Add general transpose for vivado/vitis

[calad0i]

Description

Add general N-dimensional transpose for vivado/vitis backend. Resource consumption validated to be unchanged for all 2/3d io_parallel Vivado/Vitis HLS (2020.1/2023.2), and 2d io_stream Vitis HLS.

For Vivado HLS there is a resource overhead for io_sream transpose 2d. Template overload is used to use the original impl for io_stream 2d.

Type of change

• New feature (non-breaking change which adds functionality)

Tests

test/pytest/test_transpose_concat.py/test_highdim_permute

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

[jmitrevs]

I would propose the following change with this PR since it's no longer needed:

diff --git a/hls4ml/model/optimizer/passes/convert_to_channels_last.py b/hls4ml/model/optimizer/passes/convert_to_channels_last.py
index 606f42e5..5b1a69f1 100644
--- a/hls4ml/model/optimizer/passes/convert_to_channels_last.py
+++ b/hls4ml/model/optimizer/passes/convert_to_channels_last.py
@@ -103,11 +103,6 @@ class ChannelsLastConverter(OptimizerPass):
                 input = previous_node.name
                 outshape = previous_node.get_output_variable().shape
 
-                if (model.config.config['IOType'] == 'io_stream') and len(outshape) == 3:
-                    raise Exception(
-                        'No 3D transpose available in io_stream, this model cannot be converted to channels-last'
-                    )
-
                 if len(outshape) == 2:
                     attributes = {'perm': [1, 0]}
                 else:

[jmitrevs]

I assume you tried various things, and I know C++ can be kind of clunky for stuff like this, but is there an easier/less verbose way to initialize these arrays? Would std::arrays be more readable?

[calad0i]

This is due to that vivado/vitis cannot handle constexpr array indexing with "variables" (indices can be inferenced, but not literal, like an iterator index in a fixed-boundary for loop).

If we are to drop vivado hls support completely, this can be made constexpr and initialized with one-line, and recursive template unrolling will need to be used in the caller function to do the index mapping, which only works for vitis hls and takes forever on vivado hls. I did not try std::array, as I did not expect it to be more compatible than a constexpr raw array.

[jmitrevs]

A comment or docstring would be useful to understand how this works.

[jmitrevs]

This PR looks to be quite useful and I think should be quite rapidly merged, since it makes pytorch streaming models more parsable (after removing the no longer needed graceful failure, with this PR).