A simple C++ binary to benchmark a TFLite model and its individual operators, both on desktop machines and on Android. The binary takes a TFLite model, generates random inputs and then repeatedly runs the model for specified number of runs. Aggregate latency statistics are reported after running the benchmark.
The instructions below are for running the binary on Desktop and Android, for iOS please use the iOS benchmark app.
An experimental Android APK wrapper for the benchmark model utility offers more faithful execution behavior on Android (via a foreground Activity). It is located here.
The binary takes the following required parameters:
graph:string
The path to the TFLite model file.
and the following optional parameters:
-
num_threads:int(default=-1)
The number of threads to use for running TFLite interpreter. By default, this is set to the platform default value -1. -
warmup_runs:int(default=1)
The number of warmup runs to do before starting the benchmark. -
num_runs:int(default=50)
The number of runs. Increase this to reduce variance. -
max_secs: float (default=150.0)
The maximum number of seconds the benchmark will run before being terminated. -
run_delay:float(default=-1.0)
The delay in seconds between subsequent benchmark runs. Non-positive values mean use no delay. -
run_frequency:float(default=-1.0)
The frequency of running a benchmark run as the number of prorated runs per second. If the targeted rate per second cannot be reached, the benchmark would start the next run immediately, trying its best to catch up. If set, this will override therun_delayparameter. A non-positive value means there is no delay between subsequent runs. -
enable_op_profiling:bool(default=false)
Whether to enable per-operator profiling measurement. -
max_profiling_buffer_entries:int(default=1024)
The initial max number of profiling events that will be stored during each inference run. It is only meaningful whenenable_op_profilingis set totrue. Note, the actual value of this parameter will be adjusted if the model has more nodes than the specified value of this parameter. Also, whenallow_dynamic_profiling_buffer_increaseis set totrue, the number of profiling buffer entries will be increased dynamically. -
allow_dynamic_profiling_buffer_increase:bool(default=false)
Whether allowing dynamic increase on the number of profiling buffer entries. It is only meaningful whenenable_op_profilingis set totrue. Note, allowing dynamic buffer size increase may cause more profiling overhead, thus it is preferred to setmax_profiling_buffer_entriesto a large-enough value. -
profiling_output_csv_file:str(default="")
File path to export profile data to as CSV. The results are printed tostdoutif option is not set. Requiresenable_op_profilingto betrueand the path to include the name of the output CSV; otherwise results are printed tostdout. -
print_preinvoke_state:bool(default=false)
Whether to print out the TfLite interpreter internals just before calling tflite::Interpreter::Invoke. The internals will include allocated memory size of each tensor etc. Enabling this could help understand TfLite graph and memory usage. -
print_postinvoke_state:bool(default=false)
Whether to print out the TfLite interpreter internals just before benchmark completes (i.e. after all repeated Invoke calls complete). The internals will include allocated memory size of each tensor etc. Enabling this could help understand TfLite graph and memory usage, particularly when there are dynamic-shaped tensors in the graph. -
report_peak_memory_footprint:bool(default=false)
Whether to report the peak memory footprint by periodically checking the memory footprint. Internally, a separate thread will be spawned for this periodic check. Therefore, the performance benchmark result could be affected. -
memory_footprint_check_interval_ms:int(default=50)
The interval in millisecond between two consecutive memory footprint checks. This is only used when --report_peak_memory_footprint is set to true. -
dry_run:bool(default=false)
Whether to run the tool just with simply loading the model, allocating tensors etc. but without actually invoking any op kernels. -
verbose:bool(default=false)
Whether to log parameters whose values are not set. By default, only log those parameters that are set by parsing their values from the commandline flags. -
release_dynamic_tensors:bool(default=false)
Whether to configure the Interpreter to immediately release the memory of dynamic tensors in the graph once they are not used. -
optimize_memory_for_large_tensors:int(default=0)
Whether to optimize memory usage for large tensors with sacrificing latency. When the feature is enabled,release_dynamic_tensorsis also enabled. -
enable_builtin_cast_constant_cache:bool(default=false)
Configure the builtin TFLite CAST operation to cache its output if its input is a constant tensor.WARNING: This is an experimental option that may be removed at any time.
This list of parameters is not exhaustive. See here and here for all parameters that the binary takes.
