# Tutorial: GPU Backend on SA-series — LV (GVM or PVM) ## Execution on LV (GVM or PVM) First, set up toolchain path: $ export QNN_AARCH64_OE_LINUX_GCC_93= Copy to clipboard When using qnn-model-lib-generator to build your model, use additional argument -t aarch64-oe-linux-gcc9.3: $ ${QNN_SDK_ROOT}/bin/x86_64-linux-clang/qnn-model-lib-generator \ -c ${QNN_SDK_ROOT}/examples/Models/InceptionV3/model/Inception_v3.cpp \ -b ${QNN_SDK_ROOT}/examples/Models/InceptionV3/model/Inception_v3.bin \ -o ${QNN_SDK_ROOT}/examples/Models/InceptionV3/model_libs \ # This can be any path -t aarch64-oe-linux-gcc9.3 Copy to clipboard This will produce the following artifacts: ${QNN_SDK_ROOT}/examples/Models/InceptionV3/model_libs/aarch64-oe-linux-gcc9.3/libInception_v3.so Copy to clipboard Now execute *adb root && adb remount* for your target, *adb shell* to the QNN terminal and create directories: $ mkdir -p /data/local/tmp/lib $ mkdir -p /data/local/tmp/bin Copy to clipboard Push the necessary libraries to device: $ adb push ${QNN_SDK_ROOT}/lib/aarch64-oe-linux-gcc9.3/libQnnGpu.so /data/local/tmp/lib $ adb push ${QNN_SDK_ROOT}/lib/aarch64-oe-linux-gcc9.3/libQnnGpuNetRunExtensions.so /data/local/tmp/lib $ adb push ${QNN_SDK_ROOT}/bin/aarch64-oe-linux-gcc9.3/qnn-net-run /data/local/tmp/bin $ adb push ${QNN_SDK_ROOT}/bin/aarch64-oe-linux-gcc9.3/qnn-throughput-net-run /data/local/tmp/bin $ adb push ${QNN_SDK_ROOT}/bin/aarch64-oe-linux-gcc9.3/qnn-profile-viewer /data/local/tmp/bin Copy to clipboard In the target terminal, enable permissions for QNN binaries: $ chmod 777 /data/local/tmp/bin/qnn-net-run $ chmod 777 /data/local/tmp/bin/qnn-throughput-net-run $ chmod 777 /data/local/tmp/bin/qnn-profile-viewer Copy to clipboard Now push the input data, input lists and model library to device: $ adb push ${QNN_SDK_ROOT}/examples/Models/InceptionV3/data/cropped /data/local/tmp/bin $ adb push ${QNN_SDK_ROOT}/examples/Models/InceptionV3/data/target_raw_list.txt /data/local/tmp/bin $ adb push ${QNN_SDK_ROOT}/examples/Models/InceptionV3/model_libs/aarch64-oe-linux-gcc9.3/libInception_v3.so /data/local/tmp/bin Copy to clipboard Finally, reset the target with the following command in the target terminal: $ reset Copy to clipboard Now qnn-net-run and qnn-profile-viewer can be executed using GPU and LV GVM or LV PVM. Then, *adb shell* into the target terminal and setup the execution environment: $ cd /data/local/tmp/bin $ export PATH=/data/local/tmp/bin:$PATH $ export VENDOR_LIB=/data/local/tmp/lib $ export LD_LIBRARY_PATH=$VENDOR_LIB:$LD_LIBRARY_PATH Copy to clipboard Then we can execute qnn-net-run with the following: $ ./qnn-net-run \ --backend libQnnGpu.so \ --input_list target_raw_list.txt \ --model libInception_v3.so Copy to clipboard We can also execute qnn-net-run with profiling: $ ./qnn-net-run \ --backend libQnnGpu.so \ --input_list target_raw_list.txt \ --model libInception_v3.so \ --profiling_level basic # OR $ ./qnn-net-run \ --backend libQnnGpu.so \ --input_list target_raw_list.txt \ --model libInception_v3.so \ --profiling_level detailed Copy to clipboard Profiling data can be viewed using qnn-profile-viewer by running the following command: $ ./qnn-profile-viewer \ --input_log output/qnn-profiling-data.log Copy to clipboard By default, qnn-net-run executes GPU backend in USER\_PROVIDED mode (i.e., no precision override). To explicitly set Precision Mode as FP32, FP16, or HYBRID, use the –config\_file option with a JSON configuration. Here is an example of json files: > > > 1. EXAMPLE of `gpu_config.json` file: > > > > { > "backend_extensions": { > "shared_library_path": "libQnnGpuNetRunExtensions.so", > "config_file_path": "./gpu_settings.json" > } > } > Copy to clipboard > > 2. EXAMPLE of `gpu_settings.json` file to set FP16 Mode: > > > > { > "graph_names": [""], > "precision_mode": "fp16" > } > Copy to clipboard > > 3. EXAMPLE of `gpu_settings.json` file to set FP32 Mode: > > > > { > "graph_names": [""], > "precision_mode": "fp32" > } > Copy to clipboard > > 4. EXAMPLE of `gpu_settings.json` file to set Hybrid Mode: > > > > { > "graph_names": [""], > "precision_mode": "hybrid" > } > Copy to clipboard Then we can execute qnn-net-run taking a JSON configuration file with the following: $ ./qnn-net-run \ --backend libQnnGpu.so \ --input_list target_raw_list.txt \ --model libInception_v3.so \ --config_file gpu_config.json Copy to clipboard We can execute qnn-throughput-net-run with the following: $ ./qnn-throughput-net-run \ --config throughput_config.json \ --output throughput_result.json Copy to clipboard Below is the sample throughput\_config.json file: > > > 1. EXAMPLE of `throughput_config.json` file: > > > > > > > > > { > > "backends": [ > > { > > "backendName": "gpu_backend", > > "backendPath": "libQnnGpu.so", > > "profilingLevel": "BASIC", > > "backendExtensions": "libQnnGpuNetRunExtensions.so", > > "perfProfile": "high_performance" > > } > > ], > > "models": [ > > { > > "modelName": "Inception_v3", > > "modelPath": "libInception_v3.so", > > "inputPath": "target_raw_list.txt", > > "inputDataType": "FLOAT" > > } > > ], > > "contexts": [ > > { > > "contextName": "gpu_context_1" > > } > > ], > > "testCase": { > > "iteration": 1, > > "logLevel": "error", > > "threads": [ > > { > > "threadName": "gpu_thread_1", > > "backend": "gpu_backend", > > "context": "gpu_context_1", > > "model": "Inception_v3", > > "interval": 0, > > "loopUnit": "count", > > "loop": 10 > > } > > ] > > } > > } > > Copy to clipboard Outputs from the run will be located at the default ./output directory. Exit the device and view the results: $ exit $ cd ${QNN_SDK_ROOT}/examples/Models/InceptionV3 $ adb pull /data/local/tmp/bin/output output_lv $ python3 ${QNN_SDK_ROOT}/examples/Models/InceptionV3/scripts/show_inceptionv3_classifications.py -i data/cropped/raw_list.txt \ -o output_lv/ \ -l data/imagenet_slim_labels.txt Copy to clipboard Last Published: Jun 04, 2026 [Previous Topic Tutorial: Running QNN GPU on SA-series](https://docs.qualcomm.com/bundle/publicresource/80-63442-10/topics/gpu_auto_tutorial_2.md) [Next Topic Tutorial: Running QNN on SA8797 QC Linux PVM (HTP and LPAI Backends)](https://docs.qualcomm.com/bundle/publicresource/80-63442-10/topics/htp_auto_qclinux.md)