[Scheduler] fix internal issue when scheduling producer nodes

Review changes
Download
Patches
Plain diff

Merged [Scheduler] fix internal issue when scheduling producer nodes

cguillon/aidge_core:fix-scheduler-debug into dev

Overview 14
Commits 2
Pipelines 7
Changes 2

Merged Christophe Guillon requested to merge cguillon/aidge_core:fix-scheduler-debug into dev 6 months ago

Overview 14
Commits 2
Pipelines 7
Changes 2

Context

This is a fix for a "bug" discovered in aidge_backend_cpu unit tests on scheduler forward, the bug also appears in some ONNX import+forward tests.

Refer to the bug report done in aidge_backend_cpu: aidge_backend_cpu#29 (closed)

Actually with the change introduced in @e3d14f15 Producer nodes may be considered as consumers in the main loop of generateBaseScheduling.

But the function summarizeConsumerState is not protected against node with no input or no output.

This MR is a simple fix to this function to account for case of 0-inputs or 0-outputs.

The case where graph may generate these bugs are for instance single node Producer graph:

one single Producer noe
one single graph output (the producer node)

These graphs are "degenerate" but may be generated as unit tests for ONNX Constant node for instance which will generate a single Producer node.

Or in the aidge_backend_cpu test_scheduler.py test, which is itself buggy (ref to the bug report above), and which give to the scheduler an unconnected Producer node to schedule.