#include "cdecl.h"
.include "shuffle.inc"

.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

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

#store
vmovdqa		%ymm9,(%rdi)
vmovdqa		%ymm8,32(%rdi)
vmovdqa		%ymm7,64(%rdi)
vmovdqa		%ymm6,96(%rdi)
vmovdqa		%ymm5,128(%rdi)
vmovdqa		%ymm4,160(%rdi)
vmovdqa		%ymm3,192(%rdi)
vmovdqa		%ymm11,224(%rdi)

ret

.global cdecl(PQCLEAN_DILITHIUM2_AVX2_nttunpack_avx)
.global _cdecl(PQCLEAN_DILITHIUM2_AVX2_nttunpack_avx)
cdecl(PQCLEAN_DILITHIUM2_AVX2_nttunpack_avx):
_cdecl(PQCLEAN_DILITHIUM2_AVX2_nttunpack_avx):
call		nttunpack128_avx
add		$256,%rdi
call		nttunpack128_avx
add		$256,%rdi
call		nttunpack128_avx
add		$256,%rdi
call		nttunpack128_avx
ret