Skip to content

Commit

Permalink
upstream update [full tests]
Browse files Browse the repository at this point in the history 
Signed-off-by: Basil Hess <[email protected]>
  • Loading branch information
@bhess
bhess committed Jan 7, 2025
1 parent fc27988 commit f1d7574
Show file tree
Hide file tree
Showing 88 changed files with 6,754 additions and 6,052 deletions.
2 changes: 1 addition & 1 deletion docs/algorithms/kem/ml_kem.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
- **Authors' website**: https://pq-crystals.org/kyber/ and https://csrc.nist.gov/pubs/fips/203
- **Specification version**: ML-KEM.
- **Primary Source**<a name="primary-source"></a>:
- **Source**: https://github.com/bhess/mlkem-native/commit/fc3fa5f4618ef2e5567d0595be3cdb002d4baaac
- **Source**: https://github.com/bhess/mlkem-native/commit/8f54f09f21583fc0e29103f200fd5a42ec57665d
- **Implementation license (SPDX-Identifier)**: CC0-1.0 or Apache-2.0


Expand Down
2 changes: 1 addition & 1 deletion docs/algorithms/kem/ml_kem.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@ website: https://pq-crystals.org/kyber/ and https://csrc.nist.gov/pubs/fips/203
nist-round: FIPS203
spec-version: ML-KEM
primary-upstream:
source: https://github.com/bhess/mlkem-native/commit/fc3fa5f4618ef2e5567d0595be3cdb002d4baaac
source: https://github.com/bhess/mlkem-native/commit/8f54f09f21583fc0e29103f200fd5a42ec57665d
spdx-license-identifier: CC0-1.0 or Apache-2.0
parameter-sets:
- name: ML-KEM-512
Expand Down
4 changes: 2 additions & 2 deletions scripts/copy_from_upstream/copy_from_upstream.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -33,8 +33,8 @@ upstreams:
-
name: mlkem-native
git_url: https://github.com/bhess/mlkem-native.git
git_branch: updates-2
git_commit: fc3fa5f4618ef2e5567d0595be3cdb002d4baaac
git_branch: updates-3
git_commit: 8f54f09f21583fc0e29103f200fd5a42ec57665d
kem_meta_path: '{pretty_name_full}_META.yml'
kem_scheme_path: '.'
-
Expand Down
6 changes: 3 additions & 3 deletions src/kem/ml_kem/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@ if(OQS_ENABLE_KEM_ml_kem_512_x86_64)
endif()

if(OQS_ENABLE_KEM_ml_kem_512_aarch64)
add_library(ml_kem_512_aarch64 OBJECT mlkem-native_ml-kem-512_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-512_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-512_aarch64/cbd.c mlkem-native_ml-kem-512_aarch64/debug/debug.c mlkem-native_ml-kem-512_aarch64/indcpa.c mlkem-native_ml-kem-512_aarch64/kem.c mlkem-native_ml-kem-512_aarch64/ntt.c mlkem-native_ml-kem-512_aarch64/poly.c mlkem-native_ml-kem-512_aarch64/polyvec.c mlkem-native_ml-kem-512_aarch64/rej_uniform.c mlkem-native_ml-kem-512_aarch64/verify.c mlkem-native_ml-kem-512_aarch64/zetas.c)
add_library(ml_kem_512_aarch64 OBJECT mlkem-native_ml-kem-512_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-512_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-512_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-512_aarch64/aarch64/src/rej_uniform_table.c mlkem-native_ml-kem-512_aarch64/cbd.c mlkem-native_ml-kem-512_aarch64/debug/debug.c mlkem-native_ml-kem-512_aarch64/indcpa.c mlkem-native_ml-kem-512_aarch64/kem.c mlkem-native_ml-kem-512_aarch64/ntt.c mlkem-native_ml-kem-512_aarch64/poly.c mlkem-native_ml-kem-512_aarch64/polyvec.c mlkem-native_ml-kem-512_aarch64/rej_uniform.c mlkem-native_ml-kem-512_aarch64/verify.c mlkem-native_ml-kem-512_aarch64/zetas.c)
target_include_directories(ml_kem_512_aarch64 PRIVATE ${CMAKE_CURRENT_LIST_DIR}/mlkem-native_ml-kem-512_aarch64)
target_include_directories(ml_kem_512_aarch64 PRIVATE ${PROJECT_SOURCE_DIR}/src/common/pqclean_shims)
target_compile_options(ml_kem_512_aarch64 PUBLIC -DMLKEM_K=2 -DFORCE_AARCH64 -DMLKEM_NATIVE_ARITH_BACKEND_NAME=AARCH64_OPT -DMLKEM_USE_NATIVE)
Expand All @@ -50,7 +50,7 @@ if(OQS_ENABLE_KEM_ml_kem_768_x86_64)
endif()

