matrice-streaming 0.1.60__py3-none-any.whl → 0.1.62__py3-none-any.whl

matrice_streaming/streaming_gateway/camera_streamer/nvdec.py

@@ -0,0 +1,961 @@
+#!/usr/bin/env python3
+"""Streaming Gateway - CUDA IPC Video Producer (NVDEC Hardware Decode).
+
+This module implements the producer side of the zero-copy video pipeline
+using NVDEC hardware video decoding for maximum throughput.
+
+Architecture:
+=============
+
+    ┌─────────────────────────────────────────────────────────────────────────┐
+    │                    STREAMING GATEWAY (Producer)                         │
+    ├─────────────────────────────────────────────────────────────────────────┤
+    │                                                                         │
+    │   ┌─────────────────────────────────────────────────────────────────┐   │
+    │   │                   NVDEC Decoder Pool                            │   │
+    │   │                                                                 │   │
+    │   │  ┌────────────────┐  ┌────────────────┐  ┌────────────────┐     │   │
+    │   │  │  Decoder 0     │  │  Decoder 1     │  │  Decoder N     │     │   │
+    │   │  │                │  │                │  │                │     │   │
+    │   │  │   NVDEC HW     │  │   NVDEC HW     │  │   NVDEC HW     │     │   │
+    │   │  │   decode       │  │   decode       │  │   decode       │     │   │
+    │   │  │       ↓        │  │       ↓        │  │       ↓        │     │   │
+    │   │  │  NV12 Resize   │  │  NV12 Resize   │  │  NV12 Resize   │     │   │
+    │   │  │       ↓        │  │       ↓        │  │       ↓        │     │   │
+    │   │  │   CUDA IPC     │  │   CUDA IPC     │  │   CUDA IPC     │     │   │
+    │   │  │   Ring Buf     │  │   Ring Buf     │  │   Ring Buf     │     │   │
+    │   │  │  (NV12 0.6MB)  │  │  (NV12 0.6MB)  │  (NV12 0.6MB)     │     │   │
+    │   │  └────────────────┘  └────────────────┘  └────────────────┘     │   │
+    │   │                                                                 │   │
+    │   └─────────────────────────────────────────────────────────────────┘   │
+    │                               │                                         │
+    │                    Output: NV12 (H*1.5, W) uint8 = 0.6 MB               │
+    │                    50% less IPC bandwidth than RGB                      │
+    │                               ↓                                         │
+    └───────────────────────────────┼─────────────────────────────────────────┘
+                                    │
+                         Consumer reads via CUDA IPC
+                         → NV12→RGB→CHW→FP16 in one kernel
+                         → TensorRT inference
+
+Usage:
+======
+    python streaming_gateway.py --video videoplayback.mp4 --num-streams 100
+
+Requirements:
+=============
+    - PyNvVideoCodec for NVDEC hardware decode
+    - CuPy with CUDA support
+    - cuda_shm_ring_buffer module
+"""
+
+import argparse
+import logging
+import multiprocessing as mp
+import os
+import time
+import threading
+import queue as thread_queue
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Any
+
+import numpy as np
+
+try:
+    import cupy as cp
+    CUPY_AVAILABLE = True
+except ImportError:
+    CUPY_AVAILABLE = False
+    cp = None
+
+try:
+    import PyNvVideoCodec as nvc
+    PYNVCODEC_AVAILABLE = True
+except ImportError:
+    PYNVCODEC_AVAILABLE = False
+    nvc = None
+
+try:
+    from matrice_common.stream.cuda_shm_ring_buffer import CudaIpcRingBuffer, GlobalFrameCounter
+    RING_BUFFER_AVAILABLE = True
+except ImportError:
+    RING_BUFFER_AVAILABLE = False
+
+logger = logging.getLogger(__name__)
+
+def setup_logging(quiet: bool = True):
+    """Configure logging level based on quiet mode."""
+    level = logging.WARNING if quiet else logging.INFO
+    logging.basicConfig(
+        level=level,
+        format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
+    )
+    logging.getLogger('cuda_shm_ring_buffer').setLevel(logging.WARNING if quiet else logging.INFO)
+
+
+@dataclass
+class StreamConfig:
+    """Configuration for a single video stream."""
+    camera_id: str
+    video_path: str
+    width: int = 640
+    height: int = 640
+    target_fps: int = 10
+    gpu_id: int = 0
+
+
+@dataclass
+class GatewayConfig:
+    """Configuration for the streaming gateway."""
