GPU delegate with OpenGL backend hangs after first frame on Snapdragon QCS8250 / Adreno 650

[BohdanPetrykiv]

Component: LiteRT GPU Delegate / OpenGL backend
Version: LiteRT v2.1.4

Device Information

SoC: Qualcomm Snapdragon QCS8250
GPU: Adreno 650
Android version: 13
LiteRT version: 2.1.4
Build type: Android ARM64, NDK r27c

When using the LiteRT GPU delegate with OpenGL backend, inference runs correctly on the first frame but then hangs indefinitely on subsequent frames look like it runs always on first frame.
With the OpenCL backend (LITERT_CL) on the same device and same model, inference runs correctly on all frames without hanging.

[shivvaam0001]

Hi @BohdanPetrykiv

Thank you for reporting this issue. We have noted the details, could you please share some information regarding the model and execution context to help us reproduce and investigate the hanging behavior?

1- Could you provide a minimal reproducible code snippet showing how the GPU delegate is initialized and invoked across multiple frames?
2- Could you please share the adb logcat output captured at the exact moment the hang occurs? We will look for driver level timeouts or traces.

[BohdanPetrykiv]

Hi @shivvaam0001

Thank you for quick response.

1. cmake -B build_android -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_HOME/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-26 -DCMAKE_BUILD_TYPE=Release
   cmake --build . -j

litert-gpu-inference.zip

Run: LD_LIBRARY_PATH=/data/local/tmp ./infer --video maxim_1080p.mp4 --output maxim_out_litert.mp4 --model mediapipe_face-facedetector-float_.tflite --device gpu_gl --thres 0.5 --inference-type litert

05-07 00:27:04.924 23038 23038 I infer : type=1400 audit(0.0:8984): avc: denied { map } for path="/data/local/tmp/infer" dev="dm-51" ino=20193 scontext=u:r:dropbear:s0 tcontext=u:object_r:shell_data_file:s0 tclass=file permissive=1
05-07 00:27:04.928 23038 23038 I infer : type=1400 audit(0.0:8985): avc: denied { search } for name="tests" dev="dm-51" ino=112 scontext=u:r:dropbear:s0 tcontext=u:object_r:shell_test_data_file:s0 tclass=dir permissive=1
05-07 00:27:04.928 23038 23038 I infer : type=1400 audit(0.0:8986): avc: denied { dac_read_search } for capability=2 scontext=u:r:dropbear:s0 tcontext=u:r:dropbear:s0 tclass=capability permissive=1
05-07 00:27:05.003 23038 23038 D libEGL : loaded /vendor/lib64/egl/libEGL_adreno.so
05-07 00:27:05.004 23038 23038 D libEGL : loaded /vendor/lib64/egl/libGLESv1_CM_adreno.so
05-07 00:27:05.006 23038 23038 D libEGL : loaded /vendor/lib64/egl/libGLESv2_adreno.so
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: QUALCOMM build : a4f4f78fac, Ic4d50c01ff
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Build Date : 11/27/24
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: OpenGL ES Shader Compiler Version: EV031.36.08.30
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Local Branch :
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Remote Branch : refs/tags/AU_LINUX_ANDROID_LA.VENDOR.1.0.11.00.00.770.073
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Remote Branch : NONE
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Reconstruct Branch : NOTHING
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Build Config : S P 12.1.1 AArch64
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Driver Path : /vendor/lib64/egl/libGLESv2_adreno.so
05-07 00:27:05.022 23038 23038 I AdrenoGLES-0: Driver Version : 0615.93
05-07 00:27:05.025 23038 23038 I AdrenoGLES-0: PFP: 0x016dd112, ME: 0x00000000
05-07 00:27:05.030 23038 23038 E libEGL : call to OpenGL ES API with no current context (logged once per thread)
05-07 00:27:05.030 23038 23038 I litert : [environment.cc:30] Creating LiteRT environment with options
05-07 00:27:05.030 23038 23038 W litert : [auto_registration.cc:78] NPU accelerator could not be loaded and registered: kLiteRtStatusErrorInvalidArgument.
05-07 00:27:05.030 23038 23038 I litert : [gpu_registry.cc:162] Loading GPU accelerator(/data/local/tmp/libLiteRtGpuAccelerator.so).
05-07 00:27:05.031 23038 23038 I litert : [gpu_registry.cc:162] Loading GPU accelerator(/data/local/tmp/libLiteRtClGlAccelerator.so).
05-07 00:27:05.033 23038 23038 I litert : [accelerator_registry.cc:54] RegisterAccelerator: ptr=0xb400007453b2d180, name=LiteRT GPU
05-07 00:27:05.033 23038 23038 I litert : [gpu_registry.cc:170] Dynamically loaded GPU accelerator(libLiteRtClGlAccelerator.so) registered.
05-07 00:27:05.033 23038 23038 I tflite : Loaded OpenCL library with dlopen.
05-07 00:27:05.037 23038 23038 I litert : [gpu_environment.cc:237] Created default OpenCL device.
05-07 00:27:05.038 23038 23038 I litert : [gpu_environment.cc:328] Created default OpenGL-OpenCL shared context.
05-07 00:27:05.038 23038 23038 I litert : [gpu_environment.cc:361] Created default OpenCL command queue.
05-07 00:27:05.039 23038 23038 I litert : [gpu_environment.h:152] Created LiteRT GpuEnvironment.
05-07 00:27:05.039 23038 23038 I litert : [environment.cc:41] Adding options to the existing LiteRT environment
05-07 00:27:05.039 23038 23038 I litert : [compiled_model.cc:719] Flatbuffer model initialized directly from incoming litert model.
05-07 00:27:05.039 23038 23038 I tflite : Initialized TensorFlow Lite runtime.
05-07 00:27:05.040 23038 23038 I litert : [environment.cc:41] Adding options to the existing LiteRT environment
05-07 00:27:05.041 23038 23038 E tflite : Following operations are not supported by GPU delegate:
05-07 00:27:05.041 23038 23038 E tflite : CUSTOM TFLite_Detection_PostProcess: TFLite_Detection_PostProcess
05-07 00:27:05.041 23038 23038 E tflite : 147 operations will run on the GPU, and the remaining 1 operations will run on the CPU.
05-07 00:27:05.041 23038 23038 I tflite : Replacing 147 out of 148 node(s) with delegate (LITERT_OPENGL) node, yielding 2 partitions for subgraph 0.
05-07 00:27:05.043 23038 23038 I native : I0000 00:00:1778102825.043493 23038 delegate_kernel.cc:721] Initializing OpenGL-based API from graph.

[shivvaam0001]

Hi @BohdanPetrykiv,

Thank you for the detailed report and for providing the exact CMake commands, NDK version, and logcat output.
Based on the logs you provided, we have spend some time and identified the following two log entries:

1- E libEGL : call to OpenGL ES API with no current context (logged once per thread)
2- l tflite : Replacing 147 out of 148 node(s) with delegate (LITERT_OPENGL) node, yielding 2 partitions for subgraph 0.

We will investigate this further.
Thanks!

[BohdanPetrykiv]

Hi @shivvaam0001

Thank you!
If you need my assistance, you are welcome