size_t KeccakP1600times4_12rounds_FastLoop_Absorb(void *states, unsigned int laneCount, unsigned int laneOffsetParallel, unsigned int laneOffsetSerial, const unsigned char *data, size_t dataByteLen)
{
size_t dataMinimumSize = (laneOffsetParallel*3 + laneCount)*8;
if (laneCount == 21) {
#if (KeccakP1600times4_unrolling < 12)
unsigned int i;
#endif
const unsigned char *dataStart = data;
V256 *statesAsLanes = states;
const UINT64 *dataAsLanes = (const UINT64 *)data;
KeccakP_DeclareVars;
V128 index;
copyFromState(statesAsLanes);
index = LOAD4_32(3*laneOffsetParallel, 2*laneOffsetParallel, 1*laneOffsetParallel, 0*laneOffsetParallel);
while(dataByteLen >= dataMinimumSize) {
#define Add_In( argLane, argIndex ) argLane = XOR(argLane, LOAD_GATHER4_64(index, dataAsLanes+argIndex))
Add_In( _ba, 0 );
Add_In( _be, 1 );
Add_In( _bi, 2 );
Add_In( _bo, 3 );
Add_In( _bu, 4 );
Add_In( _ga, 5 );
Add_In( _ge, 6 );
Add_In( _gi, 7 );
Add_In( _go, 8 );
Add_In( _gu, 9 );
Add_In( _ka, 10 );
Add_In( _ke, 11 );
Add_In( _ki, 12 );
Add_In( _ko, 13 );
Add_In( _ku, 14 );
Add_In( _ma, 15 );
Add_In( _me, 16 );
Add_In( _mi, 17 );
Add_In( _mo, 18 );
Add_In( _mu, 19 );
Add_In( _sa, 20 );
#undef Add_In
rounds12;
dataAsLanes += laneOffsetSerial;
dataByteLen -= laneOffsetSerial*8;
}
copyToState(statesAsLanes);
return (const unsigned char *)dataAsLanes - dataStart;
}
else {
const unsigned char *dataStart = data;
while(dataByteLen >= dataMinimumSize) {
KeccakP1600times4_AddLanesAll(states, data, laneCount, laneOffsetParallel);
KeccakP1600times4_PermuteAll_12rounds(states);
data += laneOffsetSerial*8;
dataByteLen -= laneOffsetSerial*8;
}
return data - dataStart;
}
}
void KeccakP1600times4_PermuteAll_12rounds(void *states)
{
V256 *statesAsLanes = states;
KeccakP_DeclareVars;
#if (KeccakP1600times4_unrolling < 12)
unsigned int i;
#endif
copyFromState(statesAsLanes);
rounds12;
copyToState(statesAsLanes);
}
void KeccakDoublePermutationOnWords(V128 *state)
{
declareABCDE
#if (Unrolling != 24)
unsigned int i;
#endif
copyFromState(A, state)
rounds
#if defined(UseMMX)
_mm_empty();
#endif
}