+    video_path: str
+    num_streams: int = 100
+    target_fps: int = 0  # 0 = unlimited, >0 = FPS limit per stream
+    frame_width: int = 640
+    frame_height: int = 640
+    gpu_id: int = 0
+    num_gpus: int = 1
+    duration_sec: float = 30.0
+    nvdec_pool_size: int = 8
+    nvdec_burst_size: int = 4
+    num_slots: int = 32
+
+
+@dataclass
+class StreamState:
+    """Track state for each logical stream in NVDEC pool."""
+    stream_id: int
+    camera_id: str
+    video_path: str
+    demuxer: Any
+    frames_decoded: int = 0
+    width: int = 640
+    height: int = 640
+    empty_packets: int = 0
+
+
+# =============================================================================
+# CUDA Kernel: NV12 Resize (no color conversion - 50% less bandwidth)
+# =============================================================================
+
+_nv12_resize_kernel = None
+
+
+def _get_nv12_resize_kernel():
+    """Get or compile the NV12 resize kernel.
+
+    This kernel resizes NV12 directly (no color conversion).
+    Output: concatenated Y (H×W) + UV ((H/2)×W) = H×W×1.5 bytes
+    This is 50% smaller than RGB (H×W×3 bytes).
+
+    Consumer will do: NV12→RGB→CHW→FP16 in one fused kernel.
+    """
+    global _nv12_resize_kernel
+    if _nv12_resize_kernel is None and CUPY_AVAILABLE:
+        _nv12_resize_kernel = cp.RawKernel(r'''
+extern "C" __global__ void nv12_resize(
+    const unsigned char* src_y,    // Source Y plane
+    const unsigned char* src_uv,   // Source UV plane (interleaved)
+    unsigned char* dst,            // Output: Y (H×W) followed by UV ((H/2)×W)
+    int src_h, int src_w,
+    int dst_h, int dst_w,
+    int y_stride, int uv_stride
+) {
+    int dst_x = blockIdx.x * blockDim.x + threadIdx.x;
+    int dst_y = blockIdx.y * blockDim.y + threadIdx.y;
+
+    // Total height in output: dst_h (Y) + dst_h/2 (UV) = dst_h * 1.5
+    int total_h = dst_h + dst_h / 2;
+    if (dst_x >= dst_w || dst_y >= total_h) return;
+
+    float scale_x = (float)src_w / dst_w;
+    float scale_y = (float)src_h / dst_h;
+
+    if (dst_y < dst_h) {
+        // Y plane region: resize Y
+        int src_x = min((int)(dst_x * scale_x), src_w - 1);
+        int src_y_coord = min((int)(dst_y * scale_y), src_h - 1);
+        int src_idx = src_y_coord * y_stride + src_x;
+        int dst_idx = dst_y * dst_w + dst_x;
+        dst[dst_idx] = src_y[src_idx];
+    } else {
+        // UV plane region: resize UV (UV is at half vertical resolution)
+        int uv_dst_y = dst_y - dst_h;  // 0 to dst_h/2-1
+        int uv_src_y = min((int)(uv_dst_y * scale_y), src_h / 2 - 1);
+
+        // UV is interleaved, so we copy pairs (U, V) together
+        int src_uv_x = min((int)((dst_x / 2) * 2 * scale_x), src_w - 2);
+        src_uv_x = (src_uv_x / 2) * 2;  // Ensure even
+
+        int src_idx = uv_src_y * uv_stride + src_uv_x + (dst_x % 2);
+        int dst_idx = dst_h * dst_w + uv_dst_y * dst_w + dst_x;
+        dst[dst_idx] = src_uv[src_idx];
+    }
+}
+''', 'nv12_resize')
+    return _nv12_resize_kernel
+
+
+def nv12_resize(y_plane: cp.ndarray, uv_plane: cp.ndarray,
+                y_stride: int, uv_stride: int,
+                src_h: int, src_w: int,
+                dst_h: int = 640, dst_w: int = 640) -> cp.ndarray:
+    """Resize NV12 without color conversion.
+
+    Output: concatenated Y (H×W) + UV ((H/2)×W) as single buffer.
+    Total size: H×W + (H/2)×W = H×W×1.5 bytes (50% of RGB).
+    """
+    kernel = _get_nv12_resize_kernel()
+    if kernel is None:
+        return None
+
+    total_h = dst_h + dst_h // 2
+    output = cp.empty((total_h, dst_w), dtype=cp.uint8)
+
+    block = (16, 16)
+    grid = ((dst_w + 15) // 16, (total_h + 15) // 16)
+
+    kernel(grid, block, (
+        y_plane, uv_plane, output,
+        cp.int32(src_h), cp.int32(src_w),
+        cp.int32(dst_h), cp.int32(dst_w),
+        cp.int32(y_stride), cp.int32(uv_stride)
+    ))
+
+    return output
+
+
+def surface_to_nv12(frame, target_h: int = 640, target_w: int = 640) -> Optional[cp.ndarray]:
+    """Convert NVDEC surface to resized NV12 (50% smaller than RGB).