if(OQS_ENABLE_KEM_ml_kem_768_aarch64)
add_library(ml_kem_768_aarch64 OBJECT mlkem-native_ml-kem-768_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-768_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-768_aarch64/cbd.c mlkem-native_ml-kem-768_aarch64/debug/debug.c mlkem-native_ml-kem-768_aarch64/indcpa.c mlkem-native_ml-kem-768_aarch64/kem.c mlkem-native_ml-kem-768_aarch64/ntt.c mlkem-native_ml-kem-768_aarch64/poly.c mlkem-native_ml-kem-768_aarch64/polyvec.c mlkem-native_ml-kem-768_aarch64/rej_uniform.c mlkem-native_ml-kem-768_aarch64/verify.c mlkem-native_ml-kem-768_aarch64/zetas.c)
add_library(ml_kem_768_aarch64 OBJECT mlkem-native_ml-kem-768_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-768_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-768_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-768_aarch64/aarch64/src/rej_uniform_table.c mlkem-native_ml-kem-768_aarch64/cbd.c mlkem-native_ml-kem-768_aarch64/debug/debug.c mlkem-native_ml-kem-768_aarch64/indcpa.c mlkem-native_ml-kem-768_aarch64/kem.c mlkem-native_ml-kem-768_aarch64/ntt.c mlkem-native_ml-kem-768_aarch64/poly.c mlkem-native_ml-kem-768_aarch64/polyvec.c mlkem-native_ml-kem-768_aarch64/rej_uniform.c mlkem-native_ml-kem-768_aarch64/verify.c mlkem-native_ml-kem-768_aarch64/zetas.c)
target_include_directories(ml_kem_768_aarch64 PRIVATE ${CMAKE_CURRENT_LIST_DIR}/mlkem-native_ml-kem-768_aarch64)
target_include_directories(ml_kem_768_aarch64 PRIVATE ${PROJECT_SOURCE_DIR}/src/common/pqclean_shims)
target_compile_options(ml_kem_768_aarch64 PUBLIC -DMLKEM_K=3 -DFORCE_AARCH64 -DMLKEM_NATIVE_ARITH_BACKEND_NAME=AARCH64_OPT -DMLKEM_USE_NATIVE)
Expand All @@ -76,7 +76,7 @@ if(OQS_ENABLE_KEM_ml_kem_1024_x86_64)
endif()

