Decrease Stack Size

[SamuelRiedel]

Make the L1 size a configurable parameter and decrease the default stack size to 512 bytes per core. This is plenty for the kernels I wrote and should easily be enough for most other kernels as well, but @mbertuletti, @yichao-zh, @sermazz, let me know what you think as well.

Changelog

Changed

• Decrease stack size to 128 words

Checklist

• Automated tests pass
• Changelog updated
• Code style guideline is observed

[mbertuletti]

I tested it on some of my kernels and it works safely. For me you would be free to merge.

[sermazz]

For me it also looks good. Just a note about the systolic configuration:
systolic.mk uses the default stack size (which now is going from 1024 to 512 bytes), but increases the size of the sequential memory per core (2048 bytes). This is because the first few words of each tile sequential region are allocated to Xqueue (https://github.com/pulp-platform/mempool/blob/main/software/runtime/crt0.S#L71-L74). Then the remaining sequential space not used by either Xqueue or the stack is allocated to the sequential heap (https://github.com/pulp-platform/mempool/blob/main/software/runtime/runtime.h#L79-L95).
The sequential heap is not heavily employed.. so to reduce memory waste in the systolic configuration too, I would reduce its sequential memory per core from 2048 to 1024 bytes.

[SamuelRiedel]

For me it also looks good. Just a note about the systolic configuration: systolic.mk uses the default stack size (which now is going from 1024 to 512 bytes), but increases the size of the sequential memory per core (2048 bytes). This is because the first few words of each tile sequential region are allocated to Xqueue (https://github.com/pulp-platform/mempool/blob/main/software/runtime/crt0.S#L71-L74). Then the remaining sequential space not used by either Xqueue or the stack is allocated to the sequential heap (https://github.com/pulp-platform/mempool/blob/main/software/runtime/runtime.h#L79-L95). The sequential heap is not heavily employed.. so to reduce memory waste in the systolic configuration too, I would reduce its sequential memory per core from 2048 to 1024 bytes.

Absolutely, I added this to the PR.

[sermazz]

Good to merge for me :)