Add RTMLib CUDA backends and staged timing

servers/unraid/tail2-pose-server/app/tail2/offline_benchmark.py

@@ -26,22 +26,28 @@ def run_offline_benchmark(models, threshold: float, max_people: int) -> None:
         one = cv2.imdecode(encoded, cv2.IMREAD_COLOR)
         if one is None:
             raise RuntimeError("official demo image decode failed")
-        images = {1: one, 2: np.concatenate([one, one], axis=1),
+        images = {0: np.zeros_like(one), 1: one, 2: np.concatenate([one, one], axis=1),
                   4: np.concatenate([np.concatenate([one, one], axis=1)] * 2, axis=0)}
-        for _ in range(50):
+        for _ in range(100):
             models.infer(one, threshold, max_people)
-        report = {"source_url": DEMO_URL, "warmup_iterations": 50, "measured_iterations": 60,
+        report = {"source_url": DEMO_URL, "warmup_iterations": 100, "measured_iterations": 300,
                   "image_retained": False, "profiles": {}}
         for requested, image in images.items():
-            samples = []; counts = []
-            for _ in range(60):
+            samples = []; counts = []; stages = {}
+            for _ in range(300):
                 started = time.perf_counter_ns()
-                people, det_ms, pose_ms = models.infer(image, threshold, max_people)
+                people, timing = models.infer(image, threshold, max_people)
                 samples.append((time.perf_counter_ns() - started) / 1e6); counts.append(len(people))
+                for name, value in timing.items():
+                    stages.setdefault(name, []).append(value)
             report["profiles"][str(requested)] = {
                 "requested_people": requested, "detected_people_min": min(counts), "detected_people_max": max(counts),
                 "p50_ms": round(percentile(samples, .50), 3), "p95_ms": round(percentile(samples, .95), 3),
-                "p99_ms": round(percentile(samples, .99), 3), "fps_from_p50": round(1000 / percentile(samples, .50), 3)}
+                "p99_ms": round(percentile(samples, .99), 3), "fps_from_p50": round(1000 / percentile(samples, .50), 3),
+                "stages": {name: {"p50_ms": round(percentile(vals, .50), 3),
+                                  "p95_ms": round(percentile(vals, .95), 3),
+                                  "p99_ms": round(percentile(vals, .99), 3)}
+                           for name, vals in stages.items()}}
         with open(path, "x", encoding="utf-8") as fh:
             json.dump(report, fh, indent=2)
     except Exception as exc: