void KeccakP1600times4_OverwriteLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset)
{
V256 *stateAsLanes = states;
unsigned int i;
const UINT64 *curData0 = (const UINT64 *)data;
const UINT64 *curData1 = (const UINT64 *)(data+laneOffset*SnP_laneLengthInBytes);
const UINT64 *curData2 = (const UINT64 *)(data+laneOffset*2*SnP_laneLengthInBytes);
const UINT64 *curData3 = (const UINT64 *)(data+laneOffset*3*SnP_laneLengthInBytes);
V256 lanes0, lanes1, lanes2, lanes3, lanesL01, lanesL23, lanesH01, lanesH23;
#define OverWr( argIndex ) STORE256(stateAsLanes[argIndex], LOAD4_64(curData3[argIndex], curData2[argIndex], curData1[argIndex], curData0[argIndex]))
#define OverWr4( argIndex ) lanes0 = LOAD256u( curData0[argIndex]),\
lanes1 = LOAD256u( curData1[argIndex]),\
lanes2 = LOAD256u( curData2[argIndex]),\
lanes3 = LOAD256u( curData3[argIndex]),\
INTLEAVE(),\
STORE256( stateAsLanes[argIndex+0], lanes0 ),\
STORE256( stateAsLanes[argIndex+1], lanes1 ),\
STORE256( stateAsLanes[argIndex+2], lanes2 ),\
STORE256( stateAsLanes[argIndex+3], lanes3 )
if ( laneCount >= 16 ) {
OverWr4( 0 );
OverWr4( 4 );
OverWr4( 8 );
OverWr4( 12 );
if ( laneCount >= 20 ) {
OverWr4( 16 );
for(i=20; i<laneCount; i++)
OverWr( i );
}
else {
for(i=16; i<laneCount; i++)
OverWr( i );
}
}
else {
for(i=0; i<laneCount; i++)
OverWr( i );
}
#undef OverWr
#undef OverWr4
}
void KeccakF1600_Pl2_OverwriteLanesAll(void *states, const unsigned char *data, unsigned int laneCount, unsigned int laneOffset)
{
V128 *stateAsLanes = states;
unsigned int i;
const UINT64 *curData0 = (const UINT64 *)data;
const UINT64 *curData1 = (const UINT64 *)(data+laneOffset*SnP_laneLengthInBytes);
#define OverWr( argIndex ) STORE128(stateAsLanes[argIndex], LOAD6464(curData1[argIndex], curData0[argIndex]))
if ( laneCount >= 17 ) {
OverWr( 0 );
OverWr( 1 );
OverWr( 2 );
OverWr( 3 );
OverWr( 4 );
OverWr( 5 );
OverWr( 6 );
OverWr( 7 );
OverWr( 8 );
OverWr( 9 );
OverWr( 10 );
OverWr( 11 );
OverWr( 12 );
OverWr( 13 );
OverWr( 14 );
OverWr( 15 );
OverWr( 16 );
if ( laneCount >= 21 ) {
OverWr( 17 );
OverWr( 18 );
OverWr( 19 );
OverWr( 20 );
for(i=21; i<laneCount; i++)
OverWr( i );
}
else {
for(i=17; i<laneCount; i++)
OverWr( i );
}
}
else {
for(i=0; i<laneCount; i++)
OverWr( i );
}
#undef OverWr
}