if(OQS_ENABLE_KEM_ml_kem_1024_aarch64)
add_library(ml_kem_1024_aarch64 OBJECT mlkem-native_ml-kem-1024_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-1024_aarch64/cbd.c mlkem-native_ml-kem-1024_aarch64/debug/debug.c mlkem-native_ml-kem-1024_aarch64/indcpa.c mlkem-native_ml-kem-1024_aarch64/kem.c mlkem-native_ml-kem-1024_aarch64/ntt.c mlkem-native_ml-kem-1024_aarch64/poly.c mlkem-native_ml-kem-1024_aarch64/polyvec.c mlkem-native_ml-kem-1024_aarch64/rej_uniform.c mlkem-native_ml-kem-1024_aarch64/verify.c mlkem-native_ml-kem-1024_aarch64/zetas.c)
add_library(ml_kem_1024_aarch64 OBJECT mlkem-native_ml-kem-1024_aarch64/aarch64/src/aarch64_zetas.c mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/ntt_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/ntt_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/poly_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/poly_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/polyvec_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/polyvec_opt.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/rej_uniform_asm_clean.S mlkem-native_ml-kem-1024_aarch64/aarch64/src/rej_uniform_table.c mlkem-native_ml-kem-1024_aarch64/cbd.c mlkem-native_ml-kem-1024_aarch64/debug/debug.c mlkem-native_ml-kem-1024_aarch64/indcpa.c mlkem-native_ml-kem-1024_aarch64/kem.c mlkem-native_ml-kem-1024_aarch64/ntt.c mlkem-native_ml-kem-1024_aarch64/poly.c mlkem-native_ml-kem-1024_aarch64/polyvec.c mlkem-native_ml-kem-1024_aarch64/rej_uniform.c mlkem-native_ml-kem-1024_aarch64/verify.c mlkem-native_ml-kem-1024_aarch64/zetas.c)
target_include_directories(ml_kem_1024_aarch64 PRIVATE ${CMAKE_CURRENT_LIST_DIR}/mlkem-native_ml-kem-1024_aarch64)
target_include_directories(ml_kem_1024_aarch64 PRIVATE ${PROJECT_SOURCE_DIR}/src/common/pqclean_shims)
target_compile_options(ml_kem_1024_aarch64 PUBLIC -DMLKEM_K=4 -DFORCE_AARCH64 -DMLKEM_NATIVE_ARITH_BACKEND_NAME=AARCH64_OPT -DMLKEM_USE_NATIVE)
Expand Down
13 changes: 6 additions & 7 deletions src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/aarch64_zetas.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@
* Table of zeta values used in the AArch64 forward NTT
* See autogenerate_files.py for details.
*/
const int16_t aarch64_ntt_zetas_layer01234[] = {
ALIGN const int16_t aarch64_ntt_zetas_layer01234[] = {
-1600, -15749, -749, -7373, -40, -394, -687, -6762, 630, 6201,
-1432, -14095, 848, 8347, 0, 0, 1062, 10453, 296, 2914,
-882, -8682, 0, 0, -1410, -13879, 1339, 13180, 1476, 14529,
Expand All @@ -31,7 +31,7 @@ const int16_t aarch64_ntt_zetas_layer01234[] = {
0, 0, -1583, -15582, -1355, -13338, 821, 8081, 0, 0,
};

const int16_t aarch64_ntt_zetas_layer56[] = {
ALIGN const int16_t aarch64_ntt_zetas_layer56[] = {
289, 289, 331, 331, -76, -76, -1573, -1573, 2845,
2845, 3258, 3258, -748, -748, -15483, -15483, 17, 17,
583, 583, 1637, 1637, -1041, -1041, 167, 167, 5739,
Expand Down Expand Up @@ -77,7 +77,7 @@ const int16_t aarch64_ntt_zetas_layer56[] = {
10129, 10129, -3878, -3878, -11566, -11566,
};

const int16_t aarch64_invntt_zetas_layer01234[] = {
ALIGN const int16_t aarch64_invntt_zetas_layer01234[] = {
1583, 15582, -821, -8081, 1355, 13338, 0, 0, -569, -5601,
450, 4429, 936, 9213, 0, 0, 69, 679, 447, 4400,
-535, -5266, 0, 0, 543, 5345, 1235, 12156, -1426, -14036,
Expand All @@ -88,7 +88,7 @@ const int16_t aarch64_invntt_zetas_layer01234[] = {
-848, -8347, 1432, 14095, -630, -6201, 687, 6762, 0, 0,
};

const int16_t aarch64_invntt_zetas_layer56[] = {
ALIGN const int16_t aarch64_invntt_zetas_layer56[] = {
-910, -910, -1227, -1227, 219, 219, 855, 855, -8957,
-8957, -12078, -12078, 2156, 2156, 8416, 8416, 1175, 1175,
394, 394, -1029, -1029, -1212, -1212, 11566, 11566, 3878,
Expand Down Expand Up @@ -134,7 +134,7 @@ const int16_t aarch64_invntt_zetas_layer56[] = {
-16113, -16113, -5739, -5739, -167, -167,
};