+
+    Output: (H + H/2, W) uint8 - concatenated Y + UV planes.
+    Total size: H×W×1.5 bytes (vs H×W×3 for RGB).
+    """
+    if not CUPY_AVAILABLE or frame is None:
+        return None
+
+    try:
+        cuda_views = frame.cuda()
+        if not cuda_views or len(cuda_views) < 2:
+            return None
+
+        # Extract Y plane
+        y_view = cuda_views[0]
+        y_cai = y_view.__cuda_array_interface__
+        y_shape = tuple(y_cai['shape'])
+        y_strides = tuple(y_cai['strides'])
+        y_ptr = y_cai['data'][0]
+        src_h, src_w = y_shape[:2]
+        y_stride = y_strides[0]
+
+        y_size = src_h * y_stride
+        y_mem = cp.cuda.UnownedMemory(y_ptr, y_size, owner=frame)
+        y_memptr = cp.cuda.MemoryPointer(y_mem, 0)
+        y_plane = cp.ndarray((src_h, src_w), dtype=cp.uint8, memptr=y_memptr,
+                            strides=(y_stride, 1))
+
+        # Extract UV plane
+        uv_view = cuda_views[1]
+        uv_cai = uv_view.__cuda_array_interface__
+        uv_shape = tuple(uv_cai['shape'])
+        uv_strides = tuple(uv_cai['strides'])
+        uv_ptr = uv_cai['data'][0]
+        uv_stride = uv_strides[0]
+
+        uv_h = uv_shape[0]
+        uv_w = uv_shape[1] if len(uv_shape) > 1 else src_w
+        uv_size = uv_h * uv_stride
+        uv_mem = cp.cuda.UnownedMemory(uv_ptr, uv_size, owner=frame)
+        uv_memptr = cp.cuda.MemoryPointer(uv_mem, 0)
+        uv_plane = cp.ndarray((uv_h, uv_w), dtype=cp.uint8, memptr=uv_memptr,
+                             strides=(uv_stride, 1))
+
+        # NV12 resize (no color conversion - 50% smaller output!)
+        nv12_frame = nv12_resize(y_plane, uv_plane, y_stride, uv_stride,
+                                  src_h, src_w, target_h, target_w)
+        # Add channel dimension for ring buffer compatibility: (H*1.5, W) -> (H*1.5, W, 1)
+        return nv12_frame[:, :, cp.newaxis] if nv12_frame is not None else None
+
+    except Exception as e:
+        logger.warning(f"surface_to_nv12 failed: {e}")
+        return None
+
+
+# =============================================================================
+# NVDEC Decoder Pool
+# =============================================================================
+
+class NVDECDecoderPool:
+    """Pool of NVDEC decoders that time-multiplex streams.
+
+    Each decoder is exclusively owned by one worker thread.
+    Outputs NV12: 1.5×H×W bytes (50% smaller than RGB).
+    """
+
+    def __init__(self, pool_size: int, gpu_id: int = 0):
+        self.pool_size = pool_size
+        self.gpu_id = gpu_id
+        self.decoders = []
+        self.streams_per_decoder: List[List[StreamState]] = [[] for _ in range(pool_size)]
+
+        if not PYNVCODEC_AVAILABLE:
+            raise RuntimeError("PyNvVideoCodec not available")
+
+        if CUPY_AVAILABLE:
+            cp.cuda.Device(gpu_id).use()
+
+        for i in range(pool_size):
+            try:
+                decoder = nvc.CreateDecoder(
+                    gpuid=gpu_id,
+                    codec=nvc.cudaVideoCodec.H264,
+                    usedevicememory=True
+                )
+                self.decoders.append(decoder)
+            except Exception as e:
+                logger.warning(f"Failed to create decoder {i}: {e}")
+                break
+
+        self.actual_pool_size = len(self.decoders)
+        logger.info(f"Created NVDEC pool: {self.actual_pool_size}/{pool_size} decoders on GPU {gpu_id}")
+
+    def assign_stream(self, stream_id: int, camera_id: str, video_path: str,
+                      width: int = 640, height: int = 640) -> bool:
+        """Assign a stream to a decoder (round-robin)."""
