Sotfmax bw inference fix

[calad0i]

Description

Fix edge in softmax stable impl norm_t type inference: consider fixed<2,2> *arr of range {-2,-1,0,1} max(arr) - arr can be all {0,1,2,3}, so iwidth and width shall be kept if was signed.

Type of change

• Bug fix (non-breaking change that fixes an issue)

Tests

Checklist

• all

[JanFSchulte]

These test failures seem to be reproducible, so I think this requires a further check before merging:

FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vivado_io_parallel_resource] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vivado_io_parallel_latency] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vivado_io_stream_latency] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vivado_io_stream_resource] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vitis_io_parallel_resource] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vitis_io_parallel_latency] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vitis_io_stream_latency] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)
FAILED test_cnn_mnist_qkeras.py::test_accuracy[Vitis_io_stream_resource] - assert (0.988 > 0.92 and 0.011133603238866406 < 0.01)

[calad0i]

@JanFSchulte

These test failures seem to be reproducible

Indeed they are reproducible, but the situation is rather complicated:

First of all, it is possible to have overflows without the fix. However, having an overflow here that is noticeable in practice is pretty rare.

Regarding the precision drop, by default, the table_sizes are 1024 everywhere, which means we always take 10 most significant bits from the inputs. However, since the output of the last dense is not quantized, it gets by default fixed<16,6>. Without the fix, the table effectively sees ufixed<15,-5>, but with the fix, the table now sees ufixed<16,-6>, making the lookup less accurate in the relevant region (low value).