const int16_t aarch64_zetas_mulcache_native[] = {
ALIGN const int16_t aarch64_zetas_mulcache_native[] = {
17, -17, -568, 568, 583, -583, -680, 680, 1637, -1637, 723,
-723, -1041, 1041, 1100, -1100, 1409, -1409, -667, 667, -48, 48,
233, -233, 756, -756, -1173, 1173, -314, 314, -279, 279, -1626,
Expand All @@ -149,7 +149,7 @@ const int16_t aarch64_zetas_mulcache_native[] = {
1219, -394, 394, 885, -885, -1175, 1175,
};

const int16_t aarch64_zetas_mulcache_twisted_native[] = {
ALIGN const int16_t aarch64_zetas_mulcache_twisted_native[] = {
167, -167, -5591, 5591, 5739, -5739, -6693, 6693, 16113,
-16113, 7117, -7117, -10247, 10247, 10828, -10828, 13869, -13869,
-6565, 6565, -472, 472, 2293, -2293, 7441, -7441, -11546,
Expand All @@ -168,7 +168,6 @@ const int16_t aarch64_zetas_mulcache_twisted_native[] = {
};

#else
#include "params.h"

/* Dummy declaration for compilers disliking empty compilation units */
#define empty_cu_aarch64_zetas MLKEM_NAMESPACE(empty_cu_aarch64_zetas)
Expand Down
4 changes: 3 additions & 1 deletion src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/arith_native_aarch64.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,13 +19,15 @@
MLKEM_NAMESPACE(aarch64_zetas_mulcache_native)
#define aarch64_zetas_mulcache_twisted_native \
MLKEM_NAMESPACE(aarch64_zetas_mulcache_twisted_native)
#define rej_uniform_table MLKEM_NAMESPACE(rej_uniform_table)

extern const int16_t aarch64_ntt_zetas_layer01234[];
extern const int16_t aarch64_ntt_zetas_layer56[];
extern const int16_t aarch64_invntt_zetas_layer01234[];
extern const int16_t aarch64_invntt_zetas_layer56[];
extern const int16_t aarch64_zetas_mulcache_native[];
extern const int16_t aarch64_zetas_mulcache_twisted_native[];
extern const uint8_t rej_uniform_table[];

#define ntt_asm_clean MLKEM_NAMESPACE(ntt_asm_clean)
void ntt_asm_clean(int16_t *, const int16_t *, const int16_t *);
Expand All @@ -41,7 +43,7 @@ void intt_asm_opt(int16_t *, const int16_t *, const int16_t *);

#define rej_uniform_asm_clean MLKEM_NAMESPACE(rej_uniform_asm_clean)
unsigned int rej_uniform_asm_clean(int16_t *r, const uint8_t *buf,
unsigned int buflen);
unsigned int buflen, const uint8_t *table);

#define poly_reduce_asm_clean MLKEM_NAMESPACE(poly_reduce_asm_clean)
void poly_reduce_asm_clean(int16_t *);
Expand Down
2 changes: 1 addition & 1 deletion src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/clean_impl.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,7 @@ static INLINE int rej_uniform_native(int16_t *r, unsigned int len,
{
return -1;
}
return (int)rej_uniform_asm_clean(r, buf, buflen);
return (int)rej_uniform_asm_clean(r, buf, buflen, rej_uniform_table);
}

#endif /* MLKEM_NATIVE_ARITH_PROFILE_IMPL_H */
13 changes: 0 additions & 13 deletions src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/common.i
View file

This file was deleted.

