tflite Namespace Reference

Functions
TfLiteDelegate *	tflite_plugin_create_delegate (char **options_keys, char **options_values, size_t num_options, void(*report_error)(const char *))
	Create an ArmNN delegate plugin. More...
void	tflite_plugin_destroy_delegate (TfLiteDelegate *delegate)
	Destroy a given delegate plugin. More...

Variables
std::vector< std::string >	gpu_options
	This file defines two symbols that need to be exported to use the TFLite external delegate provider. More...

Function Documentation

tflite_plugin_create_delegate()

TfLiteDelegate* tflite::tflite_plugin_create_delegate	(	char **	options_keys,
		char **	options_values,
		size_t	num_options,
		void(*)(const char *)	report_error
	)

Create an ArmNN delegate plugin.

Available options:

Option key: "backends"
Possible values: ["EthosNPU"/"GpuAcc"/"CpuAcc"/"CpuRef"]
Descriptions: A comma separated list without whitespaces of backends which should be used for execution. Falls back to next backend in list if previous doesn't provide support for operation. e.g. "GpuAcc,CpuAcc"

Option key: "logging-severity"
Possible values: ["trace"/"debug"/"info"/"warning"/"error"/"fatal"]
Description: Sets the logging severity level for ArmNN. Logging is turned off if this option is not provided.

Option key: "gpu-tuning-level"
Possible values: ["0"/"1"/"2"/"3"]
Description: 0=UseOnly(default), 1=RapidTuning, 2=NormalTuning, 3=ExhaustiveTuning. Requires option gpu-tuning-file. 1,2 and 3 will create a tuning-file, 0 will apply the tunings from an existing file

Option key: "gpu-tuning-file"
Possible values: [filenameString]
Description: File name for the tuning file.

Option key: "gpu-kernel-profiling-enabled"
Possible values: ["true"/"false"]
Description: Enables GPU kernel profiling

Option key: "reduce-fp32-to-fp16"
Possible values: ["true"/"false"]
Description: Reduce Fp32 data to Fp16 for faster processing

Option key: "reduce-fp32-to-bf16"
Possible values: ["true"/"false"]
Description: Reduce Fp32 data to Bf16 for faster processing

Option key: "debug-data"
Possible values: ["true"/"false"]
Description: Add debug data for easier troubleshooting

Option key: "memory-import"
Possible values: ["true"/"false"]
Description: Enable memory import

Parameters

[in]	option_keys	Delegate option names
[in]	options_values	Delegate option values
[in]	num_options	Number of delegate options
[in,out]	report_error	Error callback function

Returns

An ArmNN delegate if it succeeds else NULL

Definition at line 89 of file armnn_external_delegate.cpp.

References DelegateOptions::AddBackendOption(), OptimizerOptions::m_Debug, OptimizerOptions::m_ImportEnabled, OptimizerOptions::m_ReduceFp32ToBf16, OptimizerOptions::m_ReduceFp32ToFp16, DelegateOptions::SetBackends(), DelegateOptions::SetLoggingSeverity(), DelegateOptions::SetOptimizerOptions(), armnnDelegate::TfLiteArmnnDelegateCreate(), and armnnDelegate::TfLiteArmnnDelegateOptionsDefault().

{
    // Returning null indicates an error during delegate creation so we initialize with that
    TfLiteDelegate* delegate = nullptr;
    try
    {
        // (Initializes with CpuRef backend)
        armnnDelegate::DelegateOptions options = armnnDelegate::TfLiteArmnnDelegateOptionsDefault();
        armnn::OptimizerOptions optimizerOptions;
        for (size_t i = 0; i < num_options; ++i)
        {
            // Process backends
            if (std::string(options_keys[i]) == std::string("backends"))
            {
                // The backend option is a comma separated string of backendIDs that needs to be split
                std::vector<armnn::BackendId> backends;
                char* pch;
                pch = strtok(options_values[i],",");
                while (pch != NULL)
                {
                    backends.push_back(pch);
                    pch = strtok (NULL, ",");
                }
                options.SetBackends(backends);
            }
            // Process logging level
            else if (std::string(options_keys[i]) == std::string("logging-severity"))
            {
                options.SetLoggingSeverity(options_values[i]);
            }
            // Process GPU backend options
            else if (std::string(options_keys[i]) == std::string("gpu-tuning-level"))
            {
                armnn::BackendOptions option("GpuAcc", {{"TuningLevel", atoi(options_values[i])}});
                options.AddBackendOption(option);
            }
            else if (std::string(options_keys[i]) == std::string("gpu-mlgo-tuning-file"))
            {
                armnn::BackendOptions option("GpuAcc", {{"MLGOTuningFilePath", std::string(options_values[i])}});
                options.AddBackendOption(option);
            }
            else if (std::string(options_keys[i]) == std::string("gpu-tuning-file"))
            {
                armnn::BackendOptions option("GpuAcc", {{"TuningFile", std::string(options_values[i])}});
                options.AddBackendOption(option);
            }
            else if (std::string(options_keys[i]) == std::string("gpu-kernel-profiling-enabled"))
            {
                armnn::BackendOptions option("GpuAcc", {{"KernelProfilingEnabled", (*options_values[i] != '0')}});
                options.AddBackendOption(option);
            }
            // Process reduce-fp32-to-fp16 option
            else if (std::string(options_keys[i]) == std::string("reduce-fp32-to-fp16"))
            {
               optimizerOptions.m_ReduceFp32ToFp16 = *options_values[i] != '0';
            }
            // Process reduce-fp32-to-bf16 option
            else if (std::string(options_keys[i]) == std::string("reduce-fp32-to-bf16"))
            {
               optimizerOptions.m_ReduceFp32ToBf16 = *options_values[i] != '0';
            }
            // Process debug-data
            else if (std::string(options_keys[i]) == std::string("debug-data"))
            {
               optimizerOptions.m_Debug = *options_values[i] != '0';
            }
            // Process memory-import
            else if (std::string(options_keys[i]) == std::string("memory-import"))
            {
               optimizerOptions.m_ImportEnabled = *options_values[i] != '0';
            }
            else
            {
                throw armnn::Exception("Unknown option for the ArmNN Delegate given: " + std::string(options_keys[i]));
            }
        }
        options.SetOptimizerOptions(optimizerOptions);
        delegate = TfLiteArmnnDelegateCreate(options);
    }
    catch (const std::exception& ex)
    {
        if(report_error)
        {
            report_error(ex.what());
        }
    }
    return delegate;
}

tflite_plugin_destroy_delegate()

void tflite::tflite_plugin_destroy_delegate

(

TfLiteDelegate *

delegate

)

Destroy a given delegate plugin.

Parameters

[in]

delegate

Delegate to destruct

Definition at line 185 of file armnn_external_delegate.cpp.

References armnnDelegate::TfLiteArmnnDelegateDelete().

{
    armnnDelegate::TfLiteArmnnDelegateDelete(delegate);
}

Variable Documentation

gpu_options

std::vector<std::string> gpu_options

Initial value:

{"gpu-tuning-level",
                                      "gpu-tuning-file",
                                      "gpu-kernel-profiling-enabled"}

This file defines two symbols that need to be exported to use the TFLite external delegate provider.

This is a plugin that can be used for fast integration of delegates into benchmark tests and other tools. It allows loading of a dynamic delegate library at runtime.

The external delegate also has Tensorflow Lite Python bindings. Therefore the dynamic external delegate can be directly used with Tensorflow Lite Python APIs.

See tensorflow/lite/delegates/external for details or visit the tensorflow guide here

Definition at line 28 of file armnn_external_delegate.cpp.