static
void groestl_small_rounds(uint8_t *m, uint8_t q){
uint8_t r,i,j;
uint8_t tmp[8];
for(r=0; r<ROUNDS; ++r){
if(q){
for(i=0; i<64/4; ++i){
((uint32_t*)m)[i] ^= 0xffffffff;
}
for(i=0;i<8; ++i){
m[7+i*8] ^= r ^ (i<<4);
}
}else{
for(i=0;i<8; ++i){
m[i*8] ^= r ^ (i<<4);
}
}
#if DEBUG
if(r<2){
cli_putstr("\r\npost add-const");
dump_m(m);
}
#endif
for(i=0;i<8*8; ++i){
m[i] = aes_sbox[m[i]];
}
if(q){
shift_columns(m, q_shifts);
}else{
shift_columns(m, p_shifts);
}
#if DEBUG
if(r<2){
cli_putstr("\r\npost shift-bytes");
dump_m(m);
}
#endif
for(i=0; i<8; ++i){
memcpy(tmp, m+8*i, 8);
for(j=0; j<8; ++j){
m[j+i*8] = gf256mul(matrix[8*j+0],tmp[0], POLYNOM)
^ gf256mul(matrix[8*j+1],tmp[1], POLYNOM)
^ gf256mul(matrix[8*j+2],tmp[2], POLYNOM)
^ gf256mul(matrix[8*j+3],tmp[3], POLYNOM)
^ gf256mul(matrix[8*j+4],tmp[4], POLYNOM)
^ gf256mul(matrix[8*j+5],tmp[5], POLYNOM)
^ gf256mul(matrix[8*j+6],tmp[6], POLYNOM)
^ gf256mul(matrix[8*j+7],tmp[7], POLYNOM);
}
}
#if DEBUG
if(r<2){
cli_putstr("\r\npost mix-bytes");
dump_m(m);
}
#endif
}
}
void groestl_small_ctx2hash(void* dest, const groestl_small_ctx_t* ctx, uint16_t outlength_b){
uint8_t tmp[64];
memcpy(tmp, ctx->h, 64);
groestl_small_rounds(tmp, 0);
memxor(tmp, ctx->h, 64);
#if DEBUG
cli_putstr("\r\npost finalisation");
dump_m(tmp);
#endif
memcpy(dest, tmp+64-outlength_b/8, outlength_b/8);
}
void groestl_small_rounds(uint8_t *m, uint8_t q){
uint8_t r,i,j;
uint8_t tmp[8];
#if DEBUG
cli_putstr_P(PSTR("\r\n:: BEGIN "));
cli_putc(q?'Q':'P');
#endif
for(r=0; r<ROUNDS; ++r){
if(q){
for(i=0; i<8*8; ++i){
m[i] ^= 0xff;
}
for(i=0; i<8; ++i){
m[7+i*8] ^= r ^ (i<<4);
}
}else{
for(i=0; i<8; ++i){
m[i*8] ^= r ^ (i<<4);
}
}
#if DEBUG
// if(r<2){
cli_putstr_P(PSTR("\r\npost add-const"));
dump_m(m);
// }
#endif
for(i=0;i<8*8; ++i){
m[i] = pgm_read_byte(aes_sbox+m[i]);
}
if(!q){
shift_columns(m, p_shifts);
}else{
shift_columns(m, q_shifts);
}
#if DEBUG
if(r<2){
cli_putstr_P(PSTR("\r\npost shift-bytes"));
dump_m(m);
}
#endif
for(i=0; i<8; ++i){
memcpy(tmp, m+8*i, 8);
for(j=0; j<8; ++j){
m[j+i*8] = gf256mul(pgm_read_byte(matrix+8*j+0),tmp[0], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+1),tmp[1], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+2),tmp[2], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+3),tmp[3], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+4),tmp[4], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+5),tmp[5], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+6),tmp[6], POLYNOM)
^ gf256mul(pgm_read_byte(matrix+8*j+7),tmp[7], POLYNOM);
}
}
#if DEBUG
if(r<2){
cli_putstr_P(PSTR("\r\npost mix-bytes"));
dump_m(m);
}
#endif
}
}
