Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

vitis:fifo_depth_optimization fails for models with more than one input or output

Open Beginner friendly
  • Is something wrong?
#1,582 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 3 days

@dgaytanv is already working on this.

Since Oct 15, 2026.

  • #1583 by @dgaytanv — open

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
78/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
keras, python
Domain
backend

Research direction

Start in hls4ml/backends/vitis/passes/fifo_depth_optimization.py, in initialize_large_fifos, where the exclusion checks compare each variable only with the first input and first output. The issue proposes replacing those comparisons with membership tests against all inputs and outputs. Verify it with a two-input, two-output Keras model, calling initialize_large_fifos and checking that the second input and output are not returned. Done when that holds and the Vitis flow completes without the RuntimeError.

Written by the indexing model from the issue text.

Description

bug good first issue
Quick summary

The FIFO depth optimization flow fails for models with more than one input or output. initialize_large_fifos leaves out only the first input and the first output of the model from profiling. The other inputs and outputs are profiled too, but they are ports of the top function, not FIFOs inside it, so they never appear in channel_info.csv. The last step then cannot find their depths and stops.

This happens on Vitis and on VitisUnified (axi_master), and is independent of #1570, which is fixed by #1571.

hls4ml version

#1571 at e1245b0cd. Without #1571 the flow stops earlier, with the KeyError from #1570.

Frontend

Keras v2

Backend

Vitis

io_type

io_stream

HLS tool and version

Vitis HLS 2024.1

Steps to reproduce
import hls4ml
from tensorflow.keras.layers import Concatenate, Dense, Input
from tensorflow.keras.models import Model

in_a = Input((16,), name='in_a')
in_b = Input((16,), name='in_b')
h = Concatenate(name='cat')([Dense(8, name='da')(in_a), Dense(8, name='db')(in_b)])
h = Dense(8, name='dh')(h)
model = Model([in_a, in_b], [Dense(4, name='oa')(h), Dense(2, name='ob')(h)])

config = hls4ml.utils.config_from_keras_model(model, granularity='name', default_precision='ap_fixed<16, 6>')
config['Flows'] = ['vitis:fifo_depth_optimization']

hls_model = hls4ml.converters.convert_from_keras_model(
    model,
    hls_config=config,
    io_type='io_stream',
    output_dir='fifo_multi_io_prj',
    backend='Vitis',
    part='xcu250-figd2104-2L-e',
)
Expected behaviour

The flow finishes, and fifo_depths.json holds an optimized depth for every FIFO between the layers. The model's inputs and outputs are not profiled.

Actual behaviour

Synthesis and co-simulation complete, then the flow stops:

RuntimeError: FIFO depth optimization could not find the profiled depths of in_b, layer12_out in fifo_multi_io_prj/myproject_prj/solution1/.autopilot/db/channel_info.csv. Make sure that the co-simulation completed successfully.

in_b is the second input and layer12_out the second output. fifo_depths.json lists them with the FIFOs to profile:

{"in_b": {"initial": 1}, "layer3_out": {"initial": 1}, "layer5_out": {"initial": 1}, "layer7_out": {"initial": 1}, "layer8_out": {"initial": 1}, "layer12_out": {"initial": 1}, "layer14_cpy1": {"initial": 1}, "layer14_cpy2": {"initial": 1}}

channel_info.csv has rows only for the FIFOs inside the top function:

,layer3_out_U,,chan_status1.csv
,layer5_out_U,,chan_status2.csv
,layer7_out_U,,chan_status3.csv
,layer8_out_U,,chan_status4.csv
,layer14_cpy1_U,,chan_status5.csv
,layer14_cpy2_U,,chan_status6.csv

With backend='VitisUnified', board='zcu102' and axi_mode='axi_master', the same model stops with the same error for in_b and layer12_out.

Possible fix or additional context

initialize_large_fifos in hls4ml/backends/vitis/passes/fifo_depth_optimization.py compares each variable only with the first input and the first output:

        and output_variable != model.get_output_variables()[0]
        and output_variable != model.get_input_variables()[0]

Comparing with all of them fixes it:

        and output_variable not in model.get_output_variables()
        and output_variable not in model.get_input_variables()

With this change the model above goes through on Vitis, and the co-simulation with the optimized depths passes.

The Vivado FIFO optimizer has the same check (hls4ml/backends/vivado/passes/fifo_depth_optimization.py, line 76). I have not run the Vivado flow, so I don't know whether it fails the same way there.

The check can be tested without any tools: convert a model with two inputs and two outputs, call initialize_large_fifos, and check that the second input and the second output are not among the FIFOs it returns.

AI assistance
  • An AI tool helped produce this report, and I have checked its content myself.
Dominant language
Python
Stars
2.2k
Forks
605
Avg merge
7d 35m
Merged PRs (30d)
19

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from fastmachinelearning/hls4ml

All issues in fastmachinelearning/hls4ml

Similar issues

More Python issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.