Quartus Backend

[hamzajaved780]

This PR integrates Quartus backend with the hls4ml frontend (PR #195). Backend can selected by changing the backend config variable to Quartus / Vivado. Several functions have been pushed back to their backend-specific template and writer files. This allows easier integration of new backends by just defining new template and writer files. This solution still isn't perfect and a further restructuring of the front-end is needed for future seamless integrations.

In addition, the following layers are currently supported on the Quartus backend:

1. Dense Latency / Large with binary/ternary support.
2. Activations (The same as in Vivado)

There is also partial support for sparse dense layers, however there are some scaling issues with wide networks.

[jmitrevs]

Just a minor question: are the .DS_store files added on purpose?

[thesps]

In addition to the .DS_store files (which I don't think should be there), @hamzajaved780 could you remove the converted test models and associated files? They shouldn't be included.

[hamzajaved780]

Was a TF 2.3 related issue, should be fixed now!

[jmitrevs]

Was a TF 2.3 related issue, should be fixed now!

It seems to be fixed.

[jmitrevs]

I think the following change is needed:

diff --git a/hls4ml/writer/vivado_writer.py b/hls4ml/writer/vivado_writer.py
index aa8dc1c..8b68649 100644
--- a/hls4ml/writer/vivado_writer.py
+++ b/hls4ml/writer/vivado_writer.py
@@ -498,7 +498,7 @@ class VivadoWriter(Writer):
             line = line.replace('myproject',model.config.get_project_name())
 
             if 'set_part {xcku115-flvb2104-2-i}' in line:
-                line = 'set_part {{{}}}\n'.format(model.config.get_config_value('XilinxPart'))
+                line = 'set_part {{{}}}\n'.format(model.config.get_config_value('FPGAPart'))
             elif 'create_clock -period 5 -name default' in line:
                 line = 'create_clock -period {} -name default\n'.format(model.config.get_config_value('ClockPeriod'))

@@ -516,7 +516,7 @@ class VivadoWriter(Writer):
         for line in f.readlines():
             line = line.replace('myproject', model.config.get_project_name())
             if '-part' in line:
-                line = 'synth_design -top {} -part {}\n'.format(model.config.get_project_name(), model.config.get_config_value('XilinxPart'))
+                line = 'synth_design -top {} -part {}\n'.format(model.config.get_project_name(), model.config.get_config_value('FPGAPart'))
 
             fout.write(line)
         f.close()

[jmitrevs]

I think this needs to continue using env, not a hardcoded python path.

[jmitrevs]

Fixed in current version.

[jmitrevs]

should be self._show...

[jmitrevs]

should be self._show...

[jmitrevs]

Some of the front end problems that I have mentioned and related issues have been fixed in my branch:

https://github.com/jmitrevs/hls4ml/tree/quartus

[jmitrevs]

I noticed that the quartus nnet_utils are inside the firmware directory while the Vivado ones are parallel to the firmware directory. Would it make sense to use the same location?

I suppose given the different pragmas it doesn't make sense to try to unify these, though I wonder if there's too much code duplication and the potential to deviate for unnecessary reasons. Is there something better we can do?

[vloncar]

I noticed that the quartus nnet_utils are inside the firmware directory while the Vivado ones are parallel to the firmware directory. Would it make sense to use the same location?

At the same level, yes.

I suppose given the different pragmas it doesn't make sense to try to unify these, though I wonder if there's too much code duplication and the potential to deviate for unnecessary reasons. Is there something better we can do?

It's not just the pragmas, the algorithms themselves are different, just look at Dense layer and activations. I don't see nnet_utils being shared, but we can and will reuse as much of template/writer (i.e., python) code as possible.

[jmitrevs]

Closing the old quartus pull request, which has been superseded.