Skip to content

wangyx95/tpu-perf

 
 

Repository files navigation

TPU Perf

Prerequisites

Install protoc version 3.19 in order to be compatible with ufw.

mkdir build
cd build
cmake ..
make install

Usage

Refer to example for how to create test specs.

Execute from SDK environment.

python3 -m tpu_perf.build --time # To build efficiency test models
python3 -m tpu_perf.build # To build precision test models
python3 -m tpu_perf.run # To run efficency test
python3 -m tpu_perf.precision_benchmark # To run precision benchmark
python3 -m tpu_perf.make_table #To make table for model zoo test result

config.yaml

Preset variables

Field Type Usage
root Preset Repo top directory
home Preset Current model directory
workdir Preset Current model working directory
lmdb_out Preset Dataset output dir. Only set if lmdb field exists.
shape Preset Current shape list. Only set if shapes field exists.
shape_key Preset Current shape string like "1x3x224x224". Only set if shapes field exists.
shape_param Preset Current shape string like "[1,3,224,224]". Only set if shapes field exists.

Common variables

Field Type Usage
name Required Specify network name, should be unique
gops Optional Specify network FLOPs

NNTC

Field Type Usage
precision Optional Boolean type. Indicates whether to build precision model when tpu_perf.build is called.
fp_compile_options Optional Will use this command to build FP32 bmodel.
cali Optional Calibration command. tpu_perf.build will do calibrating if this variable exists.
time_only_cali Optional Calibration command. tpu_perf.build --time will do calibrating if this variable exists.
bmnetu_options Optional bmnetu compile options. Will do int8 compiling if this variable exists
bmnetu_batch_sizes Optional Specify int8 output batches
harness Optional Harness to call when tpu_perf.precision_benchmark is called.
lmdb Optional This object is provided to dataset and harness plugins to do preprocess.

MLIR

Variables in top level config.yaml are also set for all models.

About

A model benchmark tool

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 50.3%
  • C++ 42.6%
  • Dockerfile 3.9%
  • CMake 1.9%
  • C 1.2%
  • Shell 0.1%