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: preset implementation

Browse Source
This commit is contained in:
minjaesong
2025-11-24 17:40:45 +09:00
parent 6132012e74
commit 08bb33bf27
6 changed files with 152 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
video_encoder/decoder_tav.c
View File

@@ -17,7 +17,7 @@
#include "decoder_tad.h" // Shared TAD decoder library
#include "tav_avx512.h" // AVX-512 SIMD optimisations
#define DECODER_VENDOR_STRING "Decoder-TAV 20251124 (avx512)"
#define DECODER_VENDOR_STRING "Decoder-TAV 20251124 (avx512,presets)"
// TAV format constants
#define TAV_MAGIC "\x1F\x54\x53\x56\x4D\x54\x41\x56"
@@ -95,7 +95,8 @@ typedef struct {
uint8_t encoder_quality;
uint8_t channel_layout;
uint8_t entropy_coder;
uint8_t reserved[2];
uint8_t encoder_preset; // Byte 28: bit 0 = sports, bit 1 = anime
uint8_t reserved;
uint8_t device_orientation;
uint8_t file_role;
} __attribute__((packed)) tav_header_t;
@@ -394,10 +395,20 @@ static inline float tav_grain_triangular_noise(uint32_t rng_val) {
return (u1 + u2) - 1.0f;
}
// Remove grain synthesis from DWT coefficients (decoder subtracts noise)
// Apply grain synthesis from DWT coefficients (decoder subtracts noise)
// This must be called AFTER dequantisation but BEFORE inverse DWT
static void remove_grain_synthesis_decoder(float *coeffs, int width, int height,
int decomp_levels, int frame_num, int q_y_global) {
static void apply_grain_synthesis(float *coeffs, int width, int height,
int decomp_levels, int frame_num, int q_y_global, uint8_t encoder_preset, int no_grain_synthesis) {
// Command-line override: disable grain synthesis
if (no_grain_synthesis) {
return; // Skip grain synthesis entirely
}
// Anime preset: completely disable grain synthesis
if (encoder_preset & 0x02) {
return; // Skip grain synthesis entirely
}
dwt_subband_info_t subbands[32];
const int subband_count = calculate_subband_layout(width, height, decomp_levels, subbands);
@@ -412,7 +423,7 @@ static void remove_grain_synthesis_decoder(float *coeffs, int width, int height,
// Calculate band index for RNG (matches Kotlin: level + subbandType * 31 + 16777619)
uint32_t band = subband->level + subband->subband_type * 31 + 16777619;
// Remove noise from each coefficient in this subband
// Apply noise from each coefficient in this subband
for (int i = 0; i < subband->coeff_count; i++) {
const int idx = subband->coeff_start + i;
if (idx < width * height) {
@@ -1226,14 +1237,14 @@ static int get_temporal_subband_level(int frame_idx, int num_frames, int tempora
}
// Calculate temporal quantiser scale for a given temporal subband level
static float get_temporal_quantiser_scale(int temporal_level) {
static float get_temporal_quantiser_scale(uint8_t encoder_preset, int temporal_level) {
// Uses exponential scaling: 2^(BETA × level^KAPPA)
// With BETA=0.6, KAPPA=1.14:
// - Level 0 (tLL): 2^0.0 = 1.00
// - Level 1 (tH): 2^0.68 = 1.61
// - Level 2 (tHH): 2^1.29 = 2.45
const float BETA = 0.6f; // Temporal scaling exponent
const float KAPPA = 1.14f;
const float BETA = (encoder_preset & 0x01) ? 0.0f : 0.6f;
const float KAPPA = (encoder_preset & 0x01) ? 1.0f : 1.14f;
return powf(2.0f, BETA * powf(temporal_level, KAPPA));
}
@@ -1812,6 +1823,7 @@ typedef struct {
int frame_size;
int is_monoblock; // True if version 3-6 (single tile mode)
int temporal_motion_coder; // Temporal wavelet: 0=Haar, 1=CDF 5/3 (extracted from version)
int no_grain_synthesis; // Command-line flag: disable grain synthesis
// Screen masking (letterbox/pillarbox) - array of geometry changes
screen_mask_entry_t *screen_masks;
@@ -2023,10 +2035,11 @@ static int extract_audio_to_wav(const char *input_file, const char *wav_file, in
// Decoder Initialisation and Cleanup
//=============================================================================
static tav_decoder_t* tav_decoder_init(const char *input_file, const char *output_file, const char *audio_file) {
static tav_decoder_t* tav_decoder_init(const char *input_file, const char *output_file, const char *audio_file, int no_grain_synthesis) {
tav_decoder_t *decoder = calloc(1, sizeof(tav_decoder_t));
if (!decoder) return NULL;
decoder->no_grain_synthesis = no_grain_synthesis;
decoder->input_fp = fopen(input_file, "rb");
