This repository has no description
perf: parallel pose+object detection + CPU+XNNPACK default on Android
On Samsung SM-A366B the GPU delegate offloaded only 26/685 YOLO26n ops;
the CPU↔GPU roundtrip tax dominated, and in BOTH mode the GPU also contended
with the pose pipeline. Flipping the default to CPU+XNNPACK (setUseXNNPACK(true)
is required on tensorflow-lite-support 0.5.0, otherwise pure CPU is ~100× slower)
raised BOTH-mode FPS 8.8 → 9.9.
Removing the even/odd alternation in CameraView.android.kt activated the
already-existing `poseExecutor` parallel path in Utils.android.kt, so in BOTH
mode each camera frame now runs both detectors concurrently instead of every
other frame. Net: effective per-detector rate doubled.
numThreads dropped 4 → 3 for the object interpreter so concurrent pose + object
don't oversubscribe the 8-core CPU.
Runtime override preserved: `adb shell setprop debug.tflite.delegate GPU|NNAPI|CPU`.
Combined with tonight's INT8 baseline re-export, on-device BOTH-mode FPS went
8.77 → 16.02 (+83%) with effective object-detection rate rising from ~4.4 to ~16 FPS.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
author
virtualintern
co-author
Claude Opus 4.7 (1M context)
date
(Apr 18, 2026, 8:38 PM +0300)
commit
a405c0c0
a405c0c0af40f1a5bf981a3d05c3268090c4b452
parent
ad700278
ad700278cfc5cf6c57d65848854281fc0c5681d7
+59
-26
+12
-8
posedetection/src/androidMain/kotlin/com.performancecoachlab/posedetection/camera/CameraView.android.kt
+12
-8
posedetection/src/androidMain/kotlin/com.performancecoachlab/posedetection/camera/CameraView.android.kt
···
305
305
val now = sensorMs + bootEpochOffset
306
306
val area = focus
307
307
308
-
// In BOTH mode, stagger pose and object on alternate frames
309
-
// to reduce per-frame processing time and improve tracking.
310
-
val frameNum = analysisFrameCounter.incrementAndGet()
311
-
val isBothMode = currentDetectMode == DetectMode.BOTH
308
+
// In BOTH mode, run both detectors on every eligible frame.
309
+
// The underlying `ImageProxy.process()` already fans pose out
310
+
// to `poseExecutor` and runs object inference on the analyzer
311
+
// thread concurrently, so per-frame wall-clock is max(pose,
312
+
// object) instead of pose+object. The previous even/odd
313
+
// alternation cut each detector's effective frame rate in
314
+
// half for ~20% combined-FPS gain — ditched in favor of
315
+
// doubling per-detector coverage, which matters more for
316
+
// ball-near-hoop detection.
317
+
analysisFrameCounter.incrementAndGet()
312
318
val shouldRunObject = currentDetectMode.doObject() &&
313
-
(now - lastObjectRunAtMs.get() >= objectIntervalMs) &&
314
-
(!isBothMode || frameNum % 2 == 0L)
319
+
(now - lastObjectRunAtMs.get() >= objectIntervalMs)
315
320
val shouldRunPose = currentDetectMode.doPose() &&
316
-
(now - lastPoseRunAtMs.get() >= poseIntervalMs) &&
317
-
(!isBothMode || frameNum % 2 != 0L)
321
+
(now - lastPoseRunAtMs.get() >= poseIntervalMs)
318
322
319
323
// If neither detector is scheduled to run, just close quickly and reuse last results.
320
324
if (!shouldRunObject && !shouldRunPose) {
+47
-18
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/custom/CustomObjectModel.android.kt
+47
-18
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/custom/CustomObjectModel.android.kt
···
22
22
if (modelPath.androidModelPath == null) {
23
23
throw IllegalArgumentException("Android model path cannot be null")
24
24
}
25
-
// Prefer GPU, then NNAPI (API 27+), then CPU. `selectedDelegate` tracks
26
-
// which one actually ends up in the final interpreter so we can log it.
