This repository has no description
perf: add NNAPI delegate fallback for TFLite object detection
Add NNAPI delegate as an intermediate fallback between GPU and CPU
for TFLite inference. On devices where GPU delegate is unavailable,
NNAPI can route inference to the device's DSP/NPU for better
performance than CPU-only execution.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
author
virtualintern
co-author
Claude Opus 4.6 (1M context)
committer
![]()
tangled.org
date
(Apr 1, 2026, 8:15 PM UTC)
commit
b42adef6
b42adef6d539490ea03b0a813f7b9ab78ff3adaf
parent
83f417b0
83f417b006edd8e5ed50b2297c77efd00c2dda8e
+20
-4
+20
-4
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/custom/CustomObjectModel.android.kt
+20
-4
posedetection/src/androidMain/kotlin/com/performancecoachlab/posedetection/custom/CustomObjectModel.android.kt
···
4
4
import androidx.compose.ui.platform.LocalContext
5
5
import co.touchlab.kermit.Logger
6
6
import org.json.JSONObject
7
+
import android.os.Build
7
8
import org.tensorflow.lite.Interpreter
8
9
import org.tensorflow.lite.gpu.GpuDelegate
9
10
import org.tensorflow.lite.support.common.FileUtil
···
21
22
if (modelPath.androidModelPath == null) {
22
23
throw IllegalArgumentException("Android model path cannot be null")
23
24
}
24
-
// Prefer GPU if available, fall back to CPU.
25
+
// Prefer GPU, then NNAPI (API 27+), then CPU.
25
26
val (options, gpuDelegate) = runCatching {
26
27
val delegate = GpuDelegate()
27
28
val opts = Interpreter.Options().apply {
28
29
addDelegate(delegate)
29
-
// Threads often don’t matter with GPU delegate; keep small to reduce contention.
30
30
setNumThreads(2)
31
31
}
32
32
Logger.d { "TFLite GPU delegate available" }
33
33
opts to delegate
34
34
}.onFailure { t ->
35
-
Logger.w(t) { "TFLite GPU delegate not available; falling back to CPU" }
35
+
Logger.w(t) { "TFLite GPU delegate not available; trying NNAPI" }
36
36
}.getOrElse {
37
-
Interpreter.Options().apply { setNumThreads(4) } to null
37
+
if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O_MR1) {
38
+
runCatching {
39
+
val nnapiDelegate = org.tensorflow.lite.nnapi.NnApiDelegate()
40
+
val opts = Interpreter.Options().apply {
41
+
addDelegate(nnapiDelegate)
42
+
setNumThreads(2)
43
+
}
44
+
Logger.d { "TFLite NNAPI delegate available" }
45
+
opts to null
46
+
}.onFailure { t ->
47
+
Logger.w(t) { "TFLite NNAPI delegate not available; falling back to CPU" }
48
+
}.getOrElse {
49
+
Interpreter.Options().apply { setNumThreads(4) } to null
50
+
}
51
+
} else {
52
+
Interpreter.Options().apply { setNumThreads(4) } to null
53
+
}
38
54
}
39
55
40
56
val model = FileUtil.loadMappedFile(LocalContext.current, modelPath.androidModelPath)