76 changes: 33 additions & 43 deletions src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_clean.S
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,10 +26,6 @@
#include "common.h"
#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_CLEAN)

// Needed to provide ASM_LOAD directive
#include "common.i"
#include "params.h"

// Bounds:
// If C is chosen so that |src| < q * C, then |dst| < q * (0.0508 * C + 1/2)
//
Expand Down Expand Up @@ -145,38 +141,7 @@

.text

.global MLKEM_NAMESPACE(intt_asm_clean)
.global _MLKEM_NAMESPACE(intt_asm_clean)

.p2align 4
const_addr: .short 3329
.short 20159
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
ninv_addr: .short 512
.short 512
.short 512
.short 512
.short 512
.short 512
.short 512
.short 512
ninv_tw_addr: .short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040

MLKEM_NAMESPACE(intt_asm_clean):
_MLKEM_NAMESPACE(intt_asm_clean):
push_stack
.global MLKEM_ASM_NAMESPACE(intt_asm_clean)

in .req x0
r01234_ptr .req x1
Expand Down Expand Up @@ -227,16 +192,41 @@ _MLKEM_NAMESPACE(intt_asm_clean):
t2 .req v27
t3 .req v28

ASM_LOAD(xtmp, const_addr)
ld1 {consts.8h}, [xtmp]

ninv .req v29
q_ninv .req q29
ninv_tw .req v30
q_ninv_tw .req q30

/* Literal pool */
.macro dup8h c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.endm

.p2align 4
c_consts: .short 3329
.short 20159
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
c_ninv: dup8h 512
c_ninv_tw: dup8h 5040

MLKEM_ASM_NAMESPACE(intt_asm_clean):
push_stack

ASM_LOAD(xtmp, ninv_addr)
ld1r {ninv.8h}, [xtmp]
ASM_LOAD(xtmp, ninv_tw_addr)
ld1r {ninv_tw.8h}, [xtmp]
ldr q_consts, c_consts
ldr q_ninv, c_ninv
ldr q_ninv_tw, c_ninv_tw

mov inp, in
mov count, #8
Expand Down
76 changes: 33 additions & 43 deletions src/kem/ml_kem/mlkem-native_ml-kem-1024_aarch64/aarch64/src/intt_opt.S
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,10 +26,6 @@
#include "common.h"
#if defined(MLKEM_NATIVE_ARITH_BACKEND_AARCH64_OPT)

// Needed to provide ASM_LOAD directive
#include "common.i"
#include "params.h"

// Bounds:
// If C is chosen so that |src| < q * C, then |dst| < q * (0.0508 * C + 1/2)
//
Expand Down Expand Up @@ -145,38 +141,7 @@

.text

.global MLKEM_NAMESPACE(intt_asm_opt)
.global _MLKEM_NAMESPACE(intt_asm_opt)

.p2align 4
const_addr: .short 3329
.short 20159
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
ninv_addr: .short 512
.short 512
.short 512
.short 512
.short 512
.short 512
.short 512
.short 512
ninv_tw_addr: .short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040
.short 5040

MLKEM_NAMESPACE(intt_asm_opt):
_MLKEM_NAMESPACE(intt_asm_opt):
push_stack
.global MLKEM_ASM_NAMESPACE(intt_asm_opt)

in .req x0
r01234_ptr .req x1
Expand Down Expand Up @@ -227,16 +192,41 @@ _MLKEM_NAMESPACE(intt_asm_opt):
t2 .req v27
t3 .req v28

ASM_LOAD(xtmp, const_addr)
ld1 {consts.8h}, [xtmp]

ninv .req v29
q_ninv .req q29
ninv_tw .req v30
q_ninv_tw .req q30

/* Literal pool */
.macro dup8h c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.short \c
.endm

.p2align 4
c_consts: .short 3329
.short 20159
.short 0
.short 0
.short 0
.short 0
.short 0
.short 0
c_ninv: dup8h 512
c_ninv_tw: dup8h 5040

MLKEM_ASM_NAMESPACE(intt_asm_opt):
push_stack

ASM_LOAD(xtmp, ninv_addr)
ld1r {ninv.8h}, [xtmp]
ASM_LOAD(xtmp, ninv_tw_addr)
ld1r {ninv_tw.8h}, [xtmp]
ldr q_consts, c_consts
ldr q_ninv, c_ninv
ldr q_ninv_tw, c_ninv_tw

mov inp, in
mov count, #8
Expand Down
Loading

0 comments on commit f1d7574

Please sign in to comment.