if (!decoder->input_fp) {
free(decoder);
@@ -2511,8 +2524,9 @@ static int decode_i_or_p_frame(tav_decoder_t *decoder, uint8_t packet_type, uint
// Remove grain synthesis from Y channel (must happen after dequantisation, before inverse DWT)
// Phase 2: Use decoding dimensions and temporary buffer
remove_grain_synthesis_decoder(temp_dwt_y, decoder->decoding_width, decoder->decoding_height,
decoder->header.decomp_levels, decoder->frame_count, decoder->header.quantiser_y);
apply_grain_synthesis(temp_dwt_y, decoder->decoding_width, decoder->decoding_height,
decoder->header.decomp_levels, decoder->frame_count, decoder->header.quantiser_y,
decoder->header.encoder_preset, decoder->no_grain_synthesis);
// Debug: Check LL band AFTER grain removal
// if (decoder->frame_count == 32) {
@@ -2712,10 +2726,11 @@ static void print_usage(const char *prog) {
printf("Version: %s\n\n", DECODER_VENDOR_STRING);
printf("Usage: %s -i input.tav -o output.mkv\n\n", prog);
printf("Options:\n");
printf(" -i <file> Input TAV file\n");
printf(" -o <file> Output MKV file (optional, auto-generated from input)\n");
printf(" -v Verbose output\n");
printf(" -h, --help Show this help\n\n");
printf(" -i <file> Input TAV file\n");
printf(" -o <file> Output MKV file (optional, auto-generated from input)\n");
printf(" -v Verbose output\n");
printf(" --no-grain-synthesis Disable grain synthesis (override encoder preset)\n");
printf(" -h, --help Show this help\n\n");
printf("Supported features (matches TSVM decoder):\n");
printf(" - I-frames and P-frames (delta mode)\n");
printf(" - GOP unified 3D DWT (temporal compression)\n");
@@ -2740,9 +2755,11 @@ int main(int argc, char *argv[]) {
char *input_file = NULL;
char *output_file = NULL;
int verbose = 0;
int no_grain_synthesis = 0;
static struct option long_options[] = {
{"help", no_argument, 0, 'h'},
{"no-grain-synthesis", no_argument, 0, 1000},
{0, 0, 0, 0}
};
@@ -2761,6 +2778,12 @@ int main(int argc, char *argv[]) {
case 'h':
print_usage(argv[0]);
return 0;
case 1000: // --no-grain-synthesis
no_grain_synthesis = 1;
if (verbose) {
printf("Grain synthesis disabled\n");
}
break;
default:
print_usage(argv[0]);
return 1;
@@ -2819,7 +2842,7 @@ int main(int argc, char *argv[]) {
}
// Pass 2: Decode video with audio file
tav_decoder_t *decoder = tav_decoder_init(input_file, output_file, temp_audio_file);
tav_decoder_t *decoder = tav_decoder_init(input_file, output_file, temp_audio_file, no_grain_synthesis);
if (!decoder) {
fprintf(stderr, "Failed to initialise decoder\n");
unlink(temp_audio_file); // Clean up temp file
@@ -3126,7 +3149,7 @@ int main(int argc, char *argv[]) {
// EZBC mode with perceptual quantisation: coefficients are normalised
// Need to dequantise using perceptual weights (same as twobit-map mode)
const int temporal_level = get_temporal_subband_level(t, gop_size, temporal_levels);
const float temporal_scale = get_temporal_quantiser_scale(temporal_level);
const float temporal_scale = get_temporal_quantiser_scale(decoder->header.encoder_preset, temporal_level);
// FIX: Use QLUT to convert header quantiser indices to actual values
const float base_q_y = roundf(QLUT[decoder->header.quantiser_y] * temporal_scale);
@@ -3160,7 +3183,7 @@ int main(int argc, char *argv[]) {
} else if (!is_ezbc) {
// Normal mode: multiply by quantiser
const int temporal_level = get_temporal_subband_level(t, gop_size, temporal_levels);
const float temporal_scale = get_temporal_quantiser_scale(temporal_level);
const float temporal_scale = get_temporal_quantiser_scale(decoder->header.encoder_preset, temporal_level);
// CRITICAL: Must ROUND temporal quantiser to match encoder's roundf() behavior
// FIX: Use QLUT to convert header quantiser indices to actual values
@@ -3206,9 +3229,10 @@ int main(int argc, char *argv[]) {
// Phase 2: Use GOP dimensions (may be cropped) for grain removal
for (int t = 0; t < gop_size; t++) {
remove_grain_synthesis_decoder(gop_y[t], gop_width, gop_height,
apply_grain_synthesis(gop_y[t], gop_width, gop_height,
decoder->header.decomp_levels, decoder->frame_count + t,
decoder->header.quantiser_y);
decoder->header.quantiser_y, decoder->header.encoder_preset,
decoder->no_grain_synthesis);
}
// Apply inverse 3D DWT (spatial + temporal)

