Faster data unpacking

[mdanilow]

Faster implementation of finn.util.data_packing.packed_bytearray_to_finnpy. It utilizes bitshifts and is necessary for realtime processing on pynq.

[fpjentzsch]

Might be a good idea to review/merge this PR together with #1172 (@bwintermann).

[bwintermann]

I indeed ran into the same issue and opened the PR mentioned by @fpjentzsch because of it. Did you measure how much faster this is than the previous approach? I would be quite happy if we didn't have to load some C code to get fast execution speed.

[fpjentzsch]

I tried this on a MNv1 with FLOAT32 output and got this error because you expect a 5-dimensional output shape:

File "/home/xilinx/jupyter_notebooks/ras_end2end_val_manual/measurement/driver/finn/util/data_packing.py", line 451, in packed_bytearray_to_finnpy
result[:, :, :, :, i] += ret[:, :, :, :, packing*i + fold] << (packing - 1 - fold) * 8
IndexError: too many indices for array: array is 3-dimensional, but 5 were indexed

After adjusting the code from 5 to 3 dimensions, it worked for this case and sped up the output data unpacking by ~3000x.
The new bottleneck is now the ImageNet input data pre-processing, so my overall speedup was ~14x.

[auphelia]

Closing this PR because it is superseded by this merged PR: #1473
@mdanilow, please let me know if I’m overlooking anything and we can reopen it.