By default, the tool will use randomized data for model inputs. The following parameters allow users to specify customized input values to the model when running the benchmark tool:
input_layer:string
A comma-separated list of input layer names, e.g. 'input1,input2'. Note all inputs of the model graph need to be specified. However, the input name does not need to match that encoded in the model. Additionally, the order of input layer names specified here is assumed to be same with that is seen by the Tensorflow Lite interpreter. This is a bit inconvenient but the visualization tool should help to find this order.input_layer_shape:string
A colon-separated list of input layer shapes, where each shape is a comma-separated list, e.g. '1,30:1,10'. Similar toinput_layer, this parameter also requires shapes of all inputs be specified, and the order of inputs be same with that is seen by the interpreter.input_layer_value_range:string
A map-like string representing value range for integer input layers. Each item is separated by ':', and the item value consists of input layer name and integer-only range values (both low and high are inclusive) separated by ',', e.g. 'input1,1,2:input2,0,254'. Note that the input layer name must exist in the list of names specified byinput_layer.input_layer_value_files:string
A map-like string representing files that contain input values. Each item is separated by ',', and the item value consists of input layer name and the file path separated by ':', e.g. 'input1:file_path1,input2:file_path2'. In case the input layer name contains ':' e.g. "input:0", escape it with "::" literal, e.g.input::0:file_path1. If a input name appears in bothinput_layer_value_rangeandinput_layer_value_files, the corresponding input value range specified byinput_layer_value_rangewill be ignored. The file format is binary, and the content should be either a byte array or null-separated strings. Note that the input layer name must also exist in the list of names specified byinput_layer.
The tool supports all runtime/delegate parameters introduced by the delegate registrar. The following simply lists the names of these parameters and additional notes where applicable. For details about each parameter, please refer to this page.
max_delegated_partitions:int(default=0)min_nodes_per_partition:int(default=0)delegate_serialize_dir:str(default="")delegate_serialize_token:str(default="")
use_gpu:bool(default=false)gpu_precision_loss_allowed:bool(default=true)gpu_experimental_enable_quant:bool(default=true)gpu_inference_for_sustained_speed:bool(default=false)gpu_backend:string(default="")gpu_wait_type:str(default="")
use_nnapi:bool(default=false)
Note some Android P devices will fail to use NNAPI for models in/data/local/tmp/and this benchmark tool will not correctly use NNAPI.nnapi_execution_preference:str(default="")
Should be one of:fast_single_answer,sustained_speed,low_power,undefined.nnapi_execution_priority:str(default="")
Note this requires Android 11+.nnapi_accelerator_name:str(default="")
Note this requires Android 10+.disable_nnapi_cpu:bool(default=true)nnapi_allow_fp16:bool(default=false)nnapi_allow_dynamic_dimensions:bool(default=false)nnapi_use_burst_mode:bool(default=false)
use_hexagon:bool(default=false)hexagon_profiling:bool(default=false)
Note enabling this option will not produce profiling results outputs unlessenable_op_profilingis also turned on. When both parameters are set to true, the profile of ops on hexagon DSP will be added to the profile table. Note that, the reported data on hexagon is in cycles, not in ms like on cpu.hexagon_lib_path:string(default="/data/local/tmp/")
The library path for the underlying Hexagon libraries. This is where libhexagon_nn_skel*.so files should be. For libhexagon_interface.so it needs to be on a path that can be loaded from example: put it in LD_LIBRARY_PATH.
use_xnnpack:bool(default=false)
Note if this option is explicitly set tofalse, the TfLite runtime will use its original CPU kernels for model execution. In other words, after enabling the feature that the XNNPACK delegate is applied by default in TfLite runtime, explictly setting this flag tofalsewill cause the benchmark tool to disable the feature at runtime, and to use the original non-delegated CPU execution path for model benchmarking.xnnpack_force_fp16:bool(default=false)
Enforce float16 inference.
use_coreml:bool(default=false)coreml_version:int(default=0)
external_delegate_path:string(default="")external_delegate_options:string(default="")
stable_delegate_loader_settings:string(default="") A path to the JSON-encoded delegateTFLiteSettingsfile, which is defined inconfiguration.proto.
As some delegates are only available on certain platforms, when running the
benchmark tool on a particular platform, specifying --help will print out all
supported parameters.
When multiple delegates are specified to be used in the commandline flags, the order of delegates applied to the TfLite runtime will be same as their enabling commandline flag is specified. For example, "--use_xnnpack=true --use_gpu=true" means applying the XNNPACK delegate first, and then the GPU delegate secondly. In comparison, "--use_gpu=true --use_xnnpack=true" means applying the GPU delegate first, and then the XNNPACK delegate secondly.
(0) Refer to https://www.tensorflow.org/lite/guide/build_android to edit the
WORKSPACE to configure the android NDK/SDK.
(1) Build for your specific platform, e.g.:
bazel build -c opt \
--config=android_arm64 \
tensorflow/lite/tools/benchmark:benchmark_model
(2) Connect your phone. Push the binary to your phone with adb push (make the directory if required):
adb push bazel-bin/tensorflow/lite/tools/benchmark/benchmark_model /data/local/tmp
(3) Make the binary executable.
adb shell chmod +x /data/local/tmp/benchmark_model
(4) Push the compute graph that you need to test. For example:
adb push mobilenet_quant_v1_224.tflite /data/local/tmp
(5) Optionally, install Hexagon libraries on device.