+        if self.actual_pool_size == 0:
+            return False
+
+        decoder_idx = stream_id % self.actual_pool_size
+        try:
+            demuxer = nvc.CreateDemuxer(video_path)
+        except Exception as e:
+            logger.error(f"Failed to create demuxer for {camera_id}: {e}")
+            return False
+
+        stream_state = StreamState(
+            stream_id=stream_id,
+            camera_id=camera_id,
+            video_path=video_path,
+            demuxer=demuxer,
+            width=width,
+            height=height
+        )
+        self.streams_per_decoder[decoder_idx].append(stream_state)
+        return True
+
+    def decode_round(self, decoder_idx: int, frames_per_stream: int = 4,
+                     target_h: int = 640, target_w: int = 640) -> Tuple[int, List[Tuple[str, cp.ndarray]]]:
+        """Decode frames and convert to NV12.
+
+        Returns:
+            (total_frames, [(camera_id, nv12_tensor), ...])
+        """
+        if decoder_idx >= self.actual_pool_size:
+            return 0, []
+
+        decoder = self.decoders[decoder_idx]
+        streams = self.streams_per_decoder[decoder_idx]
+        total_frames = 0
+        decoded_frames = []
+
+        for stream in streams:
+            frames_this_stream = 0
+
+            while frames_this_stream < frames_per_stream:
+                try:
+                    packet = stream.demuxer.Demux()
+                    if packet is None:
+                        stream.demuxer = nvc.CreateDemuxer(stream.video_path)
+                        stream.empty_packets = 0
+                        packet = stream.demuxer.Demux()
+                        if packet is None:
+                            break
+
+                    frames_before = frames_this_stream
+                    for surface in decoder.Decode(packet):
+                        tensor = surface_to_nv12(surface, target_h, target_w)
+
+                        if tensor is not None:
+                            decoded_frames.append((stream.camera_id, tensor))
+                            frames_this_stream += 1
+                            stream.frames_decoded += 1
+                            total_frames += 1
+                            stream.empty_packets = 0
+
+                        if frames_this_stream >= frames_per_stream:
+                            break
+
+                    if frames_this_stream == frames_before:
+                        stream.empty_packets += 1
+                        if stream.empty_packets >= 3:
+                            stream.demuxer = nvc.CreateDemuxer(stream.video_path)
+                            stream.empty_packets = 0
+
+                except Exception:
+                    break
+
+                if frames_this_stream >= frames_per_stream:
+                    break
+
+        return total_frames, decoded_frames
+
+    def get_camera_ids_for_decoder(self, decoder_idx: int) -> List[str]:
+        """Get camera IDs for a decoder."""
+        if decoder_idx >= self.actual_pool_size:
+            return []
+        return [s.camera_id for s in self.streams_per_decoder[decoder_idx]]
+
+    def close(self):
+        """Close all decoders."""
+        self.decoders.clear()
+        for streams in self.streams_per_decoder:
+            streams.clear()
+
+
+# =============================================================================
+# Worker Thread
+# =============================================================================
+
+def nvdec_pool_worker(
+    worker_id: int,
+    decoder_idx: int,
+    pool: NVDECDecoderPool,
+    ring_buffers: Dict[str, CudaIpcRingBuffer],
+    frame_counter: GlobalFrameCounter,
+    duration_sec: float,
+    result_queue: thread_queue.Queue,
+    stop_event: threading.Event,
+    burst_size: int = 4,
+    target_h: int = 640,
+    target_w: int = 640,
+    target_fps: int = 0,
+    shared_frame_count: Optional[mp.Value] = None,
+):
+    """NVDEC worker thread.
+
+    Decodes frames and writes NV12 tensors to ring buffers.
+    Uses dedicated CUDA stream per worker for kernel overlap.
+    Supports FPS limiting when target_fps > 0.
