Skip to main content
The Android Neural Networks API (NNAPI) provides hardware-accelerated machine learning inference on Android devices, enabling efficient on-device ML model execution.

​
Overview

NNAPI is a C API that serves as a bridge between your application and hardware accelerators like GPUs, DSPs, and dedicated neural network processors.

​
Key benefits

  • Hardware acceleration - Automatically uses the best available hardware accelerator
  • Reduced latency - On-device inference eliminates network round trips
  • Privacy - Data stays on device
  • Offline support - Works without internet connectivity
NNAPI is available starting from Android 8.1 (API level 27), with significant improvements in subsequent releases.

​
Getting started

1

Add NNAPI to your build

In CMakeLists.txt:
find_library(neuralnetworks-lib neuralnetworks)
target_link_libraries(your-app ${neuralnetworks-lib})
For ndk-build, in Android.mk:
LOCAL_LDLIBS := -lneuralnetworks
2

Include the NNAPI header

#include <android/NeuralNetworks.h>
3

Check NNAPI availability

#include <android/api-level.h>

bool isNNAPIAvailable() {
    return android_get_device_api_level() >= 27;
}

​
Building a neural network model

​
Create the model

ANeuralNetworksModel* model = NULL;
ANeuralNetworksModel_create(&model);

// Add operands
ANeuralNetworksOperandType input_type = {
    .type = ANEURALNETWORKS_TENSOR_FLOAT32,
    .dimensionCount = 4,
    .dimensions = (uint32_t[]){1, 224, 224, 3},  // Batch, Height, Width, Channels
    .scale = 0.0f,
    .zeroPoint = 0
};

uint32_t input_index;
ANeuralNetworksModel_addOperand(model, &input_type);

​
Define model operations

// Example: Add a convolution operation
ANeuralNetworksOperandType filter_type = {
    .type = ANEURALNETWORKS_TENSOR_FLOAT32,
    .dimensionCount = 4,
    .dimensions = (uint32_t[]){32, 3, 3, 3},  // Output channels, H, W, Input channels
};

uint32_t filter_index;
ANeuralNetworksModel_addOperand(model, &filter_type);

// Set filter weights
float* filter_data = loadFilterWeights();
ANeuralNetworksModel_setOperandValue(model, filter_index, 
                                     filter_data, 
                                     sizeof(float) * 32 * 3 * 3 * 3);

// Add convolution operation
uint32_t inputs[] = {input_index, filter_index, /* bias, padding, stride, etc. */};
uint32_t outputs[] = {output_index};
ANeuralNetworksModel_addOperation(model, 
                                  ANEURALNETWORKS_CONV_2D,
                                  sizeof(inputs)/sizeof(uint32_t), inputs,
                                  sizeof(outputs)/sizeof(uint32_t), outputs);

​
Specify model inputs and outputs

uint32_t model_inputs[] = {input_index};
uint32_t model_outputs[] = {output_index};

ANeuralNetworksModel_identifyInputsAndOutputs(
    model,
    1, model_inputs,
    1, model_outputs
);

// Finish model construction
ANeuralNetworksModel_finish(model);
For complex models, consider using TensorFlow Lite with NNAPI delegation instead of building models manually with NNAPI.

​
Executing inference

​
Create compilation

ANeuralNetworksCompilation* compilation;
ANeuralNetworksCompilation_create(model, &compilation);

// Set preference
ANeuralNetworksCompilation_setPreference(compilation, 
    ANEURALNETWORKS_PREFER_SUSTAINED_SPEED);

// Finish compilation
ANeuralNetworksCompilation_finish(compilation);

​
Performance preferences

  • ANEURALNETWORKS_PREFER_LOW_POWER - Optimize for battery life
  • ANEURALNETWORKS_PREFER_FAST_SINGLE_ANSWER - Optimize for low latency
  • ANEURALNETWORKS_PREFER_SUSTAINED_SPEED - Balance performance and power for repeated inference

​
Run inference

ANeuralNetworksExecution* execution;
ANeuralNetworksExecution_create(compilation, &execution);

// Set input data
float* input_data = preprocessImage();
ANeuralNetworksExecution_setInput(execution, 0, NULL, 
                                  input_data, 
                                  sizeof(float) * 1 * 224 * 224 * 3);

// Set output buffer
float output_data[1000];  // For 1000 class classification
ANeuralNetworksExecution_setOutput(execution, 0, NULL,
                                   output_data,
                                   sizeof(output_data));

// Execute
ANeuralNetworksEvent* event = NULL;
int result = ANeuralNetworksExecution_startCompute(execution, &event);

if (result == ANEURALNETWORKS_NO_ERROR) {
    // Wait for completion
    ANeuralNetworksEvent_wait(event);
    ANeuralNetworksEvent_free(event);
    
    // Process output_data
    int predicted_class = argmax(output_data, 1000);
}

ANeuralNetworksExecution_free(execution);

​
Using TensorFlow Lite with NNAPI

TensorFlow Lite provides a higher-level API with NNAPI delegation:

​
Setup

In build.gradle: 
dependencies {
    implementation 'org.tensorflow:tensorflow-lite:2.13.0'
}

