Fix GPU encoding black frames and improve debug logging

- Add CUDA sync before encoding to ensure RGB->NV12 kernel completes
- Add debug logging for frame data validation (sum check)
- Handle GPUFrame objects in GPUHLSOutput.write()
- Fix cv2.resize for CuPy arrays (use cupyx.scipy.ndimage.zoom)
- Fix fused pipeline parameter ordering (geometric first, color second)
- Add raindrop-style ripple with random position/freq/decay/amp
- Generate final VOD playlist with #EXT-X-ENDLIST

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

streaming/sexp_to_cuda.py

@@ -11,6 +11,9 @@ import numpy as np
 from typing import Dict, List, Any, Optional, Tuple
 import hashlib
 import sys
+import logging
+
+logger = logging.getLogger(__name__)
 
 # Kernel cache
 _COMPILED_KERNELS: Dict[str, Any] = {}
@@ -72,6 +75,13 @@ def compile_frame_pipeline(effects: List[dict], width: int, height: int) -> call
 def _generate_fused_kernel(effects: List[dict], width: int, height: int) -> str:
     """Generate CUDA kernel code for fused effects pipeline."""
 
+    # Validate all ops are supported
+    SUPPORTED_OPS = {'rotate', 'zoom', 'ripple', 'invert', 'hue_shift', 'brightness'}
+    for effect in effects:
+        op = effect.get('op')
+        if op not in SUPPORTED_OPS:
+            raise ValueError(f"Unsupported CUDA kernel operation: '{op}'. Supported ops: {', '.join(sorted(SUPPORTED_OPS))}. Note: 'resize' must be handled separately before the fused kernel.")
+
     # Build the kernel
     code = r'''
 extern "C" __global__
@@ -129,7 +139,7 @@ void fused_pipeline(
 '''
         elif op == 'ripple':
             code += f'''
-    // Ripple {i}
+    // Ripple {i} - matching original formula: sin(dist/freq - phase) * exp(-dist*decay/maxdim)
     {{
         float amplitude = params[param_idx++];
         float frequency = params[param_idx++];
@@ -141,9 +151,11 @@ void fused_pipeline(
         float rdx = src_x - rcx;
         float rdy = src_y - rcy;
         float dist = sqrtf(rdx * rdx + rdy * rdy);
+        float max_dim = (float)(width > height ? width : height);
 
-        float wave = sinf(dist * frequency * 0.1f + phase);
-        float amp = amplitude * expf(-dist * decay * 0.01f);
+        // Original formula: sin(dist / frequency - phase) * exp(-dist * decay / max_dim)
+        float wave = sinf(dist / frequency - phase);
+        float amp = amplitude * expf(-dist * decay / max_dim);
 
         if (dist > 0.001f) {{
             ripple_dx += rdx / dist * wave * amp;
@@ -288,10 +300,25 @@ void fused_pipeline(
     return code
 
 
+_BUILD_PARAMS_COUNT = 0
+
 def _build_params(effects: List[dict], dynamic_params: dict) -> cp.ndarray:
-    """Build parameter array for kernel."""
+    """Build parameter array for kernel.
+
+    IMPORTANT: Parameters must be built in the same order the kernel consumes them:
+    1. First all geometric transforms (rotate, zoom, ripple) in list order
+    2. Then all color transforms (invert, hue_shift, brightness) in list order
+    """
+    global _BUILD_PARAMS_COUNT
+    _BUILD_PARAMS_COUNT += 1
+
+    # ALWAYS log first few calls - use WARNING to ensure visibility in Celery logs
+    if _BUILD_PARAMS_COUNT <= 3:
+        logger.warning(f"[BUILD_PARAMS #{_BUILD_PARAMS_COUNT}] effects={[e['op'] for e in effects]}")
+
     params = []
 
+    # First pass: geometric transforms (matches kernel's first loop)
     for effect in effects:
         op = effect['op']
 
@@ -300,16 +327,30 @@ def _build_params(effects: List[dict], dynamic_params: dict) -> cp.ndarray:
         elif op == 'zoom':
             params.append(float(dynamic_params.get('zoom_amount', effect.get('amount', 1.0))))
         elif op == 'ripple':
-            params.append(float(dynamic_params.get('ripple_amplitude', effect.get('amplitude', 10))))
-            params.append(float(effect.get('frequency', 8)))
-            params.append(float(effect.get('decay', 2)))
-            params.append(float(dynamic_params.get('ripple_phase', effect.get('phase', 0))))
-            params.append(float(effect.get('center_x', 960)))
-            params.append(float(effect.get('center_y', 540)))
-        elif op == 'invert':
-            params.append(float(effect.get('amount', 0)))
+            amp = float(dynamic_params.get('ripple_amplitude', effect.get('amplitude', 10)))
+            freq = float(effect.get('frequency', 8))
+            decay = float(effect.get('decay', 2))
+            phase = float(dynamic_params.get('ripple_phase', effect.get('phase', 0)))
+            cx = float(effect.get('center_x', 960))
+            cy = float(effect.get('center_y', 540))
+            params.extend([amp, freq, decay, phase, cx, cy])
+            if _BUILD_PARAMS_COUNT <= 10 or _BUILD_PARAMS_COUNT % 500 == 0:
+                logger.warning(f"[BUILD_PARAMS #{_BUILD_PARAMS_COUNT}] ripple amp={amp} freq={freq} decay={decay} phase={phase:.2f} cx={cx} cy={cy}")
+
+    # Second pass: color transforms (matches kernel's second loop)
+    for effect in effects:
+        op = effect['op']
+
+        if op == 'invert':
+            amt = float(effect.get('amount', 0))
+            params.append(amt)
+            if _BUILD_PARAMS_COUNT <= 10 or _BUILD_PARAMS_COUNT % 500 == 0:
+                logger.warning(f"[BUILD_PARAMS #{_BUILD_PARAMS_COUNT}] invert amount={amt}")
         elif op == 'hue_shift':
-            params.append(float(effect.get('degrees', 0)))
+            deg = float(effect.get('degrees', 0))
+            params.append(deg)
+            if _BUILD_PARAMS_COUNT <= 10 or _BUILD_PARAMS_COUNT % 500 == 0:
+                logger.warning(f"[BUILD_PARAMS #{_BUILD_PARAMS_COUNT}] hue_shift degrees={deg}")
         elif op == 'brightness':
             params.append(float(effect.get('factor', 1.0)))