- Add code for the new dataset [1] (e.g.
snu_sr.{cc,h}
) toflutter/cpp/datasets
and corresponding building rule(s) influtter/cpp/datasets/BUILD
. - Add corresponding code to
flutter/cpp/binary/main.cc
so we can test the dataset with CLI. - Add a new task specific
benchmark_setting
to backends so that accelerators could be used. E.g. adding a new setting to TFLite backend by modifyingmobile_back_tflite/cpp/backend_tflite/tflite_settings_android.h
, for an example see d23d16c - Add a new task to the
flutter/assets/tasks.pbtxt
file. - Add a new task to Flutter code part.
- Add an integration test for the new task.
See the PR #574 for an example how to add the new SNUSR dataset.
See the PR #608 for an example how to add the new super resolution task.
See the issue #595 for a complete discussion how we added the new super resolution task.
[1] What does "a dataset" mean? A dataset here is to implement LoadGen's Query Sample Library (QSL). Actually, we can share some common methods and the MLPerfDriver uses interface defined in flutter/cpp/dataset.h
, all the datasets in flutter/cpp/datasets/
now inherit the mlperf::mobile::Dataset
.