​
Load and run model

#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/kernels/register.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/delegates/nnapi/nnapi_delegate.h"

// Load model
std::unique_ptr<tflite::FlatBufferModel> model =
    tflite::FlatBufferModel::BuildFromFile("model.tflite");

// Build interpreter
tflite::ops::builtin::BuiltinOpResolver resolver;
tflite::InterpreterBuilder builder(*model, resolver);
std::unique_ptr<tflite::Interpreter> interpreter;
builder(&interpreter);

// Enable NNAPI delegate
auto nnapi_delegate = tflite::NnApiDelegate();
interpreter->ModifyGraphWithDelegate(&nnapi_delegate);

// Allocate tensors
interpreter->AllocateTensors();

// Set input
float* input = interpreter->typed_input_tensor<float>(0);
memcpy(input, input_data, input_size * sizeof(float));

// Run inference
interpreter->Invoke();

// Get output
float* output = interpreter->typed_output_tensor<float>(0);
TensorFlow Lite automatically falls back to CPU execution if NNAPI is unavailable or doesn’t support certain operations.

​
Hardware acceleration

​
Query available accelerators

// Get number of devices (Android 10+)
uint32_t numDevices = 0;
ANeuralNetworks_getDeviceCount(&numDevices);

for (uint32_t i = 0; i < numDevices; i++) {
    ANeuralNetworksDevice* device;
    ANeuralNetworks_getDevice(i, &device);
    
    const char* name;
    ANeuralNetworksDevice_getName(device, &name);
    
    int32_t type;
    ANeuralNetworksDevice_getType(device, &type);
    
    // Type can be:
    // ANEURALNETWORKS_DEVICE_ACCELERATOR
    // ANEURALNETWORKS_DEVICE_CPU
    // ANEURALNETWORKS_DEVICE_GPU
}

​
Specify execution device

// Compile for specific device (Android 10+)
ANeuralNetworksDevice* device;
ANeuralNetworks_getDevice(0, &device);

ANeuralNetworksCompilation_createForDevices(model, &device, 1, &compilation);

​
Optimizing models for NNAPI

​
Use supported operations

Not all operations are hardware-accelerated. Check the NNAPI operator support for your target API level.

​
Quantization

Quantize models to 8-bit integers for better performance: 
# Using TensorFlow Lite converter
import tensorflow as tf

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
tflite_quant_model = converter.convert()
Quantized models are typically 4x smaller and run 2-3x faster on NNAPI with minimal accuracy loss.

​
Input preprocessing

Perform preprocessing (normalization, resizing) efficiently: 
// Use Android's ImageReader or Bitmap APIs from Java/Kotlin
// Then pass preprocessed data to native code

void preprocessImage(uint8_t* rgba_data, int width, int height, float* output) {
    // Normalize to [-1, 1] or [0, 1] depending on model
    for (int i = 0; i < width * height; i++) {
        output[i * 3 + 0] = (rgba_data[i * 4 + 0] / 255.0f - 0.5f) * 2.0f;  // R
        output[i * 3 + 1] = (rgba_data[i * 4 + 1] / 255.0f - 0.5f) * 2.0f;  // G
        output[i * 3 + 2] = (rgba_data[i * 4 + 2] / 255.0f - 0.5f) * 2.0f;  // B
    }
}

​
Error handling

int result = ANeuralNetworksExecution_startCompute(execution, &event);

switch (result) {
    case ANEURALNETWORKS_NO_ERROR:
        // Success
        break;
    case ANEURALNETWORKS_OUT_OF_MEMORY:
        // Handle OOM
        break;
    case ANEURALNETWORKS_INCOMPLETE:
        // Model construction incomplete
        break;
    case ANEURALNETWORKS_UNEXPECTED_NULL:
        // Null pointer passed
        break;
    case ANEURALNETWORKS_BAD_DATA:
        // Invalid model data
        break;
    default:
        // Other error
        break;
}

​
Best practices

  • Reuse compilations - Compile once, execute many times for better performance
  • Batch processing - Process multiple inputs in a single inference when possible
  • Quantize models - Use 8-bit quantization for faster execution and smaller model size
  • Profile on target devices - Performance varies significantly across device accelerators
  • Provide CPU fallback - Not all operations are supported on all accelerators
  • Cache compiled models - Save compilation results to disk to reduce startup time (Android 10+)
  • Test API level support - Use feature detection, not just API level checks
Always test on a variety of devices. Some low-end devices may execute models faster on CPU than with their hardware accelerators.

​
Debugging and profiling

​
Enable verbose logging

adb shell setprop debug.nn.vlog 1
adb logcat -s NNAPI

​
Benchmark performance

#include <chrono>

auto start = std::chrono::high_resolution_clock::now();
ANeuralNetworksExecution_startCompute(execution, &event);
ANeuralNetworksEvent_wait(event);
auto end = std::chrono::high_resolution_clock::now();

float inference_time = std::chrono::duration<float, std::milli>(end - start).count();
printf("Inference time: %.2f ms\n", inference_time);

​
Additional resources

Build docs developers (and LLMs) love

Get started for freeTalk to us