TensorFlow Lite on GPU

TensorFlow Lite supports several hardware accelerators. This document describes how to use the GPU backend using the TensorFlow Lite delegate APIs on Android (requires OpenCL or OpenGL ES 3.1 and higher) and iOS (requires iOS 8 or later).

Benefits of GPU acceleration

Speed

GPUs are designed to have high throughput for massively parallelizable workloads. Thus, they are well-suited for deep neural nets, which consist of a huge number of operators, each working on some input tensor(s) that can be easily divided into smaller workloads and carried out in parallel. This parallelism typically results in lower latency. In the best scenario, inference on the GPU may run fast enough to become suitable for real-time applications that were not previously possible.

Accuracy

GPUs do their computation with 16-bit or 32-bit floating point numbers and (unlike the CPUs) do not require quantization for optimal performance. If decreased accuracy made quantization untenable for your models, running your neural network on a GPU may eliminate this concern.

Energy efficiency

Another benefit that comes with GPU inference is its power efficiency. A GPU carries out computations in a very efficient and optimized way, consuming less power and generating less heat than the same task run on a CPU.

Supported ops

TensorFlow Lite on GPU supports the following ops in 16-bit and 32-bit float precision:

  • ADD
  • AVERAGE_POOL_2D
  • CONCATENATION
  • CONV_2D
  • DEPTHWISE_CONV_2D v1-2
  • EXP
  • FULLY_CONNECTED
  • LOGISTIC
  • LSTM v2 (Basic LSTM only)
  • MAX_POOL_2D
  • MAXIMUM
  • MINIMUM
  • MUL
  • PAD
  • PRELU
  • RELU
  • RELU6
  • RESHAPE
  • RESIZE_BILINEAR v1-3
  • SOFTMAX
  • STRIDED_SLICE
  • SUB
  • TRANSPOSE_CONV

By default, all ops are only supported at version 1. Enabling the experimental quantization support allows the appropriate versions; for example, ADD v2.

Basic usage

There are two ways to invoke model acceleration in Android depending on if you are using Android Studio ML Model Binding or TensorFlow Lite Interpreter.

Android via TensorFlow Lite Interpreter

Add the tensorflow-lite-gpu package alongside the existing tensorflow-lite package in the existing dependencies block.

dependencies {
    ...
    implementation 'org.tensorflow:tensorflow-lite:2.3.0'
    implementation 'org.tensorflow:tensorflow-lite-gpu:2.3.0'
}

Then run TensorFlow Lite on GPU with TfLiteDelegate. In Java, you can specify the GpuDelegate through Interpreter.Options.

Kotlin

    import org.tensorflow.lite.Interpreter
    import org.tensorflow.lite.gpu.CompatibilityList
    import org.tensorflow.lite.gpu.GpuDelegate

    val compatList = CompatibilityList()

    val options = Interpreter.Options().apply{
        if(compatList.isDelegateSupportedOnThisDevice){
            // if the device has a supported GPU, add the GPU delegate
            val delegateOptions = compatList.bestOptionsForThisDevice
            this.addDelegate(GpuDelegate(delegateOptions))
        } else {
            // if the GPU is not supported, run on 4 threads
            this.setNumThreads(4)
        }
    }

    val interpreter = Interpreter(model, options)

    // Run inference
    writeToInput(input)
    interpreter.run(input, output)
    readFromOutput(output)
      

Java

    import org.tensorflow.lite.Interpreter;
    import org.tensorflow.lite.gpu.CompatibilityList;
    import org.tensorflow.lite.gpu.GpuDelegate;

    // Initialize interpreter with GPU delegate
    Interpreter.Options options = new Interpreter.Options();
    CompatibilityList compatList = CompatibilityList();

    if(compatList.isDelegateSupportedOnThisDevice()){
        // if the device has a supported GPU, add the GPU delegate
        GpuDelegate.Options delegateOptions = compatList.getBestOptionsForThisDevice();
        GpuDelegate gpuDelegate = new GpuDelegate(delegateOptions);
        options.addDelegate(gpuDelegate);
    } else {
        // if the GPU is not supported, run on 4 threads
        options.setNumThreads(4);
    }

    Interpreter interpreter = new Interpreter(model, options);

    // Run inference
    writeToInput(input);
    interpreter.run(input, output);
    readFromOutput(output);
      

Android (C/C++)

For C/C++ usage of TensorFlow Lite GPU on Android, the GPU delegate can be created with TfLiteGpuDelegateV2Create() and destroyed with TfLiteGpuDelegateV2Delete().

