#define A3 T0
#define B3 T1
#define C3 T2
#define D3 T3
// Register and stack allocation for the AVX2 code
#define rsStoreAVX2 (0*32)(BP)
#define state1StoreAVX2 (1*32)(BP)
#define state2StoreAVX2 (2*32)(BP)
#define ctr0StoreAVX2 (3*32)(BP)
#define ctr1StoreAVX2 (4*32)(BP)
#define ctr2StoreAVX2 (5*32)(BP)
#define ctr3StoreAVX2 (6*32)(BP)
#define tmpStoreAVX2 (7*32)(BP) // 256 bytes on stack
#define AA0 Y0
#define AA1 Y5
#define AA2 Y6