"""
Генератор ROM Atari на CUDA
"""

import cupy as cp
import numpy as np
import time
from pathlib import Path

# Константы
ROM_SIZE = 4094  # Вектор сброса тестируется позже
BATCH_SIZE = 1024 * 256

# Пороговые значения, основанные на выявленных паттернах
OPCODE_THRESHOLD = 0.58
TIA_THRESHOLD = 50
RIOT_THRESHOLD = 13
BRANCH_THRESHOLD = 150
JUMP_THRESHOLD = 37
INSTRUCTION_VARIETY = 100
MIN_SCORE = 0.52

# Валидные опкоды 6502 (всего 151)
VALID_OPCODES = np.array([
    0x00, 0x01, 0x05, 0x06, 0x08, 0x09, 0x0A, 0x0D, 0x0E, 0x10, 0x11, 0x15, 0x16, 0x18,
    0x19, 0x1D, 0x1E, 0x20, 0x21, 0x24, 0x25, 0x26, 0x28, 0x29, 0x2A, 0x2C, 0x2D, 0x2E,
    0x30, 0x31, 0x35, 0x36, 0x38, 0x39, 0x3D, 0x3E, 0x40, 0x41, 0x45, 0x46, 0x48, 0x49,
    0x4A, 0x4C, 0x4D, 0x4E, 0x50, 0x51, 0x55, 0x56, 0x58, 0x59, 0x5D, 0x5E, 0x60, 0x61,
    0x65, 0x66, 0x68, 0x69, 0x6A, 0x6C, 0x6D, 0x6E, 0x70, 0x71, 0x75, 0x76, 0x78, 0x79,
    0x7D, 0x7E, 0x81, 0x84, 0x85, 0x86, 0x88, 0x8A, 0x8C, 0x8D, 0x8E, 0x90, 0x91, 0x94,
    0x95, 0x96, 0x98, 0x99, 0x9A, 0x9D, 0xA0, 0xA1, 0xA2, 0xA4, 0xA5, 0xA6, 0xA8, 0xA9,
    0xAA, 0xAC, 0xAD, 0xAE, 0xB0, 0xB1, 0xB4, 0xB5, 0xB6, 0xB8, 0xB9, 0xBA, 0xBC, 0xBD,
    0xBE, 0xC0, 0xC1, 0xC4, 0xC5, 0xC6, 0xC8, 0xC9, 0xCA, 0xCC, 0xCD, 0xCE, 0xD0, 0xD1,
    0xD5, 0xD6, 0xD8, 0xD9, 0xDD, 0xDE, 0xE0, 0xE1, 0xE4, 0xE5, 0xE6, 0xE8, 0xE9, 0xEA,
    0xEC, 0xED, 0xEE, 0xF0, 0xF1, 0xF5, 0xF6, 0xF8, 0xF9, 0xFD, 0xFE
], dtype=np.uint8)

# Опкоды потока управления
BRANCH_OPCODES = np.array([0x10, 0x30, 0x50, 0x70, 0x90, 0xB0, 0xD0, 0xF0], dtype=np.uint8)
JUMP_OPCODES = np.array([0x4C, 0x6C, 0x20], dtype=np.uint8)

def create_lookup_tables():
    """Создаём таблицы поиска GPU для анализа ROM"""
    valid_lut = cp.zeros(256, dtype=cp.bool_)
    valid_lut[VALID_OPCODES] = True
    
    branch_lut = cp.zeros(256, dtype=cp.bool_)
    branch_lut[BRANCH_OPCODES] = True
    
    jump_lut = cp.zeros(256, dtype=cp.bool_)
    jump_lut[JUMP_OPCODES] = True
    
    # Поиск команд TIA
    tia_store_lut = cp.zeros(256, dtype=cp.bool_)
    tia_store_lut[[0x85, 0x86, 0x84, 0x95, 0x96, 0x94]] = True
    
    tia_load_lut = cp.zeros(256, dtype=cp.bool_)
    tia_load_lut[[0xA5, 0xA6, 0xA4, 0xB5, 0xB6, 0xB4]] = True
    
    tia_abs_lut = cp.zeros(256, dtype=cp.bool_)
    tia_abs_lut[[0x8D, 0x8E, 0x8C, 0xAD, 0xAE, 0xAC]] = True
    