46
video_encoder/encoder_tav.c
View File

@@ -19,7 +19,7 @@
#include <float.h>
#include "tav_avx512.h" // AVX-512 SIMD optimisations
#define ENCODER_VENDOR_STRING "Encoder-TAV 20251124 (3d-dwt,tad,ssf-tc,cdf53-motion,avx512)"
#define ENCODER_VENDOR_STRING "Encoder-TAV 20251124 (3d-dwt,tad,ssf-tc,cdf53-motion,avx512,presets)"
// TSVM Advanced Video (TAV) format constants
#define TAV_MAGIC "\x1F\x54\x53\x56\x4D\x54\x41\x56" // "\x1FTSVM TAV"
@@ -1835,6 +1835,7 @@ typedef struct tav_encoder_s {
int pcm8_audio; // 1 = use 8-bit PCM audio (packet 0x21), 0 = use MP2 (default)
int tad_audio; // 1 = use TAD audio (packet 0x24), 0 = use MP2/PCM8 (default, quality follows quality_level)
int enable_crop_encoding; // 1 = encode cropped active region only (Phase 2), 0 = encode full frame (default)
uint8_t encoder_preset; // Encoder preset flags: bit 0 = sports (finer temporal quantisation), bit 1 = anime (no grain)
// Active region tracking (for Phase 2 crop encoding)
uint16_t active_mask_top, active_mask_right, active_mask_bottom, active_mask_left;
@@ -2432,6 +2433,9 @@ static void show_usage(const char *program_name) {
printf(" --dump-frame N Dump quantised coefficients for frame N (creates .bin files)\n");
printf(" --wavelet N Wavelet filter: 0=LGT 5/3, 1=CDF 9/7, 2=CDF 13/7, 16=DD-4, 255=Haar (default: 1)\n");
printf(" --zstd-level N Zstd compression level 1-22 (default: %d, higher = better compression but slower)\n", DEFAULT_ZSTD_LEVEL);
printf(" --preset PRESET Encoder presets (comma-separated, e.g., 'sports,anime'):\n");
printf(" sports (or sport): Finer temporal quantisation for better motion detail\n");
printf(" anime (or animation): Disable grain synthesis for cleaner animated content\n");
printf(" --help Show this help\n\n");
printf("Audio Rate by Quality:\n ");
@@ -3355,8 +3359,9 @@ static void quantise_3d_dwt_coefficients(tav_encoder_t *enc,
int spatial_size,
int base_quantiser,
int is_chroma) {
const float BETA = 0.6f; // Temporal scaling exponent (aggressive for temporal high-pass)
const float KAPPA = 1.14f;
// Sports preset: use finer temporal quantisation (less aggressive)
const float BETA = (enc->encoder_preset & 0x01) ? 0.0f : 0.6f;
const float KAPPA = (enc->encoder_preset & 0x01) ? 1.0f : 1.14f;
// Process each temporal subband independently (separable approach)
for (int t = 0; t < num_frames; t++) {
@@ -7528,8 +7533,10 @@ static int write_tav_header(tav_encoder_t *enc) {
// Entropy Coder (0 = Twobit-map, 1 = EZBC, 2 = Raw)
fputc(enc->preprocess_mode, enc->output_fp);
// Reserved bytes (2 bytes)
fputc(0, enc->output_fp);
// Encoder Preset (byte 28): bit 0 = sports, bit 1 = anime
fputc(enc->encoder_preset, enc->output_fp);
// Reserved byte (1 byte)
fputc(0, enc->output_fp);
// Device Orientation (default: 0 = no rotation)
@@ -10775,6 +10782,7 @@ int main(int argc, char *argv[]) {
{"tad-audio", no_argument, 0, 1028},
{"raw-coeffs", no_argument, 0, 1029},
{"single-pass", no_argument, 0, 1050}, // disable two-pass encoding with wavelet-based scene detection
{"preset", required_argument, 0, 1051}, // Encoder presets: sports, anime (comma-separated)
{"enable-crop-encoding", no_argument, 0, 1052}, // Phase 2: encode cropped active region only (experimental)
{"help", no_argument, 0, '?'},
{0, 0, 0, 0}
@@ -11012,6 +11020,34 @@ int main(int argc, char *argv[]) {
enc->two_pass_mode = 0;
printf("Two-pass wavelet-based scene change detection disabled\n");
break;
case 1051: { // --preset
char *preset_str = strdup(optarg);
char *token = strtok(preset_str, ",");
while (token != NULL) {
// Trim leading/trailing whitespace
while (*token == ' ' || *token == '\t') token++;
char *end = token + strlen(token) - 1;
while (end > token && (*end == ' ' || *end == '\t')) {
*end = '\0';
end--;
}
// Check for presets and aliases
if (strcmp(token, "sports") == 0 || strcmp(token, "sport") == 0) {
enc->encoder_preset |= 0x01;
printf("Preset 'sports' enabled: finer temporal quantisation (BETA=0.25, KAPPA=1.0)\n");
} else if (strcmp(token, "anime") == 0 || strcmp(token, "animation") == 0) {
enc->encoder_preset |= 0x02;
printf("Preset 'anime' enabled: grain synthesis disabled\n");
} else {
fprintf(stderr, "Warning: Unknown preset '%s' (valid: sports, anime)\n", token);
}
token = strtok(NULL, ",");
}
free(preset_str);
break;
}
case 1052: // --enable-crop-encoding
enc->enable_crop_encoding = 1;
printf("Phase 2 crop encoding enabled (experimental)\n");