+    """
+    if CUPY_AVAILABLE:
+        cp.cuda.Device(pool.gpu_id).use()
+        cuda_stream = cp.cuda.Stream(non_blocking=True)
+    else:
+        cuda_stream = None
+
+    local_frames = 0
+    local_errors = 0
+    frames_since_counter_update = 0
+    counter_batch_size = 100
+    start_time = time.perf_counter()
+    last_log_time = start_time
+    camera_ids = pool.get_camera_ids_for_decoder(decoder_idx)
+    num_streams = len(camera_ids)
+
+    # FPS limiting: calculate frames per second target for this worker
+    # Each worker handles num_streams cameras at target_fps each
+    fps_limit_enabled = target_fps > 0 and num_streams > 0
+    if fps_limit_enabled:
+        # Total target frames per second for all streams handled by this worker
+        worker_target_fps = target_fps * num_streams
+        frame_interval = 1.0 / worker_target_fps
+        next_frame_time = start_time
+        fps_mode = f", FPS limit={target_fps}/stream"
+    else:
+        frame_interval = 0
+        next_frame_time = 0
+        fps_mode = ", unlimited FPS"
+
+    logger.info(f"Worker {worker_id}: decoder={decoder_idx}, cams={num_streams}{fps_mode}")
+
+    while not stop_event.is_set():
+        if time.perf_counter() - start_time >= duration_sec:
+            break
+
+        now = time.perf_counter()
+        if now - last_log_time >= 5.0:
+            elapsed = now - start_time
+            fps = local_frames / elapsed if elapsed > 0 else 0
+            logger.info(f"Worker {worker_id}: {local_frames} frames, {fps:.0f} FPS")
+            last_log_time = now
+
+        # FPS limiting: wait until next scheduled frame time
+        if fps_limit_enabled:
+            current_time = time.perf_counter()
+            if current_time < next_frame_time:
+                sleep_time = next_frame_time - current_time
+                if sleep_time > 0.0001:  # Only sleep if > 100us
+                    time.sleep(sleep_time)
+
+        try:
+            with cuda_stream:
+                num_frames, decoded_frames = pool.decode_round(
+                    decoder_idx,
+                    frames_per_stream=burst_size,
+                    target_h=target_h,
+                    target_w=target_w
+                )
+
+                for cam_id, tensor in decoded_frames:
+                    if cam_id in ring_buffers:
+                        try:
+                            ring_buffers[cam_id].write_frame_fast(tensor, sync=False)
+                            local_frames += 1
+                            frames_since_counter_update += 1
+
+                            # Update shared counter for real-time progress
+                            if shared_frame_count is not None:
+                                with shared_frame_count.get_lock():
+                                    shared_frame_count.value += 1
+
+                            # Update next frame time for FPS limiting
+                            if fps_limit_enabled:
+                                next_frame_time += frame_interval
+
+                        except Exception as e:
+                            local_errors += 1
+                            if local_errors <= 3:
+                                logger.error(f"Worker {worker_id} write error: {e}")
+
+                if decoded_frames and len(ring_buffers) > 0:
+                    next(iter(ring_buffers.values())).sync_writes()
+
+            if num_frames == 0:
+                time.sleep(0.0001)
+                continue
+
+            if frames_since_counter_update >= counter_batch_size:
+                frame_counter.increment()
+                frames_since_counter_update = 0
+
+        except Exception as e:
+            local_errors += 1
+            if local_errors <= 3:
+                logger.error(f"Worker {worker_id} error: {e}")
+
+    if frames_since_counter_update > 0:
+        frame_counter.increment()
+
+    elapsed = time.perf_counter() - start_time
+    result_queue.put({
+        "worker_id": worker_id,
+        "decoder_idx": decoder_idx,
+        "elapsed_sec": elapsed,
+        "total_frames": local_frames,
+        "total_errors": local_errors,
+        "num_streams": len(camera_ids),
+        "fps": local_frames / elapsed if elapsed > 0 else 0,
+    })
+
+
+# =============================================================================
+# GPU Process
+# =============================================================================
+
+def nvdec_pool_process(
+    process_id: int,
+    camera_configs: List[StreamConfig],
+    pool_size: int,
+    duration_sec: float,
+    result_queue: mp.Queue,
+    stop_event: mp.Event,
+    burst_size: int = 4,
+    num_slots: int = 32,
+    target_fps: int = 0,
+    shared_frame_count: Optional[mp.Value] = None,
+):
+    """NVDEC process for one GPU.
+
+    Creates NV12 ring buffers: (H*1.5, W) = 0.6 MB/frame.