    # Поиск команд RIOT
    riot_access_lut = cp.zeros(256, dtype=cp.bool_)
    riot_access_lut[[0x85, 0x86, 0x84, 0xA5, 0xA6, 0xA4]] = True
    
    # Маски диапазонов адресов
    tia_range_mask = cp.arange(256, dtype=cp.uint8) <= 0x2F
    riot_timer_mask = (cp.arange(256, dtype=cp.uint8) >= 0x80) & (cp.arange(256, dtype=cp.uint8) <= 0x87)
    riot_io_mask = (cp.arange(256, dtype=cp.uint8) >= 0x94) & (cp.arange(256, dtype=cp.uint8) <= 0x97)
    
    return {
        'valid': valid_lut,
        'branch': branch_lut,
        'jump': jump_lut,
        'tia_store': tia_store_lut,
        'tia_load': tia_load_lut,
        'tia_abs': tia_abs_lut,
        'riot_access': riot_access_lut,
        'tia_range': tia_range_mask,
        'riot_timer': riot_timer_mask,
        'riot_io': riot_io_mask
    }

def analyze_roms(roms, lut):
    """Анализируем ROM на наличие похожих на игры паттернов"""
    batch_size = roms.shape[0]
    
    # Анализ опкодов
    valid_opcodes_count = cp.sum(lut['valid'][roms], axis=1)
    opcode_ratio = valid_opcodes_count.astype(cp.float32) / ROM_SIZE
    
    # Анализ потока выполнения
    branch_count = cp.sum(lut['branch'][roms], axis=1)
    jump_count = cp.sum(lut['jump'][roms], axis=1)
    
    # Анализ TIA
    tia_accesses = cp.zeros(batch_size, dtype=cp.int32)
    
    # Адресация в режиме нулевой страницы
    tia_store_zp = lut['tia_store'][roms[:, :-1]] & lut['tia_range'][roms[:, 1:]]
    tia_load_zp = lut['tia_load'][roms[:, :-1]] & lut['tia_range'][roms[:, 1:]]
    tia_zp_total = cp.sum(tia_store_zp | tia_load_zp, axis=1)
    tia_accesses += tia_zp_total
    
    # Абсолютная адресация
    tia_abs_positions = lut['tia_abs'][roms[:, :-2]]
    tia_abs_targets = lut['tia_range'][roms[:, 1:-1]] & (roms[:, 2:] == 0x00)
    tia_abs_total = cp.sum(tia_abs_positions & tia_abs_targets, axis=1)
    tia_accesses += tia_abs_total
    
    # Анализ RIOT
    riot_accesses = cp.zeros(batch_size, dtype=cp.int32)
    
    # Доступ к таймеру
    riot_timer_positions = lut['riot_access'][roms[:, :-1]]
    riot_timer_targets = lut['riot_timer'][roms[:, 1:]]
    riot_timer_hits = cp.sum(riot_timer_positions & riot_timer_targets, axis=1)
    riot_accesses += riot_timer_hits
    
    # Доступ к вводу-выводу
    riot_io_positions = lut['riot_access'][roms[:, :-1]]
    riot_io_targets = lut['riot_io'][roms[:, 1:]]
    riot_io_hits = cp.sum(riot_io_positions & riot_io_targets, axis=1)
    riot_accesses += riot_io_hits
    
    # Подсчёт уникальных опкодов в первом 1 КБ (разделе кода)
    unique_opcodes = cp.zeros(batch_size, dtype=cp.int32)
    first_kb = roms[:, :1024]  # Первый 1 КБ, где обычно находится код
    
    # Подсчёт уникальных опкодов в разделе кода
    for opcode in VALID_OPCODES:
        has_opcode = cp.any(first_kb == opcode, axis=1)
        unique_opcodes += has_opcode.astype(cp.int32)
    
