Reshape fixes: don't repack stream for flatten; remove final reshape

[jmduarte]

2 reshape-related fixes:

• Fix to ReshapeStream optimizer: don't repack stream for flatten; Addresses Flatten layer creating a lot of logic #408
  • Resources for TinyML model are consistent with skipping optimizer:

+-----------------+---------+-------+--------+-------+-----+
|       Name      | BRAM_18K| DSP48E|   FF   |  LUT  | URAM|
+-----------------+---------+-------+--------+-------+-----+
|DSP              |        -|      -|       -|      -|    -|
|Expression       |        -|      -|       0|      2|    -|
|FIFO             |      163|      -|    6241|  13136|    -|
|Instance         |       99|      0|   30015|  34856|    0|
|Memory           |        -|      -|       -|      -|    -|
|Multiplexer      |        -|      -|       -|      9|    -|
|Register         |        -|      -|       1|      -|    -|
+-----------------+---------+-------+--------+-------+-----+
|Total            |      262|      0|   36257|  48003|    0|
+-----------------+---------+-------+--------+-------+-----+
|Available        |      280|    220|  106400|  53200|    0|
+-----------------+---------+-------+--------+-------+-----+
|Utilization (%)  |       93|      0|      34|     90|    0|
+-----------------+---------+-------+--------+-------+-----+

• Introduce RemoveFinalReshape optimizer to remove reshape if it's the final layer; Addresses Compilation error with concatenate layer #415
  • Currently, this is in the same file repack_stream.py, but it could be moved
  • This script now works out of the box: https://gist.github.com/jmduarte/4c91f28ab570036aa4cef63db0eb2322
  • Should we add a pytest similar to the above script?

[thesps]

I'm happy with the state of both developments:

• Don't repack stream for flatten: tested the example-models/keras/qkeras_mnist_cnn on pynq-z2, getting identical performance to master (NN accuracy & throughput), but for fewer resources
• RemoveFinalReshape works as expected

Only question: do you want the CI test for the RemoveFinalReshape from #444 ? (It's a PR to this branch)

Some detail on the resource saving for that model:

Version	Total LUTs	Logic LUTs	LUTRAMs	SRLs	FFs	RAMB36	RAMB18	DSP48 Blocks
incl. repack-4-Flatten (master)	33264(62.53%)	31843(59.86%)	22(0.13%)	1399(8.04%)	64112(60.26%)	4(2.86%)	112(40.00%)	58(26.36%)
repack-4-Flatten removed (this PR)	29404(55.27%)	27974(52.58%)	22(0.13%)	1408(8.09%)	44830(42.13%)	4(2.86%)	111(39.64%)	58(26.36%)

[jmduarte]

Yes, now merged!