+    """
+    if not camera_configs:
+        return
+
+    gpu_id = camera_configs[0].gpu_id
+    target_h = camera_configs[0].height
+    target_w = camera_configs[0].width
+
+    if CUPY_AVAILABLE:
+        cp.cuda.Device(gpu_id).use()
+
+    # Initialize global frame counter
+    frame_counter = GlobalFrameCounter(is_producer=True)
+    if process_id == 0:
+        frame_counter.initialize()
+        logger.info(f"Process {process_id}: GlobalFrameCounter initialized")
+    else:
+        max_retries = 50
+        for retry in range(max_retries):
+            try:
+                if os.path.exists("/dev/shm/global_frame_counter"):
+                    frame_counter.connect()
+                    logger.info(f"Process {process_id}: Connected to GlobalFrameCounter")
+                    break
+            except Exception:
+                if retry == max_retries - 1:
+                    raise
+            time.sleep(0.1)
+        else:
+            raise RuntimeError(f"Process {process_id}: GlobalFrameCounter not found")
+
+    # Create decoder pool
+    try:
+        pool = NVDECDecoderPool(pool_size, gpu_id)
+    except Exception as e:
+        logger.error(f"Process {process_id}: Failed to create decoder pool: {e}")
+        result_queue.put({
+            "process_id": process_id,
+            "error": str(e),
+            "total_frames": 0,
+            "total_errors": 1,
+        })
+        return
+
+    if pool.actual_pool_size == 0:
+        result_queue.put({
+            "process_id": process_id,
+            "error": "No decoders created",
+            "total_frames": 0,
+            "total_errors": 1,
+        })
+        return
+
+    # Create NV12 ring buffers: (H + H/2, W, 1) = 0.6 MB/frame
+    ring_buffers: Dict[str, CudaIpcRingBuffer] = {}
+    frame_size_mb = target_h * target_w * 1.5 / 1e6
+
+    try:
+        for i, config in enumerate(camera_configs):
+            rb = CudaIpcRingBuffer.create_producer(
+                config.camera_id,
+                gpu_id=config.gpu_id,
+                num_slots=num_slots,
+                width=config.width,
+                height=config.height + config.height // 2,  # H * 1.5 for NV12
+                channels=1,
+            )
+            ring_buffers[config.camera_id] = rb
+
+            pool.assign_stream(
+                stream_id=i,
+                camera_id=config.camera_id,
+                video_path=config.video_path,
+                width=config.width,
+                height=config.height
+            )
+
+        logger.info(f"Process {process_id}: {pool.actual_pool_size} decoders, "
+                   f"{len(camera_configs)} streams, NV12 ({frame_size_mb:.1f} MB/frame)")
+
+        thread_stop_event = threading.Event()
+        thread_result_queue = thread_queue.Queue()
+
+        threads = []
+        for decoder_idx in range(pool.actual_pool_size):
+            t = threading.Thread(
+                target=nvdec_pool_worker,
+                args=(
+                    process_id * 100 + decoder_idx,
+                    decoder_idx,
+                    pool,
+                    ring_buffers,
+                    frame_counter,
+                    duration_sec,
+                    thread_result_queue,
+                    thread_stop_event,
+                    burst_size,
+                    target_h,
+                    target_w,
+                    target_fps,
+                    shared_frame_count,
+                )
+            )
+            t.start()
+            threads.append(t)
+
+        start_time = time.perf_counter()
+        while not stop_event.is_set():
+            if time.perf_counter() - start_time >= duration_sec:
+                break
+            time.sleep(0.1)
+
+        thread_stop_event.set()
+
+        for t in threads:
+            t.join(timeout=30.0)
+
+        total_frames = 0
+        total_errors = 0
+        elapsed = time.perf_counter() - start_time
+
+        while not thread_result_queue.empty():
+            try:
+                r = thread_result_queue.get_nowait()
+                total_frames += r.get("total_frames", 0)
+                total_errors += r.get("total_errors", 0)
+            except:
+                break
+
+        pool.close()
+        for rb in ring_buffers.values():
+            rb.close()
+
+        result_queue.put({
+            "process_id": process_id,
+            "elapsed_sec": elapsed,
+            "total_frames": total_frames,
+            "total_errors": total_errors,
+            "num_streams": len(camera_configs),
+            "pool_size": pool.actual_pool_size,
+            "fps": total_frames / elapsed if elapsed > 0 else 0,
+            "per_stream_fps": total_frames / elapsed / len(camera_configs) if elapsed > 0 and camera_configs else 0,
+        })
+
+    except Exception as e:
+        logger.error(f"Process {process_id} error: {e}")
+        import traceback
+        traceback.print_exc()
+
+        pool.close()
+        for rb in ring_buffers.values():
+            rb.close()
+
+        result_queue.put({
+            "process_id": process_id,
+            "error": str(e),
+            "total_frames": 0,
+            "total_errors": 1,
+        })
+
+
+# =============================================================================
+# Streaming Gateway
+# =============================================================================
+
+class StreamingGateway:
+    """Multi-stream video producer outputting NV12 tensors (minimal IPC payload)."""
+
+    def __init__(self, config: GatewayConfig):
+        self.config = config
+        self._workers: List[mp.Process] = []
+        self._stop_event = mp.Event()
+        self._result_queue = mp.Queue()
+
+    def start(self) -> Dict:
+        """Start the gateway."""
+        if not CUPY_AVAILABLE:
+            raise RuntimeError("CuPy is required")
+        if not RING_BUFFER_AVAILABLE:
+            raise RuntimeError("CUDA IPC ring buffer not available")
+        if not PYNVCODEC_AVAILABLE:
+            raise RuntimeError("PyNvVideoCodec required")
+        return self._start_nvdec_pool()
+
+    def _start_nvdec_pool(self) -> Dict:
+        """Start NVDEC pool across GPUs."""
