Question on Register Allocation for Large Convolution Layers in PUMA Compiler

[EVA113x]

Hi PUMA Team,

While compiling a model with large convolution layers, I encountered register-allocation errors. From the debug logs, it appears the allocator may be killing a live register that is still in use. In addition, the allocator seems to require a contiguous memory segment for registers, but the effective register file size is limited to N_INPUT_REGISTERS + N_OUTPUT_REGISTERS. Under these conditions, the allocation frequently exceeds the limit.

Could you confirm whether this behavior is expected, or if there have been recent updates to the register allocator? It would also help to know if there is a recommended workaround or configuration for large conv layers.

Looking forward to your reply.

Best regards,
H. Wu

[ielhajj]

Hi H. Wu. It's hard to tell where the bug is without thorough debugging. It is possible for very large layers to run out of memory. Please note that this project is no longer under active development and support.