    # Составная оценка
    scores = (
        opcode_ratio * 0.25 + 
        cp.minimum(tia_accesses / 150.0, 1.0) * 0.30 +
        cp.minimum(riot_accesses / 50.0, 1.0) * 0.20 +
        cp.minimum(branch_count / 200.0, 1.0) * 0.15 +
        cp.minimum(jump_count / 40.0, 1.0) * 0.10
    )
    
    # Распознавание перспективных ROM
    promising = (
        (opcode_ratio >= OPCODE_THRESHOLD) &
        (tia_accesses >= TIA_THRESHOLD) &
        (riot_accesses >= RIOT_THRESHOLD) &
        (branch_count >= BRANCH_THRESHOLD) &
        (jump_count >= JUMP_THRESHOLD) &
        (unique_opcodes >= INSTRUCTION_VARIETY) &
        (scores >= MIN_SCORE)
    )
    
    return {
        'scores': scores,
        'opcode_ratio': opcode_ratio,
        'tia_accesses': tia_accesses,
        'riot_accesses': riot_accesses,
        'branch_count': branch_count,
        'jump_count': jump_count,
        'unique_opcodes': unique_opcodes,
        'promising': promising
    }

def save_promising_rom(rom_data, score, rom_id, output_dir):
    """Сохраняем перспективные ROM в формате number_score_timestamp.bin"""
    timestamp = int(time.time())
    filename = f"{rom_id:06d}_{score:.3f}_{timestamp}.bin"
    filepath = output_dir / filename
    
    with open(filepath, 'wb') as f:
        f.write(rom_data.tobytes())
    
    return filename

def main():
    print("Finite Atari Machine - Streamlined CUDA Generator")
    print("=" * 60)
    print(f"Batch size: {BATCH_SIZE:,} ROMs per batch")
    print(f"ROM size: {ROM_SIZE:,} bytes")
    print()
    print("Thresholds:")
    print(f"  Opcodes: {OPCODE_THRESHOLD:.1%}")
    print(f"  TIA: {TIA_THRESHOLD}+")
    print(f"  RIOT: {RIOT_THRESHOLD}+")
    print(f"  Branches: {BRANCH_THRESHOLD}+")
    print(f"  Jumps: {JUMP_THRESHOLD}+")
    print(f"  Unique opcodes: {INSTRUCTION_VARIETY}+")
    print(f"  Min score: {MIN_SCORE:.2f}")
    print()
    
    # Информация о GPU
    try:
        gpu_props = cp.cuda.runtime.getDeviceProperties(0)
        gpu_name = gpu_props['name'].decode()
        total_mem = cp.cuda.runtime.memGetInfo()[1] // 1024**2
        print(f"GPU: {gpu_name}")
        print(f"Memory: {total_mem:,} MB")
    except Exception:
        print("GPU: CuPy device detected")
    
    print("\nInitializing lookup tables...")
    
    # Подготовка
    output_dir = Path("finite_atari_roms")
    output_dir.mkdir(exist_ok=True)
    
    lookup_tables = create_lookup_tables()
    
    # Статистика
    total_generated = 0
    promising_found = 0
    start_time = time.time()
    last_report = start_time
    best_score_ever = 0.0
    
    print("Starting ROM generation...")
    print("=" * 60)
    
    try:
        while True:
            batch_start = time.time()
            
            # Генерируем батч ROM
            roms = cp.random.randint(0, 256, size=(BATCH_SIZE, ROM_SIZE), dtype=cp.uint8)
            
            # Анализируем ROM
            analysis = analyze_roms(roms, lookup_tables)
            
            # Отслеживаем наилучшую оценку
            current_best = float(cp.max(analysis['scores']))
            if current_best > best_score_ever:
                best_score_ever = current_best
            
