5f * quality * quality + 1.0f)) * (level - 4.0f); // just a line that passes (4,1)
}
// Get perceptual weight for specific subband - Data-driven model based on coefficient variance analysis // Get perceptual weight for specific subband - Data-driven model based on coefficient variance analysis
static float get_perceptual_weight_model2(int level, int subband_type, int is_chroma, int max_levels) { static float get_perceptual_weight_model2(int level, int subband_type, int is_chroma, int max_levels) {
// Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity // Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity
@@ -897,10 +902,12 @@ static float get_perceptual_weight_model2(int level, int subband_type, int is_ch
} }
#define FOUR_PIXEL_DETAILER 0.88f #define FOUR_PIXEL_DETAILER 0.88f
#define TWO_PIXEL_DETAILER 0.92f
// level is one-based index
static float get_perceptual_weight(tav_encoder_t *enc, int level, int subband_type, int is_chroma, int max_levels) { static float get_perceptual_weight(tav_encoder_t *enc, int level, int subband_type, int is_chroma, int max_levels) {
// Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity // Psychovisual model based on DWT coefficient statistics and Human Visual System sensitivity
// strategy: JPEG quantisation table + real-world statistics from the encoded videos // strategy: more horizontal detail
if (!is_chroma) { if (!is_chroma) {
// LL subband - contains most image energy, preserve carefully // LL subband - contains most image energy, preserve carefully
if (subband_type == 0) if (subband_type == 0)
@@ -914,42 +921,26 @@ static float get_perceptual_weight(tav_encoder_t *enc, int level, int subband_ty
// HL subband - vertical details // HL subband - vertical details
float HL = perceptual_model3_HL(enc->quality_level, LH); float HL = perceptual_model3_HL(enc->quality_level, LH);
if (subband_type == 2) if (subband_type == 2)
return HL * (level == 3 ? FOUR_PIXEL_DETAILER : 1.0f); return HL * (level == 2 ? TWO_PIXEL_DETAILER : level == 3 ? FOUR_PIXEL_DETAILER : 1.0f);
// HH subband - diagonal details // HH subband - diagonal details
else return perceptual_model3_HH(LH, HL) * (level == 3 ? FOUR_PIXEL_DETAILER : 1.0f); else return perceptual_model3_HH(LH, HL) * (level == 2 ? TWO_PIXEL_DETAILER : level == 3 ? FOUR_PIXEL_DETAILER : 1.0f);
} else { } else {
// CHROMA CHANNELS: Less critical for human perception, more aggressive quantization // CHROMA CHANNELS: Less critical for human perception, more aggressive quantization
// strategy: mimic 4:2:2 chroma subsampling // strategy: more horizontal detail
//// mimic 4:4:0 (you heard that right!) chroma subsampling (4:4:4 for higher q, 4:2:0 for lower q)
//// because our eyes are apparently sensitive to horizontal chroma diff as well?
float base = perceptual_model3_chroma_basecurve(enc->quality_level, level);
if (subband_type == 0) { // LL chroma - still important but less than luma if (subband_type == 0) { // LL chroma - still important but less than luma
return 1.0f; return 1.0f;
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 } else if (subband_type == 1) { // LH chroma - horizontal chroma details
return 1.8f; return FCLAMP(base, 1.0f, 100.0f);
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) } else if (subband_type == 2) { // HL chroma - vertical chroma details (even less critical)
return 1.3f; return FCLAMP(base * ANISOTROPY_MULT[enc->quality_level], 1.0f, 100.0f);
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) } else { // HH chroma - diagonal chroma details (most aggressive)
return 2.5f; return FCLAMP(base * ANISOTROPY_MULT[enc->quality_level] + ANISOTROPY_BIAS_CHROMA[enc->quality_level], 1.0f, 100.0f);
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;
} }
} }
} }
@@ -1009,28 +1000,12 @@ static void quantise_dwt_coefficients_perceptual_per_coeff(tav_encoder_t *enc,