+        num_gpus = min(self.config.num_gpus, 8)
+        streams_per_gpu = self.config.num_streams // num_gpus
+        extra_streams = self.config.num_streams % num_gpus
+
+        logger.info(f"Starting NVDEC on {num_gpus} GPU(s): {self.config.num_streams} streams, "
+                   f"pool_size={self.config.nvdec_pool_size}/GPU, output=NV12 (0.6 MB)")
+
+        ctx = mp.get_context("spawn")
+        self._stop_event = ctx.Event()
+        self._result_queue = ctx.Queue()
+
+        # Shared counter for real-time FPS tracking (use 'L' for large counts)
+        shared_frame_count = ctx.Value('L', 0)
+
+        stream_idx = 0
+        for gpu_id in range(num_gpus):
+            n_streams = streams_per_gpu + (1 if gpu_id < extra_streams else 0)
+
+            gpu_configs = []
+            for i in range(n_streams):
+                config = StreamConfig(
+                    camera_id=f"cam_{stream_idx:04d}",
+                    video_path=self.config.video_path,
+                    width=self.config.frame_width,
+                    height=self.config.frame_height,
+                    target_fps=self.config.target_fps,
+                    gpu_id=gpu_id,
+                )
+                gpu_configs.append(config)
+                stream_idx += 1
+
+            p = ctx.Process(
+                target=nvdec_pool_process,
+                args=(gpu_id, gpu_configs, self.config.nvdec_pool_size,
+                      self.config.duration_sec, self._result_queue, self._stop_event,
+                      self.config.nvdec_burst_size, self.config.num_slots,
+                      self.config.target_fps, shared_frame_count)
+            )
+            p.start()
+            self._workers.append(p)
+            logger.info(f"GPU {gpu_id}: {n_streams} streams")
+            time.sleep(0.1)
+
+        # Progress monitoring loop - print progress every 5 seconds
+        start_time = time.perf_counter()
+        last_report_time = start_time
+        last_frame_count = 0
+        report_interval = 5.0  # seconds
+        processing_start_time = None  # Track when actual processing starts
+        frames_at_processing_start = 0
+
+        print(f"  [  0.0s] Started {num_gpus} GPU workers...")
+
+        while any(p.is_alive() for p in self._workers):
+            time.sleep(0.5)
+            current_time = time.perf_counter()
+
+            # Periodic progress report with real-time FPS
+            if current_time - last_report_time >= report_interval:
+                elapsed = current_time - start_time
+                remaining = max(0, self.config.duration_sec - elapsed)
+
+                # Read current frame count
+                current_frames = shared_frame_count.value
+                interval_frames = current_frames - last_frame_count
+                interval_fps = interval_frames / report_interval  # Current throughput
+                per_stream_fps = interval_fps / self.config.num_streams if self.config.num_streams > 0 else 0
+
+                # Track when processing actually starts (exclude warmup from avg)
+                if processing_start_time is None and current_frames > 0:
+                    processing_start_time = last_report_time  # Use previous report time
+                    frames_at_processing_start = last_frame_count
+
+                # Calculate average FPS excluding warmup time
+                if processing_start_time is not None:
+                    processing_elapsed = current_time - processing_start_time
+                    processing_frames = current_frames - frames_at_processing_start
+                    avg_fps = processing_frames / processing_elapsed if processing_elapsed > 0 else 0
+                    print(f"  [{elapsed:5.1f}s] {current_frames:,} frames | cur: {interval_fps:,.0f} FPS ({per_stream_fps:.1f}/stream) | avg: {avg_fps:,.0f} FPS | {remaining:.0f}s left")
+                else:
+                    print(f"  [{elapsed:5.1f}s] Warming up... | {remaining:.0f}s left")
+
+                last_report_time = current_time
+                last_frame_count = current_frames
+
+        # Wait for all workers to fully complete
+        for p in self._workers:
+            p.join(timeout=5)
+
+        results = []
+        while not self._result_queue.empty():
+            results.append(self._result_queue.get())
+
+        for r in results:
+            if "error" in r:
+                logger.error(f"NVDEC error: {r['error']}")
+
+        total_frames = sum(r.get("total_frames", 0) for r in results)
+        total_errors = sum(r.get("total_errors", 0) for r in results)
+        total_elapsed = max((r.get("elapsed_sec", 0) for r in results), default=0)
+
+        aggregate_fps = total_frames / total_elapsed if total_elapsed > 0 else 0
+        per_stream_fps = aggregate_fps / self.config.num_streams if self.config.num_streams > 0 else 0
+
+        return {
+            "num_streams": self.config.num_streams,
+            "num_gpus": num_gpus,
+            "pool_size": self.config.nvdec_pool_size,
+            "duration_sec": total_elapsed,
+            "total_frames": total_frames,
+            "total_errors": total_errors,
+            "aggregate_fps": aggregate_fps,
+            "per_stream_fps": per_stream_fps,
+            "gpu_results": results,
+        }
+
+    def stop(self):
+        """Stop all workers."""
