Vivado HLS generating multiple output streams

[lgrothh]

Hello again,

when synthesizing my design, Vivado HLS breaks up the result vector into individual AXI streams. I would assume that the intended behavior would be to have one input and one output stream to the IP Core.

In my mnist example, HLS generates 10 individual AXI streams for the return Values:

I've also seen my other projects behave in the same way. Is this an issue with the Vivado HLS Version I'm using (2020.1) or do I have something configured incorrectly in Vivado HLS?

I would appreciate some help with this,
Best Regards,
Lukas

[thesps]

Hi, which io_type are you using? To get streams (hls::streams not AXI Streams) use io_type = 'io_stream'. To really get an AXI Stream interface, use backend='VivadoAccelerator' (in current master branch not v0.5.0), with interface='axi_stream' (which is the default value).

I would assume that the intended behavior would be to have one input and one output stream to the IP Core.

For context, it depends on the application, there are those which provide parallel input (hence io_type='io_parallel' option).

[bhitesh311]

backend='VivadoAccelerator' no longer works for Vivado tools after 2021.2

[bo3z]

Can you try out the Vitis Accelerator (#991) or Coyote Accelerator backends (#1347). They work with Vitis backends.

[bhitesh311]

import hls4ml
print(hls4ml.backends.get_available_backends())

when I run the above command it gives me the output as
['vivado', 'vivadoaccelerator', 'vitis', 'quartus', 'catapult', 'symbolicexpression', 'oneapi']

The vivado version I have installed is 2024.1, I guess this shouldnt be an issue

[bo3z]

They are not merged yet. You will have to check out the branch from the respective PRs mentioned above and install hls4ml locally (by doing cd hls4ml && pip install .)

[bhitesh311]

I did that. But I guess only alveo-u55c is supported. I happen to use PYNQ-Z2.

[bo3z]

You can try the Vitis Unified backend (#1376) or Vitis IP flow (#1134); these are targetting edge devices.