kris/pqc
1
1
Fork 0
Code Issues Pull Requests Releases Wiki Activity
ebcc71c51a
pqc/crypto_kem/kyber512/avx2/shuffle.S

262 lines
4.6 KiB
ArmAsm
Raw Normal View History

Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
#include "cdecl.h"
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
.include "fq.inc"
.include "shuffle.inc"
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
/*
nttpack_avx:
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
#load
vmovdqa (%rdi),%ymm4
vmovdqa 32(%rdi),%ymm5
vmovdqa 64(%rdi),%ymm6
vmovdqa 96(%rdi),%ymm7
vmovdqa 128(%rdi),%ymm8
vmovdqa 160(%rdi),%ymm9
vmovdqa 192(%rdi),%ymm10
vmovdqa 224(%rdi),%ymm11
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
shuffle1 4,5,3,5
shuffle1 6,7,4,7
shuffle1 8,9,6,9
shuffle1 10,11,8,11
shuffle2 3,4,10,4
shuffle2 6,8,3,8
shuffle2 5,7,6,7
shuffle2 9,11,5,11
shuffle4 10,3,9,3
shuffle4 6,5,10,5
shuffle4 4,8,6,8
shuffle4 7,11,4,11
shuffle8 9,10,7,10
shuffle8 6,4,9,4
shuffle8 3,5,6,5
shuffle8 8,11,3,11
#store
vmovdqa %ymm7,(%rdi)
vmovdqa %ymm9,32(%rdi)
vmovdqa %ymm6,64(%rdi)
vmovdqa %ymm3,96(%rdi)
vmovdqa %ymm10,128(%rdi)
vmovdqa %ymm4,160(%rdi)
vmovdqa %ymm5,192(%rdi)
vmovdqa %ymm11,224(%rdi)
ret
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
*/
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
.text
nttunpack128_avx:
#load
vmovdqa (%rdi),%ymm4
vmovdqa 32(%rdi),%ymm5
vmovdqa 64(%rdi),%ymm6
vmovdqa 96(%rdi),%ymm7
vmovdqa 128(%rdi),%ymm8
vmovdqa 160(%rdi),%ymm9
vmovdqa 192(%rdi),%ymm10
vmovdqa 224(%rdi),%ymm11
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
shuffle8 4,8,3,8
shuffle8 5,9,4,9
shuffle8 6,10,5,10
shuffle8 7,11,6,11
shuffle4 3,5,7,5
shuffle4 8,10,3,10
shuffle4 4,6,8,6
shuffle4 9,11,4,11
shuffle2 7,8,9,8
shuffle2 5,6,7,6
shuffle2 3,4,5,4
shuffle2 10,11,3,11
shuffle1 9,5,10,5
shuffle1 8,4,9,4
shuffle1 7,3,8,3
shuffle1 6,11,7,11
#store
vmovdqa %ymm10,(%rdi)
vmovdqa %ymm5,32(%rdi)
vmovdqa %ymm9,64(%rdi)
vmovdqa %ymm4,96(%rdi)
vmovdqa %ymm8,128(%rdi)
vmovdqa %ymm3,160(%rdi)
vmovdqa %ymm7,192(%rdi)
vmovdqa %ymm11,224(%rdi)
ret
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
.global cdecl(PQCLEAN_KYBER512_AVX2_nttunpack_avx)
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
.global _cdecl(PQCLEAN_KYBER512_AVX2_nttunpack_avx)
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
cdecl(PQCLEAN_KYBER512_AVX2_nttunpack_avx):
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
_cdecl(PQCLEAN_KYBER512_AVX2_nttunpack_avx):
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
call nttunpack128_avx
add $256,%rdi
call nttunpack128_avx
ret
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
ntttobytes128_avx:
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
#load
vmovdqa (%rsi),%ymm5
vmovdqa 32(%rsi),%ymm6
vmovdqa 64(%rsi),%ymm7
vmovdqa 96(%rsi),%ymm8
vmovdqa 128(%rsi),%ymm9
vmovdqa 160(%rsi),%ymm10
vmovdqa 192(%rsi),%ymm11
vmovdqa 224(%rsi),%ymm12
#csubq
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
csubq 5,13
Round 3 Kyber
2020-10-27 13:48:42 +00:00
csubq 6,13
csubq 7,13
csubq 8,13
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
csubq 9,13
Round 3 Kyber
2020-10-27 13:48:42 +00:00
csubq 10,13
csubq 11,13
csubq 12,13
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
#bitpack
vpsllw $12,%ymm6,%ymm4
vpor %ymm4,%ymm5,%ymm4
vpsrlw $4,%ymm6,%ymm5
vpsllw $8,%ymm7,%ymm6
vpor %ymm5,%ymm6,%ymm5
vpsrlw $8,%ymm7,%ymm6
