[Fix] Graph's input memory offset

Context

The memory manager generates offsets for each output of each layer within the graph, following a given layout (optimized, wrapping, ...).
Then each kernel within the CPP Export needs to get this memory layout for both its inputs and outputs.

However in the case of the very first layer of the graph, the memory layout of the input was not generated (as it is not the output of any layer).
Then default values were set within the _meminfo.jinja template file, assuming the input would always be contiguous data.
However these default values were not correct, causing buffer overflow and segfault.

Changes

Two changes have been made to correct these default offsets :

• in_size was replaced by in_chan when needed, according to what was found within the memory manager functions;
• in_sizeof was added to make sure that the values actually are in bytes as it has recently been chosen.

changed milestone to %aidge v0.7.0

added Fix 🔥🔥 Help Wanted StatusReview Ready TopicExport labels

requested review from @olivierbichler

assigned to @pineapple and @axelfarr

However this fix may only work with NHWC kernels...

mentioned in merge request aidge_core!467 (merged)

added 1 commit

• 4b251f96 - Fixed standalone tensors dimensions

Compare with previous version

@axelfarr Should be OK, but I am not sure of the template syntax notably for the in_dformat check and it may not be tested in unit tests... I think you should add the revelant unit tests also in this MR.

@axelfarr Can you add a unit test please for both cases ?

added 16 commits

• 4b251f96...2466d41d - 14 commits from branch dev
• 10bbbf50 - [Fix] Change the default offset values for the input of the graph
• fdba9ac4 - Fixed standalone tensors dimensions

Compare with previous version

added 1 commit

• ce9b1582 - Fixed template syntax

Compare with previous version

There is an issue with FC export with these changes: the test_FC_image_in test fails.

added 1 commit

• 05524c28 - Added PaddedConvDepthWise tests

Compare with previous version

mentioned in merge request !59 (merged)

added 1 commit

• 4a90314f - Improved FC export

Compare with previous version

@axelfarr Could you check if everything works now? Also, do not hesite to add unit tests for your cases. This MR must make it in 0.7.0.

approved this merge request

merged