That step is only needed when using the Hexagon delegate.
bazel build --config=android_arm64 \
tensorflow/lite/delegates/hexagon/hexagon_nn:libhexagon_interface.so
adb push bazel-bin/tensorflow/lite/delegates/hexagon/hexagon_nn/libhexagon_interface.so /data/local/tmp
adb push libhexagon_nn_skel*.so /data/local/tmp
(6) Run the benchmark. For example:
adb shell /data/local/tmp/benchmark_model \
--graph=/data/local/tmp/mobilenet_quant_v1_224.tflite \
--num_threads=4
(1) build the binary
bazel build -c opt tensorflow/lite/tools/benchmark:benchmark_model
(2) Run on your compute graph, similar to the Android case but without the need of adb shell. For example:
bazel-bin/tensorflow/lite/tools/benchmark/benchmark_model \
--graph=mobilenet_quant_v1_224.tflite \
--num_threads=4
The MobileNet graph used as an example here may be downloaded from here.
Most modern Android phones use ARM big.LITTLE
architecture where some cores are more power hungry but faster than other cores.
When running benchmarks on these phones there can be significant variance
between different runs of the benchmark. One way to reduce variance between runs
is to set the CPU affinity
before running the benchmark. On Android this can be done using the taskset
command.
E.g. for running the benchmark on big cores on Pixel 2 with a single thread one
can use the following command:
adb shell taskset f0 /data/local/tmp/benchmark_model \
--graph=/data/local/tmp/mobilenet_quant_v1_224.tflite \
--num_threads=1
where f0 is the affinity mask for big cores on Pixel 2.
Note: The affinity mask varies with the device.
The benchmark model binary also allows you to profile operators and give
execution times of each operator. To do this, pass the flag
--enable_op_profiling=true to benchmark_model during invocation, e.g.,
adb shell taskset f0 /data/local/tmp/benchmark_model \
--graph=/data/local/tmp/mobilenet_quant_v1_224.tflite \
--enable_op_profiling=true
When enabled, the benchmark_model binary will produce detailed statistics for
each operation similar to those shown below:
============================== Run Order ==============================
[node type] [start] [first] [avg ms] [%] [cdf%] [mem KB] [times called] [Name]
CONV_2D 0.000 4.269 4.269 0.107% 0.107% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_0/Relu6]
DEPTHWISE_CONV_2D 4.270 2.150 2.150 0.054% 0.161% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_1_depthwise/Relu6]
CONV_2D 6.421 6.107 6.107 0.153% 0.314% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_1_pointwise/Relu6]
DEPTHWISE_CONV_2D 12.528 1.366 1.366 0.034% 0.348% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_2_depthwise/Relu6]
CONV_2D 13.895 4.195 4.195 0.105% 0.454% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_2_pointwise/Relu6]
DEPTHWISE_CONV_2D 18.091 1.260 1.260 0.032% 0.485% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_3_depthwise/Relu6]
CONV_2D 19.352 6.652 6.652 0.167% 0.652% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_3_pointwise/Relu6]
DEPTHWISE_CONV_2D 26.005 0.698 0.698 0.018% 0.670% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_4_depthwise/Relu6]
CONV_2D 26.703 3.344 3.344 0.084% 0.754% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_4_pointwise/Relu6]
DEPTHWISE_CONV_2D 30.047 0.646 0.646 0.016% 0.770% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_5_depthwise/Relu6]
CONV_2D 30.694 5.800 5.800 0.145% 0.915% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_5_pointwise/Relu6]
DEPTHWISE_CONV_2D 36.495 0.331 0.331 0.008% 0.924% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_6_depthwise/Relu6]
CONV_2D 36.826 2.838 2.838 0.071% 0.995% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_6_pointwise/Relu6]
DEPTHWISE_CONV_2D 39.665 0.439 0.439 0.011% 1.006% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_7_depthwise/Relu6]
CONV_2D 40.105 5.293 5.293 0.133% 1.139% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_7_pointwise/Relu6]
DEPTHWISE_CONV_2D 45.399 0.352 0.352 0.009% 1.147% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_8_depthwise/Relu6]
CONV_2D 45.752 5.322 5.322 0.133% 1.281% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_8_pointwise/Relu6]
DEPTHWISE_CONV_2D 51.075 0.357 0.357 0.009% 1.290% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_9_depthwise/Relu6]
CONV_2D 51.432 5.693 5.693 0.143% 1.433% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_9_pointwise/Relu6]
