Add AVX2 optimizations for ML-DSA NTT operations - #3
Closed
mdcornu wants to merge 3 commits into
Closed
Conversation
mdcornu
force-pushed
the
ml_dsa_ntt
branch
3 times, most recently
from
February 24, 2026 10:36
7fd35f2 to
7d850a5
Compare
This commit introduces AVX2-vectorized assembly implementations of the Number Theoretic Transform (NTT) operations used in ML-DSA (FIPS 204). These optimizations improve performance of ML-DSA key generation, signing, and verification operations on x86_64 platforms with AVX2 support. The implementation adds the following functions: - ml_dsa_poly_ntt_avx2: Forward NTT transformation - ml_dsa_poly_ntt_inverse_avx2: Inverse NTT transformation - ml_dsa_poly_ntt_mult_avx2: NTT-domain polynomial multiplication Key implementation details: - Uses YMM registers to process 8 32-bit coefficients in parallel - Employs Montgomery reduction for modular arithmetic - Implements NTT butterfly operations across multiple transform levels - Includes dedicated zeta table for INTT to reduce cycles - Runtime capability check via ml_dsa_ntt_avx2_capable() using OPENSSL_ia32cap_P to detect AVX2 support The C code in ml_dsa_ntt.c is updated to dispatch to AVX2 implementations at runtime when available, with automatic fallback to the portable C implementation on platforms without AVX2 support. Build system changes: - Added GENERATE rule for ml_dsa_ntt-x86_64.s from Perl assembly - Conditional assembly inclusion based on target architecture - Works with both libcrypto and FIPS provider builds Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com> Signed-off-by: Tomasz Kantecki <tomasz.kantecki@intel.com>
Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com>
Signed-off-by: Marcel Cornu <marcel.d.cornu@intel.com>
mdcornu
pushed a commit
that referenced
this pull request
Jun 10, 2026
TSAN seems to be having a problem with atomic_load_ptr and
atomic_store_ptr. Both are, by default, __ATOMIC_RELAXED operations.
According to the tsan docs, it flags these operations as a race because,
while they are indivisible, they create no happens-before constraint,
meaning they can be reordered.
An exemplar race that is reported is:
WARNING: ThreadSanitizer: data race (pid=2139404)
Read of size 4 at 0x723400002308 by thread T39:
#0 EVP_MD_up_ref crypto/evp/digest.c:995 (threadstest+0x45032d) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#1 evp_md_up_ref crypto/evp/digest.c:974 (threadstest+0x450242) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#2 ossl_method_up_ref crypto/property/property.c:201 (threadstest+0x4b7a55) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#3 ossl_method_store_cache_get_locked crypto/property/property.c:941 (threadstest+0x4b9922) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#4 ossl_method_store_cache_get crypto/property/property.c:974 (threadstest+0x4b9a47) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#5 inner_evp_generic_fetch crypto/evp/evp_fetch.c:314 (threadstest+0x458186) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#6 evp_generic_fetch crypto/evp/evp_fetch.c:404 (threadstest+0x4586dc) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#7 EVP_MD_fetch crypto/evp/digest.c:985 (threadstest+0x4502d7) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#8 derive_kdk crypto/rsa/rsa_ossl.c:472 (threadstest+0x4cf738) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#9 rsa_ossl_private_decrypt crypto/rsa/rsa_ossl.c:646 (threadstest+0x4d0174) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#10 RSA_private_decrypt crypto/rsa/rsa_crpt.c:48 (threadstest+0x4c6971) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#11 rsa_decrypt providers/implementations/asymciphers/rsa_enc.c:321 (threadstest+0x51cab7) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#12 EVP_PKEY_decrypt crypto/evp/asymcipher.c:280 (threadstest+0x44a9ca) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#13 thread_shared_evp_pkey test/threadstest.c:966 (threadstest+0x404be7) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#14 thread_run test/threadstest.h:67 (threadstest+0x40132d) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
Previous write of size 8 at 0x723400002308 by main thread (mutexes: write M0):
#0 memset <null> (libtsan.so.2+0x4c1eb) (BuildId: 40906101a3a1e1f1ececafafda314aee009d688a)
#1 CRYPTO_zalloc crypto/mem.c:228 (threadstest+0x48679d) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#2 evp_md_new crypto/evp/digest.c:758 (threadstest+0x44f35e) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#3 evp_md_from_algorithm crypto/evp/digest.c:839 (threadstest+0x44f885) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#4 construct_evp_method crypto/evp/evp_fetch.c:230 (threadstest+0x457ec9) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#5 ossl_method_construct_this crypto/core_fetch.c:110 (threadstest+0x4801bf) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#6 algorithm_do_map crypto/core_algorithm.c:77 (threadstest+0x47f7a3) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#7 algorithm_do_this crypto/core_algorithm.c:122 (threadstest+0x47f987) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
#8 ossl_provider_doall_activated crypto/provider_core.c:1609 (threadstest+0x49a42a) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#9 ossl_algorithm_do_all crypto/core_algorithm.c:164 (threadstest+0x47fb14) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#10 ossl_method_construct crypto/core_fetch.c:157 (threadstest+0x4803d0) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#11 inner_evp_generic_fetch crypto/evp/evp_fetch.c:333 (threadstest+0x4583a2) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#12 evp_generic_fetch crypto/evp/evp_fetch.c:404 (threadstest+0x4586dc) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#13 EVP_MD_fetch crypto/evp/digest.c:985 (threadstest+0x4502d7) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#14 derive_kdk crypto/rsa/rsa_ossl.c:472 (threadstest+0x4cf738) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#15 rsa_ossl_private_decrypt crypto/rsa/rsa_ossl.c:646 (threadstest+0x4d0174) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#16 RSA_private_decrypt crypto/rsa/rsa_crpt.c:48 (threadstest+0x4c6971) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#17 rsa_decrypt providers/implementations/asymciphers/rsa_enc.c:321 (threadstest+0x51cab7) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
openssl#18 EVP_PKEY_decrypt crypto/evp/asymcipher.c:280 (threadstest+0x44a9ca) (BuildId: f34377d95e3c1d13ab9aa3204d2f1f7840d1c84a)
What tsan is saying here is that the memset in evp_md_new may get
re-ordered such that the contents of the EVP_MD may still be getting
zeroed at the time we have (a) found the EVP_MD in the method store
cache, and (b) attempted to do an up_ref on it.
This is plainly impossible, especially given that, in order to reach the
method store cache, it must be places in the method store algorithm
sparse array, which still requires the taking of the method store write
lock. But for some reason tsan fails to see the memory fence that
creates.
It seems the simplest solution to correct this is, if we are running
under tsan, use __ATOMIC_ACQUIRE and __ATOMIC_RELEASE on
CRYPTO_atomic_[load|store]_ptr to make sure tsan sees the proper memory
ordering.
Reviewed-by: Saša Nedvědický <sashan@openssl.org>
Reviewed-by: Bob Beck <beck@openssl.org>
Reviewed-by: Nikola Pajkovsky <nikolap@openssl.org>
MergeDate: Tue Jun 9 18:17:19 2026
(Merged from openssl#31018)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Checklist