float effective_base_q = base_quantizer; float effective_base_q = base_quantizer;
effective_base_q = FCLAMP(effective_base_q, 1.0f, 255.0f); effective_base_q = FCLAMP(effective_base_q, 1.0f, 255.0f);
// Debug coefficient analysis for (int i = 0; i < size; i++) {
if (frame_count == 1 || frame_count == 120) { // Apply perceptual weight based on coefficient's position in DWT layout
int nonzero = 0; float weight = get_perceptual_weight_for_position(enc, i, width, height, decomp_levels, is_chroma);
for (int i = 0; i < size; i++) { float effective_q = effective_base_q * weight;
// Apply perceptual weight based on coefficient's position in DWT layout float quantised_val = coeffs[i] / effective_q;
float weight = get_perceptual_weight_for_position(enc, i, width, height, decomp_levels, is_chroma); quantised[i] = (int16_t)CLAMP((int)(quantised_val + (quantised_val >= 0 ? 0.5f : -0.5f)), -32768, 32767);
float effective_q = effective_base_q * weight;
float quantised_val = coeffs[i] / effective_q;
quantised[i] = (int16_t)CLAMP((int)(quantised_val + (quantised_val >= 0 ? 0.5f : -0.5f)), -32768, 32767);
if (quantised[i] != 0) nonzero++;
}
printf("DEBUG: Frame 120 - %s channel: %d/%d nonzero coeffs after perceptual per-coeff quantization\n",
is_chroma ? "Chroma" : "Luma", nonzero, size);
} else {
// Normal quantization loop
for (int i = 0; i < size; i++) {
// Apply perceptual weight based on coefficient's position in DWT layout
float weight = get_perceptual_weight_for_position(enc, i, width, height, decomp_levels, is_chroma);
float effective_q = effective_base_q * weight;
float quantised_val = coeffs[i] / effective_q;
quantised[i] = (int16_t)CLAMP((int)(quantised_val + (quantised_val >= 0 ? 0.5f : -0.5f)), -32768, 32767);
}
} }
} }
@@ -1373,7 +1348,7 @@ static size_t compress_and_write_frame(tav_encoder_t *enc, uint8_t packet_type)
}*/ }*/
// Debug: Check Y data before DWT transform // Debug: Check Y data before DWT transform
if (enc->frame_count == 120 && enc->verbose) { /*if (enc->frame_count == 120 && enc->verbose) {
float max_y_before = 0.0f; float max_y_before = 0.0f;
int nonzero_before = 0; int nonzero_before = 0;
int total_pixels = enc->monoblock ? (enc->width * enc->height) : (PADDED_TILE_SIZE_X * PADDED_TILE_SIZE_Y); int total_pixels = enc->monoblock ? (enc->width * enc->height) : (PADDED_TILE_SIZE_X * PADDED_TILE_SIZE_Y);
@@ -1383,7 +1358,7 @@ static size_t compress_and_write_frame(tav_encoder_t *enc, uint8_t packet_type)
if (abs_val > 0.1f) nonzero_before++; if (abs_val > 0.1f) nonzero_before++;
} }
printf("DEBUG: Y data before DWT: max=%.2f, nonzero=%d/%d\n", max_y_before, nonzero_before, total_pixels); printf("DEBUG: Y data before DWT: max=%.2f, nonzero=%d/%d\n", max_y_before, nonzero_before, total_pixels);
} }*/
// Apply DWT transform to each channel // Apply DWT transform to each channel
if (enc->monoblock) { if (enc->monoblock) {
@@ -1399,14 +1374,14 @@ static size_t compress_and_write_frame(tav_encoder_t *enc, uint8_t packet_type)
} }
// Debug: Check Y data after DWT transform for high-frequency content // Debug: Check Y data after DWT transform for high-frequency content
if (enc->frame_count == 120 && enc->verbose) { /*if (enc->frame_count == 120 && enc->verbose) {
printf("DEBUG: Y data after DWT (some high-freq samples): "); printf("DEBUG: Y data after DWT (some high-freq samples): ");
int sample_indices[] = {47034, 47035, 47036, 47037, 47038}; // HH1 start + some samples int sample_indices[] = {47034, 47035, 47036, 47037, 47038}; // HH1 start + some samples
for (int i = 0; i < 5; i++) { for (int i = 0; i < 5; i++) {
printf("%.3f ", tile_y_data[sample_indices[i]]); printf("%.3f ", tile_y_data[sample_indices[i]]);
} }
printf("\n"); printf("\n");
} }*/
// Serialise tile // Serialise tile
size_t tile_size = serialise_tile_data(enc, tile_x, tile_y, size_t tile_size = serialise_tile_data(enc, tile_x, tile_y,