`test_softmax.py` does not test `argmax` and `latency` implementations; `latency` fails

[vasdommes]

Quick summary

1. test_softmax.py does not actually test non-default implementations (argmax and latency).
2. If we fix that, tests for latency fail, reporting low accuracy.

Details

Steps to Reproduce

1. Clone the hls4ml repository
2. Checkout the main branch, commit hash: 6900257
3. Run pytest test/pytest/test_softmax.py
4. Go e.g. to test/pytest/test_softmax_test_softmax_9,6-input_shape6-18,8-io_parallel-False-latency-Vivado/parameters.h, line 29, see:

    static const nnet::softmax_implementation implementation = nnet::softmax_implementation::stable;

instead of expected nnet::softmax_implementation::latency.

5. Now, if we replace the line
   

   hls4ml/test/pytest/test_softmax.py

   Line 47 in 6900257

   cfg['LayerName']['softmax']['Strategy'] = strategy

   
   with

    cfg['LayerName']['softmax']['implementation'] = strategy

then all latency tests fail with messages like:

Accuracy hls4ml relative to keras: 0.1908

6. Now, if we replace

   hls4ml/test/pytest/test_softmax.py

   Line 19 in 6900257

   return np.clip(d, -32, 31)

   
   with

    return np.clip(d, -32, 0)

the the tests pass.

(Catapult backend fails on my laptop because it doesn't create .so file, but that's unrelated).