Auto generated convolution tests

The python scripts in test/gen generate a large number of distinct test cases to cover a number of corner cases and different configurations of the different operations provided by portDNN.

Generating the tests

We provide a simple script to generate all tests using python (both 2 and 3 are supported) in the test/gen directory.

python test/gen/generate_all.py

Dependencies

The script uses TensorFlow to generate the expected output values, as such it requires the following:

• TensorFlow 2.x
• Numpy

Changing the test parameters

The tests generators generate different test cases based on a number of parameters. These can be changed in the scripts to generate different tests, but any changes to the filenames and generated files must also be reflected in changes to the build system, otherwise new tests will not be compiled or run.

Conv2D tests

The main parameters in the generation of the conv2d tests are the window and stride sizes. The window stride pairs used in the test generation are defined in the WINDOW_LIST and STRIDE_LIST lists defined at the start of the generate_conv2d_tests.py script.

WINDOW_LIST = [1, 1, 3, 3, 5, 5]
STRIDE_LIST = [1, 2, 1, 2, 1, 2]

Any changes here should be mirrored in the corresponding CMakeLists.txt file, where the generated tests are added to the build system.

set(_windows 1 1 3 3 5 5)
set(_strides 1 2 1 2 1 2)

Changing the test code

The code which will be included in the generated tests is provided as a number of strings in the python script. If the generated code needs to be changed then these strings should be changed.

For example if a new selector type is introduced and should be added to the generated tests then it should be appended to the selector list:

using Selectors = sycldnn::types::TypeList<sycldnn::conv2d::DirectSelector,
                                           sycldnn::conv2d::TiledSelector>;