4
video_encoder/tav_avx512.h
View File

@@ -456,7 +456,7 @@ static inline void quantise_dwt_coefficients_avx512(
quant = _mm512_mask_blend_ps(dead_mask, quant, zero_vec);
}
// Manual rounding to match scalar behavior (round away from zero)
// Manual rounding to match scalar behaviour (round away from zero)
// First add 0.5 or -0.5 based on sign
__mmask16 pos_mask = _mm512_cmp_ps_mask(quant, zero_vec, _CMP_GE_OQ);
__m512 round_val = _mm512_mask_blend_ps(pos_mask, nhalf_vec, half_vec);
@@ -510,7 +510,7 @@ static inline void quantise_dwt_coefficients_perceptual_avx512(
__m512 effective_q = _mm512_mul_ps(base_q_vec, weight);
__m512 quant = _mm512_div_ps(coeff, effective_q);
// Manual rounding to match scalar behavior
// Manual rounding to match scalar behaviour
__mmask16 pos_mask = _mm512_cmp_ps_mask(quant, zero_vec, _CMP_GE_OQ);
__m512 round_val = _mm512_mask_blend_ps(pos_mask, nhalf_vec, half_vec);
quant = _mm512_add_ps(quant, round_val);

16
video_encoder/tav_inspector.c
View File

@@ -514,6 +514,7 @@ int main(int argc, char *argv[]) {
uint8_t quality = header[25];
uint8_t channel_layout = header[26];
uint8_t entropy_coder = header[27];
uint8_t encoder_preset = header[28];
static const int QLUT[] = {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126,128,132,136,140,144,148,152,156,160,164,168,172,176,180,184,188,192,196,200,204,208,212,216,220,224,228,232,236,240,244,248,252,256,264,272,280,288,296,304,312,320,328,336,344,352,360,368,376,384,392,400,408,416,424,432,440,448,456,464,472,480,488,496,504,512,528,544,560,576,592,608,624,640,656,672,688,704,720,736,752,768,784,800,816,832,848,864,880,896,912,928,944,960,976,992,1008,1024,1056,1088,1120,1152,1184,1216,1248,1280,1312,1344,1376,1408,1440,1472,1504,1536,1568,1600,1632,1664,1696,1728,1760,1792,1824,1856,1888,1920,1952,1984,2016,2048,2112,2176,2240,2304,2368,2432,2496,2560,2624,2688,2752,2816,2880,2944,3008,3072,3136,3200,3264,3328,3392,3456,3520,3584,3648,3712,3776,3840,3904,3968,4032,4096};
static const char* CLAYOUT[] = {"Luma-Chroma", "Luma-Chroma-Alpha", "Luma", "Luma-Alpha", "Chroma", "Chroma-Alpha"};
@@ -548,6 +549,21 @@ static const char* TEMPORAL_WAVELET[] = {"Haar", "CDF 5/3"};
printf(" Quality: n/a\n");
printf(" Channel layout: %s\n", CLAYOUT[channel_layout]);
printf(" Entropy coder: %s\n", entropy_coder == 0 ? "Twobit-map" : "EZBC");
printf(" Encoder preset: ");
if (encoder_preset == 0) {
printf("Default\n");
} else {
int first = 1;
if (encoder_preset & 0x01) {
printf("%sSports", first ? "" : ", ");
first = 0;
}
if (encoder_preset & 0x02) {
printf("%sAnime", first ? "" : ", ");
first = 0;
}
printf("\n");
}
printf(" Flags:\n");
printf(" Has audio: %s\n", (extra_flags & 0x01) ? "Yes" : "No");
printf(" Has subtitles: %s\n", (extra_flags & 0x02) ? "Yes" : "No");