// Set up interpreter.
auto model = FlatBufferModel::BuildFromFile(model_path);
if (!model) return false;
ops::builtin::BuiltinOpResolver op_resolver;
std::unique_ptr<Interpreter> interpreter;
InterpreterBuilder(*model, op_resolver)(&interpreter);

// NEW: Prepare GPU delegate.
auto* delegate = TfLiteGpuDelegateV2Create(/*default options=*/nullptr);
if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) return false;

// Run inference.
WriteToInputTensor(interpreter->typed_input_tensor<float>(0));
if (interpreter->Invoke() != kTfLiteOk) return false;
ReadFromOutputTensor(interpreter->typed_output_tensor<float>(0));

// NEW: Clean up.
TfLiteGpuDelegateV2Delete(delegate);

Take a look at TfLiteGpuDelegateOptionsV2 to create a delegate instance with custom options. You can initialize the default options with TfLiteGpuDelegateOptionsV2Default() and then modify them as necessary.

TFLite GPU for Android C/C++ uses the Bazel build system. The delegate can be built, for example, using the following command:

bazel build -c opt --config android_arm64 tensorflow/lite/delegates/gpu:delegate                           # for static library
bazel build -c opt --config android_arm64 tensorflow/lite/delegates/gpu:libtensorflowlite_gpu_delegate.so  # for dynamic library

iOS (Swift)

Initialize TensorFlow Lite interpreter with the GPU delegate.

import TensorFlowLite

let delegate = MetalDelegate()
if let interpreter = try Interpreter(modelPath: modelPath,
                                     delegates: [delegate]) {

  // Run inference ...
}

iOS (Objective-C)

To use TensorFlow Lite on GPU, get the GPU delegate via TFLGpuDelegateCreate() and then pass it to Interpreter::ModifyGraphWithDelegate() (instead of calling Interpreter::AllocateTensors()).

// Set up interpreter.
auto model = FlatBufferModel::BuildFromFile(model_path);
if (!model) return false;
tflite::ops::builtin::BuiltinOpResolver op_resolver;
std::unique_ptr<Interpreter> interpreter;
InterpreterBuilder(*model, op_resolver)(&interpreter);

// NEW: Prepare GPU delegate.

auto* delegate = TFLGpuDelegateCreate(/*default options=*/nullptr);
if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) return false;

// Run inference.
WriteToInputTensor(interpreter->typed_input_tensor<float>(0));
if (interpreter->Invoke() != kTfLiteOk) return false;
ReadFromOutputTensor(interpreter->typed_output_tensor<float>(0));

// Clean up.
TFLGpuDelegateDelete(delegate);

Advanced usage

Delegate Options for iOS

TFLGpuDelegateCreate() accepts a struct of options. (C API, Swift API)

Passing nullptr(C API) or nothing (Swift API) to the initializer sets the default options (which are explicated in the Basic Usage example above).

Swift API


// THIS:
var options = MetalDelegate.Options()
options.isPrecisionLossAllowed = false
options.waitType = .passive
options.isQuantizationEnabled = false
let delegate = MetalDelegate(options: options)

// IS THE SAME AS THIS:
let delegate = MetalDelegate()

C API (also used for Objective-C)


// THIS:
const TFLGpuDelegateOptions options = {
  .allow_precision_loss = false,
  .wait_type = TFLGpuDelegateWaitType::TFLGpuDelegateWaitTypePassive,
  .enable_quantization = false,
};

auto* delegate = TFLGpuDelegateCreate(options);

// IS THE SAME AS THIS:
auto* delegate = TFLGpuDelegateCreate(nullptr);

While it is convenient to use nullptr, we recommend that you explicitly set the options, to avoid any unexpected behavior if default values are changed in the future.

Running quantized models on GPU

This section explains how the GPU delegate accelerates 8-bit quantized models. This includes all flavors of quantization, including:

To optimize performance, use models that have floating-point input & output tensors.

How does this work?

Since the GPU backend only supports floating-point execution, we run quantized models by giving it a ‘floating-point view’ of the original model. At a high-level, this entails the following steps:

  • Constant tensors (such as weights/biases) are dequantized once into the GPU memory. This happens when the delegate is applied to the TFLite Interpreter.

