After having integrated parts of the Dilithium aarch64 code into my library to provide a performance boost, I had to apply the patch [1]. The key of the patch is the change of the NTT_params.h:constants definition from:

-Q1prime

to

Q1prime

Considering the C implementation which uses Q1prime instead of the negative value, I am wondering why this negative value is in the aarch64 code base of PQClean.

Please note that after applying that change, the aarch64 code base calculates the very same signatures as my C or AVX2 implementations. I tested it on an Apple M2, and on an Raspi4 where the code is both compiled for user space and Linux kernel space.

Unfortunately I was not yet able to develop a small test application linking with PQClean analyzing whether the current PQClean code indeed returns the same results as leancrypto.

[1] smuellerDD/leancrypto@c370539