Low bit support

This MR add new implementations and features to support low bit export on devices based on Arm cortexM processor.

• Add configuration files for STM32F7 in the available boards

• Add implementations and custom datatypes to support low bit inference Custom datatypes :

• data<7> (int7) & udata<7> (uint7)

• data<6> (int6) & udata<6> (uint6)

• data<5> (int5) & udata<5> (uint5)

• data<4> (int4) & udata<4> (uint4)

• data<3> (int3) & udata<3> (uint3)

• data<2> (int2) & udata<2> (uint2)

• data<1> (binary)

• Add registry for low bit implementations in operator.py

• Add jinja templates for lowbit implementations

• Add data_conversion map for ARM cortexM lowbit export in data_conversion.py. This map is passed as argument in the instanciation of NodeExport in the registry for low bit implementations in operator.py.

Closes : #22 (closed)

requested review from @cmoineau

assigned to @thibaultallenet

changed the description

added 9 commits

• 749352a8 - Add configuration files for STM32F7
• f7f2394a - Add data conversion map for low bit datatypes
• a4dd7842 - Update typedefs with new custom datatypes
• a7efeb7e - Update export.py to include new F7 board, fix a typo & remove unecessary copy tree
• 9cb10213 - Update export registry with new tensor entries
• 6acae3ed - Add ExportNodeCpp classes registry for lowbit implementations
• dfef8d16 - Add Implementations for low bit kernels
• e73291d9 - Add jinja templates
• ec1d6f8e - Update parameter.jinja

Compare with previous version

marked this merge request as ready

added 10 commits

• ec1d6f8e...1a4b72ef - 8 commits from branch dev
• 9de26a0c - Merge branch 'dev' into low_bit_support
• c012abcc - solve conflict

Compare with previous version

added Feature 🚀 label

changed the description

mentioned in issue #22 (closed)

changed the description

added 1 commit

• 10bfb0ac - Rename "custom" into "Lowbit"

Compare with previous version

mentioned in commit 5f1820a1

merged