25
+
// Default delegate is CPU+XNNPACK on Android. Measured on Samsung SM-A366B:
26
+
// CPU+XNNPACK at 12.3 FPS object-only and 9.9 FPS both-mode BEAT GPU
27
+
// (12.1 and 8.8 FPS respectively) because the GPU delegate only offloads
28
+
// 26/685 YOLO26n ops and the CPU↔GPU roundtrip overhead dominates — plus
29
+
// GPU contends with the pose pipeline in BOTH mode. See the delegate
30
+
// section of collab/experiment_log.md for the full test.
31
+
//
32
+
// Override for experiments via `adb shell setprop debug.tflite.delegate
33
+
// GPU|NNAPI|CPU`. Chosen path is logged.
34
+
val requestedDelegate = runCatching {
35
+
val cls = Class.forName("android.os.SystemProperties")
36
+
val get = cls.getMethod("get", String::class.java, String::class.java)
37
+
(get.invoke(null, "debug.tflite.delegate", "CPU") as? String).orEmpty()
38
+
.uppercase().ifBlank { "CPU" }
39
+
}.getOrElse { "CPU" }
40
+
Logger.i { "TFLite: requestedDelegate=$requestedDelegate (debug.tflite.delegate)" }
41
+
27
42
var selectedDelegate = "CPU"
28
-
val (options, gpuDelegate) = runCatching {
29
-
val delegate = GpuDelegate()
30
-
val opts = Interpreter.Options().apply {
31
-
addDelegate(delegate)
32
-
setNumThreads(2)
43
+
val (options, gpuDelegate) = when (requestedDelegate) {
44
+
"CPU" -> {
45
+
selectedDelegate = "CPU"
46
+
// numThreads=3 (not 4) leaves ~3 cores for concurrent pose detection
47
+
// in BOTH mode. With numThreads=4 and pose also multithreaded, the
48
+
// 8-core Samsung oversubscribes and thermal-throttles within ~30s.
49
+
Logger.i { "TFLite: forced CPU (no delegate, numThreads=3, XNNPACK on)" }
50
+
Interpreter.Options().apply {
51
+
setNumThreads(3)
52
+
setUseXNNPACK(true)
53
+
} to null
33
54
}
34
-
selectedDelegate = "GPU"
35
-
Logger.i { "TFLite: GPU delegate constructed" }
36
-
opts to delegate
37
-
}.onFailure { t ->
38
-
Logger.w(t) { "TFLite: GPU delegate not available; trying NNAPI" }
39
-
}.getOrElse {
40
-
if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O_MR1) {
55
+
"NNAPI" -> if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O_MR1) {
41
56
runCatching {
42
57
val nnapiDelegate = org.tensorflow.lite.nnapi.NnApiDelegate()
43
58
val opts = Interpreter.Options().apply {
···
45
60
setNumThreads(2)
46
61
}
47
62
selectedDelegate = "NNAPI"
48
-
Logger.i { "TFLite: NNAPI delegate constructed" }
63
+
Logger.i { "TFLite: NNAPI delegate constructed (forced)" }
49
64
opts to null
50
-
}.onFailure { t ->
51
-
Logger.w(t) { "TFLite: NNAPI delegate not available; falling back to CPU" }
52
-
}.getOrElse {
65
+
}.getOrElse { t ->
66
+
Logger.w(t) { "TFLite: NNAPI forced but unavailable; falling back to CPU" }
53
67
selectedDelegate = "CPU"
54
68
Interpreter.Options().apply { setNumThreads(4) } to null
55
69
}
56
70
} else {
71
+
selectedDelegate = "CPU"
72
+
Logger.w { "TFLite: NNAPI requested but SDK too old; CPU" }
73
+
Interpreter.Options().apply { setNumThreads(4) } to null
74
+
}
75
+
else -> runCatching { // GPU (default)
76
+
val delegate = GpuDelegate()
77
+
val opts = Interpreter.Options().apply {
78
+
addDelegate(delegate)
79
+
setNumThreads(2)
80
+
}
81
+
selectedDelegate = "GPU"
82
+
Logger.i { "TFLite: GPU delegate constructed" }
83
+
opts to delegate
84
+
}.getOrElse { t ->
85
+
Logger.w(t) { "TFLite: GPU unavailable; CPU" }
57
86
selectedDelegate = "CPU"
58
87
Interpreter.Options().apply { setNumThreads(4) } to null
59
88
}