            # Проверяем перспективные ROM
            promising_indices = cp.where(analysis['promising'])[0]
            
            if len(promising_indices) > 0:
                # Сохраняем перспективные ROM
                promising_roms = cp.asnumpy(roms[promising_indices])
                promising_scores = cp.asnumpy(analysis['scores'][promising_indices])
                
                for i in range(len(promising_indices)):
                    filename = save_promising_rom(
                        promising_roms[i], promising_scores[i], promising_found, output_dir
                    )
                    promising_found += 1
            
            total_generated += BATCH_SIZE
            batch_time = time.time() - batch_start
            
            # Отчёт о прогрессе
            current_time = time.time()
            if current_time - last_report >= 4:
                # Получаем наилучшую статистику ROM для этого батча
                scores = cp.asnumpy(analysis['scores'])
                best_idx = np.argmax(scores)
                best_opcodes = float(analysis['opcode_ratio'][best_idx])
                best_tia = int(analysis['tia_accesses'][best_idx])
                best_riot = int(analysis['riot_accesses'][best_idx])
                best_branches = int(analysis['branch_count'][best_idx])
                best_jumps = int(analysis['jump_count'][best_idx])
                
                elapsed = current_time - start_time
                rate = total_generated / elapsed
                success_rate = promising_found / total_generated * 100 if total_generated > 0 else 0
                
                print(f"\rGenerated: {total_generated:,} | Found: {promising_found} | "
                      f"Success: {success_rate:.8f}% | Rate: {rate:,.0f}/sec | "
                      f"Best: {best_score_ever:.3f} | "
                      f"Op:{best_opcodes:.1%} TIA:{best_tia} RIOT:{best_riot} Br:{best_branches} Jmp:{best_jumps}", 
                      end="", flush=True)
                
                last_report = current_time
    
    except KeyboardInterrupt:
        elapsed = time.time() - start_time
        rate = total_generated / elapsed
        success_rate = promising_found / total_generated * 100 if total_generated > 0 else 0
        
        print(f"\n\nStopped after {elapsed:.1f} seconds")
        print(f"Total ROMs generated: {total_generated:,}")
        print(f"Promising ROMs found: {promising_found}")
        print(f"Success rate: {success_rate:.8f}%")
        print(f"Average rate: {rate:,.0f} ROMs/second")
        print(f"Best score achieved: {best_score_ever:.4f}")
        print(f"Results saved in: {output_dir}")

if __name__ == "__main__":
    main()

# ==========================================================================
#
#   Единый конвейер Finite Atari (4 КБ, сброс @ $F000) - ИСПРАВЛЕННАЯ ВЕРСИЯ
#
#   Этот скрипт генерирует случайные ROM Atari 2600 на  GPU при помощи CUDA,
#   фильтрует их на GPU при помощи эвристик, и если они интересные, запускает
#   их в безголовом MAME на 2 с, чтобы проверить, есть ли в них динамическое видео.
#
#   Скрипт предназначен для работы на GPU с CUDA и с MAME, установленным в PATH.
#
#   Вывод сохраняется в папку "finite_atari_roms".
#
# ===========================================================================

from __future__ import annotations
import cupy as cp, numpy as np, hashlib, subprocess, tempfile, time, textwrap
from pathlib import Path
from PIL import Image

# ─── 1.  Глобальные константы ────────────────────────────────────────────────────
ROM_SIZE       = 4096
PAYLOAD_BYTES  = ROM_SIZE - 2
RESET_VECTOR   = (0x00, 0xF0)         # $F000 в little-endian 
BATCH_SIZE     = 1024 * 256           # примерно 256 тысяч ROM на батч GPU
STATUS_EVERY   = 10                   # количество батчей между выводом статуса
OUTPUT_DIR     = Path("finite_atari_roms"); OUTPUT_DIR.mkdir(exist_ok=True)

# Пороговые значения для видео
BLACK_LEVEL          = 15             # 0-255 серый; ≤ означает "чёрный"
NONBLACK_THRESHOLD   = 0.005          # ≥ 0,5 % пикселей ярче ⇒ видео
DYNAMIC_THRESHOLD    = 0.01           # ≥ 1 % хэшированных пикселей различаются ⇒ движение

# Пороговые значения эвристик (те же, что и для предыдущих скриптов Stella/MAME)
OPCODE_THRESHOLD    = 0.58
TIA_THRESHOLD       = 50
RIOT_THRESHOLD      = 13
BRANCH_THRESHOLD    = 150
JUMP_THRESHOLD      = 37
INSTRUCTION_VARIETY = 100
MIN_SCORE           = 0.52