vpsllw $4,%ymm8,%ymm7
vpor %ymm6,%ymm7,%ymm6
vpsllw $12,%ymm10,%ymm7
vpor %ymm7,%ymm9,%ymm7
vpsrlw $4,%ymm10,%ymm8
vpsllw $8,%ymm11,%ymm9
vpor %ymm8,%ymm9,%ymm8
vpsrlw $8,%ymm11,%ymm9
vpsllw $4,%ymm12,%ymm10
vpor %ymm9,%ymm10,%ymm9
shuffle1 4,5,3,5
shuffle1 6,7,4,7
shuffle1 8,9,6,9
shuffle2 3,4,8,4
shuffle2 6,5,3,5
shuffle2 7,9,6,9
shuffle4 8,3,7,3
shuffle4 6,4,8,4
shuffle4 5,9,6,9
shuffle8 7,8,5,8
shuffle8 6,3,7,3
shuffle8 4,9,6,9
#store
vmovdqu %ymm5,(%rdi)
vmovdqu %ymm7,32(%rdi)
vmovdqu %ymm6,64(%rdi)
vmovdqu %ymm8,96(%rdi)
vmovdqu %ymm3,128(%rdi)
vmovdqu %ymm9,160(%rdi)
ret
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
.global cdecl(PQCLEAN_KYBER512_AVX2_ntttobytes_avx)
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
.global _cdecl(PQCLEAN_KYBER512_AVX2_ntttobytes_avx)
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
cdecl(PQCLEAN_KYBER512_AVX2_ntttobytes_avx):
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
_cdecl(PQCLEAN_KYBER512_AVX2_ntttobytes_avx):
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
#consts
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
vmovdqa _16XQ*2(%rdx),%ymm0
call ntttobytes128_avx
add $256,%rsi
add $192,%rdi
call ntttobytes128_avx
ret
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
nttfrombytes128_avx:
Adds AVX2 variants of Kyber512, Kyber768, Kyber1024 (#225) * Integrate Kyber-AVX2 into PQClean * Fix types and formatting in Kyber * Workaround a valgrind crash * Remove comment in shuffle.s * Remove some extraneous truncations * fixup! Fix types and formatting in Kyber
2019-09-10 10:45:01 +01:00
#load
vmovdqu (%rsi),%ymm4
vmovdqu 32(%rsi),%ymm5
vmovdqu 64(%rsi),%ymm6
vmovdqu 96(%rsi),%ymm7
vmovdqu 128(%rsi),%ymm8
vmovdqu 160(%rsi),%ymm9
shuffle8 4,7,3,7
shuffle8 5,8,4,8
shuffle8 6,9,5,9
shuffle4 3,8,6,8
shuffle4 7,5,3,5
shuffle4 4,9,7,9
shuffle2 6,5,4,5
shuffle2 8,7,6,7
shuffle2 3,9,8,9
shuffle1 4,7,10,7
shuffle1 5,8,4,8
shuffle1 6,9,5,9
#bitunpack
vpsrlw $12,%ymm10,%ymm11
vpsllw $4,%ymm7,%ymm12
vpor %ymm11,%ymm12,%ymm11
vpand %ymm0,%ymm10,%ymm10
vpand %ymm0,%ymm11,%ymm11
vpsrlw $8,%ymm7,%ymm12
vpsllw $8,%ymm4,%ymm13
vpor %ymm12,%ymm13,%ymm12
vpand %ymm0,%ymm12,%ymm12
vpsrlw $4,%ymm4,%ymm13
vpand %ymm0,%ymm13,%ymm13
vpsrlw $12,%ymm8,%ymm14
vpsllw $4,%ymm5,%ymm15
vpor %ymm14,%ymm15,%ymm14
vpand %ymm0,%ymm8,%ymm8
vpand %ymm0,%ymm14,%ymm14
vpsrlw $8,%ymm5,%ymm15
vpsllw $8,%ymm9,%ymm1
vpor %ymm15,%ymm1,%ymm15
vpand %ymm0,%ymm15,%ymm15
vpsrlw $4,%ymm9,%ymm1
vpand %ymm0,%ymm1,%ymm1
#store
vmovdqa %ymm10,(%rdi)
vmovdqa %ymm11,32(%rdi)
vmovdqa %ymm12,64(%rdi)
vmovdqa %ymm13,96(%rdi)
vmovdqa %ymm8,128(%rdi)
vmovdqa %ymm14,160(%rdi)
vmovdqa %ymm15,192(%rdi)
vmovdqa %ymm1,224(%rdi)
ret
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
.global cdecl(PQCLEAN_KYBER512_AVX2_nttfrombytes_avx)
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
.global _cdecl(PQCLEAN_KYBER512_AVX2_nttfrombytes_avx)
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
cdecl(PQCLEAN_KYBER512_AVX2_nttfrombytes_avx):
Prepare for round 3 update with packaging script
2020-10-27 00:05:07 +00:00
_cdecl(PQCLEAN_KYBER512_AVX2_nttfrombytes_avx):
Update Kyber from upstream Makes Kyber-AVX run on MacOS (#251)
2020-07-31 07:17:42 +01:00
#consts
vmovdqa _16XMASK*2(%rdx),%ymm0
call nttfrombytes128_avx
add $256,%rdi
add $192,%rsi
call nttfrombytes128_avx
ret
Reference in New Issue Copy Permalink