  • Inputs and outputs to the GPU program, if 8-bit quantized, are dequantized and quantized (respectively) for each inference. This is done on the CPU using TFLite’s optimized kernels.

  • The GPU program is modified to mimic quantized behavior by inserting quantization simulators between operations. This is necessary for models where ops expect activations to follow bounds learnt during quantization.

This feature can be enabled using delegate options as follows:

Android

Android APIs support quantized models by default. To disable, do the following:

C++ API

TfLiteGpuDelegateOptionsV2 options = TfLiteGpuDelegateOptionsV2Default();
options.experimental_flags = TFLITE_GPU_EXPERIMENTAL_FLAGS_NONE;

auto* delegate = TfLiteGpuDelegateV2Create(options);
if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) return false;

Java API

GpuDelegate delegate = new GpuDelegate(new GpuDelegate.Options().setQuantizedModelsAllowed(false));

Interpreter.Options options = (new Interpreter.Options()).addDelegate(delegate);

iOS

iOD APIs support quantized models by default. To disable, do the following:

Swift API

var options = MetalDelegate.Options()
options.isQuantizationEnabled = false
let delegate = MetalDelegate(options: options)

C API (also used for Objective-C)

TFLGpuDelegateOptions options = TFLGpuDelegateOptionsDefault();
options.enable_quantization = false;

auto* delegate = TFLGpuDelegateCreate(options);

Input/Output Buffers (iOS only)

To do computation on the GPU, data must be made available to the GPU. This often requires performing a memory copy. It is desirable not to cross the CPU/GPU memory boundary if possible, as this can take up a significant amount of time. Usually, such crossing is inevitable, but in some special cases, one or the other can be omitted.

If the network's input is an image already loaded in the GPU memory (for example, a GPU texture containing the camera feed) it can stay in the GPU memory without ever entering the CPU memory. Similarly, if the network's output is in the form of a renderable image (for example, image style transfer_) it can be directly displayed on the screen.

To achieve best performance, TensorFlow Lite makes it possible for users to directly read from and write to the TensorFlow hardware buffer and bypass avoidable memory copies.

Assuming the image input is in GPU memory, it must first be converted to a MTLBuffer object for Metal. You can associate a TfLiteTensor to a user-prepared MTLBuffer with TFLGpuDelegateBindMetalBufferToTensor(). Note that TFLGpuDelegateBindMetalBufferToTensor() must be called after Interpreter::ModifyGraphWithDelegate(). Additionally, the inference output is, by default, copied from GPU memory to CPU memory. This behavior can be turned off by calling Interpreter::SetAllowBufferHandleOutput(true) during initialization.

#include "tensorflow/lite/delegates/gpu/metal_delegate.h"
#include "tensorflow/lite/delegates/gpu/metal_delegate_internal.h"

// ...

// Prepare GPU delegate.
auto* delegate = TFLGpuDelegateCreate(nullptr);

if (interpreter->ModifyGraphWithDelegate(delegate) != kTfLiteOk) return false;

interpreter->SetAllowBufferHandleOutput(true);  // disable default gpu->cpu copy
if (!TFLGpuDelegateBindMetalBufferToTensor(
        delegate, interpreter->inputs()[0], user_provided_input_buffer)) {
  return false;
}
if (!TFLGpuDelegateBindMetalBufferToTensor(
        delegate, interpreter->outputs()[0], user_provided_output_buffer)) {
  return false;
}

// Run inference.
if (interpreter->Invoke() != kTfLiteOk) return false;

Tips and Tricks

  • Some operations that are trivial on the CPU may be high cost on a GPU. One class of such operation includes various forms of reshape operations (including BATCH_TO_SPACE, SPACE_TO_BATCH, SPACE_TO_DEPTH, and similar operation). If these operations are not required (for example, they were inserted to help the network architect reason about the system but do not otherwise affect output), it is worth removing them for performance.

  • On a GPU, tensor data is sliced into 4-channels. Thus, a computation on a tensor of shape [B, H, W, 5] will perform about the same on a tensor of shape [B, H, W, 8], but significantly worse than [B, H, W, 4].

    • For example, if the camera hardware supports image frames in RGBA, feeding that 4-channel input is significantly faster, because a memory copy (from 3-channel RGB to 4-channel RGBX) can be avoided.
  • For best performance, do not hesitate to re-train your classifier with mobile-optimized network architecture. That is a significant part of optimization for on-device inference.