# ─── 2.  Таблицы поиска опкодов ────────────────────────────────────────────────
# Валидные опкоды 6502 для контекста самодельных игр 2600
VALID_OPCODES = np.array([
    0x00,0x01,0x05,0x06,0x08,0x09,0x0A,0x0D,0x0E,0x10,0x11,0x15,0x16,0x18,
    0x19,0x1D,0x1E,0x20,0x21,0x24,0x25,0x26,0x28,0x29,0x2A,0x2C,0x2D,0x2E,
    0x30,0x31,0x35,0x36,0x38,0x39,0x3D,0x3E,0x40,0x41,0x45,0x46,0x48,0x49,
    0x4A,0x4C,0x4D,0x4E,0x50,0x51,0x55,0x56,0x58,0x59,0x5D,0x5E,0x60,0x61,
    0x65,0x66,0x68,0x69,0x6A,0x6C,0x6D,0x6E,0x70,0x71,0x75,0x76,0x78,0x79,
    0x7D,0x7E,0x81,0x84,0x85,0x86,0x88,0x8A,0x8C,0x8D,0x8E,0x90,0x91,0x94,
    0x95,0x96,0x98,0x99,0x9A,0x9D,0xA0,0xA1,0xA2,0xA4,0xA5,0xA6,0xA8,0xA9,
    0xAA,0xAC,0xAD,0xAE,0xB0,0xB1,0xB4,0xB5,0xB6,0xB8,0xB9,0xBA,0xBC,0xBD,
    0xBE,0xC0,0xC1,0xC4,0xC5,0xC6,0xC8,0xC9,0xCA,0xCC,0xCD,0xCE,0xD0,0xD1,
    0xD5,0xD6,0xD8,0xD9,0xDD,0xDE,0xE0,0xE1,0xE4,0xE5,0xE6,0xE8,0xE9,0xEA,
    0xEC,0xED,0xEE,0xF0,0xF1,0xF5,0xF6,0xF8,0xF9,0xFD,0xFE], dtype=np.uint8)

BRANCH_OPCODES = np.array([0x10,0x30,0x50,0x70,0x90,0xB0,0xD0,0xF0], dtype=np.uint8)
JUMP_OPCODES   = np.array([0x4C,0x6C,0x20], dtype=np.uint8)

def create_luts():
    """Возвращает список булеву таблицу поиска с 256 элементами (cupy)."""
    lut = {}
    lut["valid"]  = cp.zeros(256, cp.bool_); lut["valid"][VALID_OPCODES]  = True
    lut["branch"] = cp.zeros(256, cp.bool_); lut["branch"][BRANCH_OPCODES] = True
    lut["jump"]   = cp.zeros(256, cp.bool_); lut["jump"][JUMP_OPCODES]     = True

    # Особенности адресации 2600 для выявления операций доступа к TIA/RIOT
    lut["tia_store"] = cp.zeros(256, cp.bool_)
    lut["tia_store"][[0x84,0x85,0x86, 0x94,0x95,0x96]] = True   # STY/STA/STX (zp & zp,x)
    lut["tia_load"]  = cp.zeros(256, cp.bool_)
    lut["tia_load" ][[0xA4,0xA5,0xA6, 0xB4,0xB5,0xB6]] = True   # LDY/LDA/LDX (zp & zp,x)
    lut["tia_abs"]   = cp.zeros(256, cp.bool_)
    lut["tia_abs"  ][[0x8C,0x8D,0x8E, 0xAC,0xAD,0xAE]] = True   # аналоги для абсолютных значений

    lut["riot_acc"]  = cp.zeros(256, cp.bool_)
    lut["riot_acc"][[0x84,0x85,0x86, 0xA4,0xA5,0xA6]] = True

    addr = cp.arange(256, dtype=cp.uint8)
    lut["tia_range"] = addr <= 0x2F
    lut["riot_tmr"]  = (addr >= 0x80) & (addr <= 0x87)
    lut["riot_io"]   = (addr >= 0x94) & (addr <= 0x97)
    return lut

