.global PQCLEAN_DILITHIUM4_AVX2_reduce_avx
PQCLEAN_DILITHIUM4_AVX2_reduce_avx:
#consts
vmovdqa		_PQCLEAN_DILITHIUM4_AVX2_8x23ones(%rip),%ymm0

xor		%eax,%eax
_looptop_rdc32:
#load
vmovdqa		(%rdi),%ymm1
vmovdqa		32(%rdi),%ymm3
vmovdqa		64(%rdi),%ymm5
vmovdqa		96(%rdi),%ymm7

#reduce
vpsrld		$23,%ymm1,%ymm2
vpsrld		$23,%ymm3,%ymm4
vpsrld		$23,%ymm5,%ymm6
vpsrld		$23,%ymm7,%ymm8
vpand		%ymm0,%ymm1,%ymm1
vpand		%ymm0,%ymm3,%ymm3
vpand		%ymm0,%ymm5,%ymm5
vpand		%ymm0,%ymm7,%ymm7
vpsubd		%ymm2,%ymm1,%ymm1
vpsubd		%ymm4,%ymm3,%ymm3
vpsubd		%ymm6,%ymm5,%ymm5
vpsubd		%ymm8,%ymm7,%ymm7
vpslld		$13,%ymm2,%ymm2
vpslld		$13,%ymm4,%ymm4
vpslld		$13,%ymm6,%ymm6
vpslld		$13,%ymm8,%ymm8
vpaddd		%ymm2,%ymm1,%ymm1
vpaddd		%ymm4,%ymm3,%ymm3
vpaddd		%ymm6,%ymm5,%ymm5
vpaddd		%ymm8,%ymm7,%ymm7

#store
vmovdqa		%ymm1,(%rdi)
vmovdqa		%ymm3,32(%rdi)
vmovdqa		%ymm5,64(%rdi)
vmovdqa		%ymm7,96(%rdi)

add		$128,%rdi
add		$1,%eax
cmp		$8,%eax
jb _looptop_rdc32

ret

.global PQCLEAN_DILITHIUM4_AVX2_csubq_avx
PQCLEAN_DILITHIUM4_AVX2_csubq_avx:
#consts
vmovdqa		_PQCLEAN_DILITHIUM4_AVX2_8xq(%rip),%ymm0

xor		%eax,%eax
_looptop_csubq:
#load
vmovdqa		(%rdi),%ymm1
vmovdqa		32(%rdi),%ymm3
vmovdqa		64(%rdi),%ymm5
vmovdqa		96(%rdi),%ymm7

#PQCLEAN_DILITHIUM4_AVX2_csubq
vpsubd		%ymm0,%ymm1,%ymm1
vpsubd		%ymm0,%ymm3,%ymm3
vpsubd		%ymm0,%ymm5,%ymm5
vpsubd		%ymm0,%ymm7,%ymm7
vpsrad		$31,%ymm1,%ymm2
vpsrad		$31,%ymm3,%ymm4
vpsrad		$31,%ymm5,%ymm6
vpsrad		$31,%ymm7,%ymm8
vpand		%ymm0,%ymm2,%ymm2
vpand		%ymm0,%ymm4,%ymm4
vpand		%ymm0,%ymm6,%ymm6
vpand		%ymm0,%ymm8,%ymm8
vpaddd		%ymm2,%ymm1,%ymm1
vpaddd		%ymm4,%ymm3,%ymm3
vpaddd		%ymm6,%ymm5,%ymm5
vpaddd		%ymm8,%ymm7,%ymm7

#store
vmovdqa		%ymm1,(%rdi)
vmovdqa		%ymm3,32(%rdi)
vmovdqa		%ymm5,64(%rdi)
vmovdqa		%ymm7,96(%rdi)

add		$128,%rdi
add		$1,%eax
cmp		$8,%eax
jb _looptop_csubq

ret