DEPTHWISE_CONV_2D 57.126 0.366 0.366 0.009% 1.442% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_10_depthwise/Relu6]
CONV_2D 57.493 5.472 5.472 0.137% 1.579% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_10_pointwise/Relu6]
DEPTHWISE_CONV_2D 62.966 0.364 0.364 0.009% 1.588% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_11_depthwise/Relu6]
CONV_2D 63.330 5.404 5.404 0.136% 1.724% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_11_pointwise/Relu6]
DEPTHWISE_CONV_2D 68.735 0.155 0.155 0.004% 1.728% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_12_depthwise/Relu6]
CONV_2D 68.891 2.970 2.970 0.074% 1.802% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_12_pointwise/Relu6]
DEPTHWISE_CONV_2D 71.862 0.206 0.206 0.005% 1.807% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_13_depthwise/Relu6]
CONV_2D 72.069 5.888 5.888 0.148% 1.955% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_13_pointwise/Relu6]
AVERAGE_POOL_2D 77.958 0.036 0.036 0.001% 1.956% 0.000 0 [MobilenetV1/Logits/AvgPool_1a/AvgPool]
CONV_2D 77.994 1.445 1.445 0.036% 1.992% 0.000 0 [MobilenetV1/Logits/Conv2d_1c_1x1/BiasAdd]
RESHAPE 79.440 0.002 0.002 0.000% 1.992% 0.000 0 [MobilenetV1/Predictions/Reshape]
SOFTMAX 79.443 0.029 0.029 0.001% 1.993% 0.000 0 [MobilenetV1/Predictions/Softmax]
============================== Top by Computation Time ==============================
[node type] [start] [first] [avg ms] [%] [cdf%] [mem KB] [times called] [Name]
CONV_2D 19.352 6.652 6.652 0.167% 0.167% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_3_pointwise/Relu6]
CONV_2D 6.421 6.107 6.107 0.153% 0.320% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_1_pointwise/Relu6]
CONV_2D 72.069 5.888 5.888 0.148% 0.468% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_13_pointwise/Relu6]
CONV_2D 30.694 5.800 5.800 0.145% 0.613% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_5_pointwise/Relu6]
CONV_2D 51.432 5.693 5.693 0.143% 0.756% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_9_pointwise/Relu6]
CONV_2D 57.493 5.472 5.472 0.137% 0.893% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_10_pointwise/Relu6]
CONV_2D 63.330 5.404 5.404 0.136% 1.029% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_11_pointwise/Relu6]
CONV_2D 45.752 5.322 5.322 0.133% 1.162% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_8_pointwise/Relu6]
CONV_2D 40.105 5.293 5.293 0.133% 1.295% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_7_pointwise/Relu6]
CONV_2D 0.000 4.269 4.269 0.107% 1.402% 0.000 0 [MobilenetV1/MobilenetV1/Conv2d_0/Relu6]
Number of nodes executed: 31
============================== Summary by node type ==============================
[Node type] [count] [avg ms] [avg %] [cdf %] [mem KB] [times called]
CONV_2D 15 1.406 89.270% 89.270% 0.000 0
DEPTHWISE_CONV_2D 13 0.169 10.730% 100.000% 0.000 0
SOFTMAX 1 0.000 0.000% 100.000% 0.000 0
RESHAPE 1 0.000 0.000% 100.000% 0.000 0
AVERAGE_POOL_2D 1 0.000 0.000% 100.000% 0.000 0
Timings (microseconds): count=50 first=79449 curr=81350 min=77385 max=88213 avg=79732 std=1929
Memory (bytes): count=0
31 nodes observed
Average inference timings in us: Warmup: 83235, Init: 38467, Inference: 79760.9
A convenient and simple C++ binary is also provided to benchmark multiple
performance options in a single run. This binary is built based on the
aforementioned benchmark tool that could only benchmark a single performance
option at a time. They share the same build/install/run process, but the BUILD
target name of this binary is benchmark_model_performance_options and it takes
some additional parameters as detailed below.
perf_options_list:string(default='all')
A comma-separated list of TFLite performance options to benchmark.option_benchmark_run_delay:float(default=-1.0)
The delay between two consecutive runs of benchmarking performance options in seconds.random_shuffle_benchmark_runs:bool(default=true)
Whether to perform all benchmark runs, each of which has different performance options, in a random order.
You can build the benchmark tool with Tensorflow operators support.
To build the tool, you need to use 'benchmark_model_plus_flex' target with '--config=monolithic' option.
bazel build -c opt \
--config=monolithic \
tensorflow/lite/tools/benchmark:benchmark_model_plus_flex
Tensorflow ops support just works the benchmark tool is built with Tensorflow ops support. It doesn't require any additional option to use it.
bazel-bin/tensorflow/lite/tools/benchmark/benchmark_model_plus_flex \
--graph=model_converted_with_TF_ops.tflite \