# ─── 3.  Эвристический фильтр на GPU ────────────────────────────────────────────────
def analyse_batch(roms: cp.ndarray, lut) -> tuple[np.ndarray, cp.ndarray]:
    """
    Возвращает (interesting_mask, scores) для 2D-массива uint8 ROM.
    Каждая строка = один ROM.
    """
    valid_cnt    = cp.sum(lut["valid"][roms], axis=1)
    opcode_ratio = valid_cnt.astype(cp.float32) / ROM_SIZE
    branch_cnt   = cp.sum(lut["branch"][roms], axis=1)
    jump_cnt     = cp.sum(lut["jump"  ][roms], axis=1)

    # --- Операции доступа к TIA --------------------------------------------------------
    tia_acc  = cp.sum((lut["tia_store"][roms[:,:-1]] | lut["tia_load"][roms[:,:-1]])
                      & lut["tia_range"][roms[:,1:]], axis=1)
    tia_acc += cp.sum(lut["tia_abs"][roms[:,:-2]]
                      & lut["tia_range"][roms[:,1:-1]]
                      & (roms[:,2:] == 0x00), axis=1)

    # --- Операции доступа к RIOT -------------------------------------------------------
    riot_acc  = cp.sum(lut["riot_acc"][roms[:,:-1]] & lut["riot_tmr"][roms[:,1:]], axis=1)
    riot_acc += cp.sum(lut["riot_acc"][roms[:,:-1]] & lut["riot_io" ][roms[:,1:]], axis=1)

    # --- Разнообразие опкодов в первом 1 КБ --------------------------------------
    uniq = cp.zeros(roms.shape[0], dtype=cp.int32)
    first_kb = roms[:, :1024]
    for op in VALID_OPCODES:
        uniq += cp.any(first_kb == op, axis=1)

    scores = (opcode_ratio * 0.25 +
              cp.minimum(tia_acc / 150.0, 1.0) * 0.30 +
              cp.minimum(riot_acc / 50.0, 1.0)  * 0.20 +
              cp.minimum(branch_cnt / 200.0, 1.0) * 0.15 +
              cp.minimum(jump_cnt / 40.0, 1.0)   * 0.10)

    interesting = ((opcode_ratio >= OPCODE_THRESHOLD) &
                   (tia_acc      >= TIA_THRESHOLD) &
                   (riot_acc     >= RIOT_THRESHOLD) &
                   (branch_cnt   >= BRANCH_THRESHOLD) &
                   (jump_cnt     >= JUMP_THRESHOLD) &
                   (uniq         >= INSTRUCTION_VARIETY) &
                   (scores       >= MIN_SCORE))

    return interesting, scores

# ─── 4.  Вспомогательный скрипт на Lua (снэпшот двух кадров) - ИСПРАВЛЕНО ────────────────────
SNAPSHOT_LUA = textwrap.dedent("""
    local s = manager.machine.screens[":screen"]
    local frame_count = 0
    emu.register_frame_done(function ()
        frame_count = frame_count + 1
        if     frame_count == 1  then s:snapshot("first.png")
        elseif frame_count == 60 then s:snapshot("second.png"); manager.machine:exit() end
    end, "snapper")
""")

# ─── 5. Вспомогательные функции анализа видео ──────────────────────────────────────────────
def _hash16(img: Path) -> str:
    with Image.open(img) as im:
        im = im.convert("L").resize((16,16), Image.NEAREST)
        return hashlib.sha1(im.tobytes()).hexdigest()

def _frame_is_nonblack(img: Path) -> bool:
    with Image.open(img) as im:
        g = np.asarray(im.convert("L"))
    return (g > BLACK_LEVEL).mean() >= NONBLACK_THRESHOLD

def rom_video_flags(rom: bytes, *, mame="mame", seconds=2.0) -> tuple[bool,bool]:
    """
    Возвращает (has_video, is_dynamic).
    • has_video  → первый кадр не чёрный
    • is_dynamic → ≥ 1 % хэшированных пикселей различаются между кадром  1 и 60
    """
    with tempfile.TemporaryDirectory() as td_s:
        td = Path(td_s)
        (td / "test.bin").write_bytes(rom)
        (td / "snapshot.lua").write_text(SNAPSHOT_LUA)