+        self._stop_event.set()
+        for p in self._workers:
+            p.join(timeout=5)
+            if p.is_alive():
+                p.terminate()
+
+
+# =============================================================================
+# CLI
+# =============================================================================
+
+def main():
+    parser = argparse.ArgumentParser(description="Streaming Gateway - CUDA IPC Producer (NV12)")
+    parser.add_argument("--video", "-v", required=True, help="Video file path")
+    parser.add_argument("--num-streams", "-n", type=int, default=100, help="Number of streams")
+    parser.add_argument("--fps", type=int, default=0, help="Target FPS limit per stream (0=unlimited)")
+    parser.add_argument("--width", type=int, default=640, help="Frame width")
+    parser.add_argument("--height", type=int, default=640, help="Frame height")
+    parser.add_argument("--duration", "-d", type=float, default=30.0, help="Duration in seconds")
+    parser.add_argument("--gpu", type=int, default=0, help="Primary GPU ID")
+    parser.add_argument("--num-gpus", "-g", type=int, default=1, help="Number of GPUs (1-8)")
+    parser.add_argument("--pool-size", type=int, default=8, help="NVDEC pool size per GPU")
+    parser.add_argument("--burst-size", type=int, default=4, help="Frames per stream before rotating")
+    parser.add_argument("--slots", type=int, default=32, help="Ring buffer slots per camera")
+    parser.add_argument("--quiet", "-q", action="store_true", help="Quiet mode - only show final results")
+    args = parser.parse_args()
+
+    # Setup logging based on quiet mode
+    setup_logging(quiet=args.quiet)
+
+    config = GatewayConfig(
+        video_path=args.video,
+        num_streams=args.num_streams,
+        target_fps=args.fps,
+        frame_width=args.width,
+        frame_height=args.height,
+        gpu_id=args.gpu,
+        num_gpus=args.num_gpus,
+        duration_sec=args.duration,
+        nvdec_pool_size=args.pool_size,
+        nvdec_burst_size=args.burst_size,
+        num_slots=args.slots,
+    )
+
+    frame_size = args.width * args.height * 1.5
+    output_fmt = f"NV12 ({args.width}x{args.height}x1.5 = {frame_size/1e6:.1f} MB/frame)"
+    fps_limit_str = f"{args.fps} FPS/stream" if args.fps > 0 else "unlimited"
+
+    if not args.quiet:
+        print("\n" + "=" * 60)
+        print("      STREAMING GATEWAY - CUDA IPC Producer (NV12)")
+        print("=" * 60)
+        print(f"  Video:      {args.video}")
+        print(f"  Streams:    {args.num_streams}")
+        print(f"  GPUs:       {args.num_gpus}")
+        print(f"  Pool size:  {args.pool_size} NVDEC decoders/GPU")
+        print(f"  FPS limit:  {fps_limit_str}")
+        print(f"  Output:     {output_fmt}")
+        print(f"  Duration:   {args.duration}s")
+        print("=" * 60)
+
+    gateway = StreamingGateway(config)
+
+    try:
+        results = gateway.start()
+        # Clean summary output
+        print("\n")
+        print("=" * 60)
+        print("         STREAMING GATEWAY BENCHMARK RESULTS")
+        print("=" * 60)
+        print(f"  Video:        {args.video}")
+        print(f"  Streams:      {args.num_streams}")
+        print(f"  GPUs:         {args.num_gpus}")
+        print(f"  FPS limit:    {fps_limit_str}")
+        print(f"  Duration:     {args.duration}s")
+        print("-" * 60)
+        print(f"  Total Frames: {results['total_frames']:,}")
+        print("-" * 60)
+        print(f"  >>> AGGREGATE FPS: {results['aggregate_fps']:,.0f} <<<")
+        print(f"  >>> PER-STREAM FPS: {results['per_stream_fps']:.1f} <<<")
+        print("=" * 60)
+        print()
+    except KeyboardInterrupt:
+        gateway.stop()
+        print("\nStopped")
+
+
+if __name__ == "__main__":
+    main()