chore(concrete_cpu): fix nightly feature which was not forwarded to TFHE-rs #654
Conversation
- this will allow to have AVX512 on x86 machines which have the hardware feature available if concrete-cpu is built with a rust nightly toolchain and the nightly feature enabled
IceTDrinker
force-pushed
the
am/fix/concrete-cpu-nightly
branch
from
59a92cd to
5833b23
Compare
|
@BourgerieQuentin , what about we organize a peer review by TFHE-rs people in Q1, about our integration? To be sure we've the best setting |
|
And if it's better for speed, let's make sure we take it for Q4 release! |
|
@slab-ci compiler-cpu-bench |
|
Failed to run command: command 'compiler-cpu-bench' not found in config file |
|
@slab-ci compiler-cpu-benchmark |
|
where the avx512 instruction are leverage? only in the fft? from benchmark result is difficult to see improvement (is on the line) |
it is code around the fft and data conversion mostly, it's a mismatch compared to what it was in concrete-cpu. I added the nightly feature to be sure that the same code was sure to be used in TFHE-rs (e.g. in concrete-cpu the convert/x86.rs file was making use of the nightly feature https://github.com/zama-ai/concrete/pull/605/files#diff-f2a36b075a72ef8dfb6ec4f65d38151bdadb460c0117eaf78079a1bff204ba60) in TFHE-rs this would be in tfhe/src/core_crypto/fft_impl/fft64/math/fft/x86.rs with nightly-avx512 If there is no benefit then I guess it could be left off but be aware that TFHE-rs won't be using AVX512 if available on the machine without that feature |
as discussed and as proposed by @bcm-at-zama I pushed this small fix to make the AVX512 feature from TFHE-rs available in concrete-cpu when the nightly feature is enbaled.