        base = [mame, "a2600", "-cart", "test.bin",
                "-seconds_to_run", str(seconds),
                "-nothrottle", "-window", "-sound", "none", "-skip_gameinfo"]

        for flag in ("-autoboot_script", "-script"):
            try:
                subprocess.run(base + [flag, "snapshot.lua"],
                               cwd=td, stdout=subprocess.DEVNULL,
                               stderr=subprocess.DEVNULL, timeout=seconds*5,
                               check=True)
                break
            except (subprocess.CalledProcessError, subprocess.TimeoutExpired):
                if flag == "-autoboot_script":
                    continue
                return (False, False)

        # Проверяем наличие кадров в корневой папке и в подпапке snap
        f1, f2 = td / "first.png", td / "second.png"
        snap_dir = td / "snap"
        if not f1.exists() and snap_dir.exists():
            snap_f1 = snap_dir / "first.png"
            snap_f2 = snap_dir / "second.png"
            if snap_f1.exists():
                f1 = snap_f1
            if snap_f2.exists():
                f2 = snap_f2

        if not f1.exists():
            return (False, False)

        nonblack = _frame_is_nonblack(f1)
        if not nonblack or not f2.exists():
            return (nonblack, False)

        diff_bits = bin(int(_hash16(f1),16) ^ int(_hash16(f2),16)).count("1")
        dynamic = diff_bits / 256.0 >= DYNAMIC_THRESHOLD
        return (nonblack, dynamic)

# ─── 6.  Генератор ROM ───────────────────────────────────────────────────────
def generate_batch(n: int) -> np.ndarray:
    payload = np.random.randint(0, 256, size=(n, PAYLOAD_BYTES), dtype=np.uint8)
    reset   = np.tile(np.array(RESET_VECTOR, dtype=np.uint8), (n,1))
    return np.hstack((payload, reset))

# ─── 7.  Основной цикл ───────────────────────────────────────────────────────────
def main():
    lut = create_luts()
    tot_gen = tot_int = tot_vid = tot_dyn = 0
    batch_idx = 0
    start = time.perf_counter()

    try:
        while True:
            roms_cpu = generate_batch(BATCH_SIZE); tot_gen += BATCH_SIZE

            roms_gpu = cp.asarray(roms_cpu)
            keep, _ = analyse_batch(roms_gpu, lut)
            keep = keep.get(); del roms_gpu
            interesting = roms_cpu[keep]
            tot_int += len(interesting)

            for rom in interesting:
                has_vid, is_dyn = rom_video_flags(rom.tobytes())

                # Сохраняем, если ЛЮБОЕ из условий истинно
                if has_vid or is_dyn:
                    sha = hashlib.sha1(rom).hexdigest()[:12]
                    (OUTPUT_DIR / f"{sha}.bin").write_bytes(rom.tobytes())

                # Отдельное управление
                if has_vid:
                    tot_vid += 1           # первый кадр не чёрный
                if is_dyn:
                    tot_dyn += 1           # обнаружена анимация

            batch_idx += 1
            if batch_idx % STATUS_EVERY == 0:
                elapsed = time.perf_counter() - start
                rate = int(tot_gen / elapsed) if elapsed else 0
                print(f"{tot_gen:,d} generated | "
                      f"{tot_int:,d} interesting | {tot_vid:,d} with video | "
                      f"{tot_dyn:,d} dynamic | {rate:,d} ROM/s", flush=True)

    except KeyboardInterrupt:
        pass  # беспроблемный выход

    elapsed = time.perf_counter() - start
    rate = int(tot_gen / elapsed) if elapsed else 0
    print("─"*72)
    print(f"TOTAL: {tot_gen:,d} generated | {tot_int:,d} interesting | "
          f"{tot_vid:,d} with video | {tot_dyn:,d} dynamic | {rate:,d} ROM/s")

if __name__ == "__main__":
    main()