void KeccakP1600times4_ExtractLanesAll(const void *states, unsigned char *data, unsigned int laneCount, unsigned int laneOffset)
{
const V256 *stateAsLanes256 = states;
const V512 *stateAsLanes512 = states;
UINT64 *dataAsLanes = (UINT64 *)data;
unsigned int i;
V256 index512;
V128 index256;
#define Extr1( argIndex ) STORE_SCATTER4_64(dataAsLanes+argIndex, index256, stateAsLanes256[argIndex])
#define Extr2( argIndex ) STORE_SCATTER8_64(dataAsLanes+argIndex, index512, stateAsLanes512[argIndex/2])
index256 = LOAD4_32(3*laneOffset, 2*laneOffset, 1*laneOffset, 0*laneOffset);
index512 = LOAD8_32(3*laneOffset+1, 2*laneOffset+1, 1*laneOffset+1, 0*laneOffset+1, 3*laneOffset, 2*laneOffset, 1*laneOffset, 0*laneOffset);
if ( laneCount >= 16 ) {
Extr2( 0 );
Extr2( 2 );
Extr2( 4 );
Extr2( 6 );
Extr2( 8 );
Extr2( 10 );
Extr2( 12 );
Extr2( 14 );
if ( laneCount >= 20 ) {
Extr2( 16 );
Extr2( 18 );
for(i=20; i<laneCount; i++)
Extr1( i );
}
else {
for(i=16; i<laneCount; i++)
Extr1( i );
}
}
else {
for(i=0; i<laneCount; i++)
Extr1( i );
}
#undef Extr1
#undef Extr2
}
void KeccakF1600_Pl2_ExtractLanesAll(const void *states, unsigned char *data, unsigned int laneCount, unsigned int laneOffset)
{
const V128 *stateAsLanes = states;
V128 lanes;
unsigned int i;
UINT64 *curData0 = (UINT64 *)data;
UINT64 *curData1 = (UINT64 *)(data+laneOffset*SnP_laneLengthInBytes);
#define Extr( argIndex ) lanes = LOAD128( stateAsLanes[argIndex] ), \
STORE64L( curData0[argIndex], lanes ), \
STORE64H( curData1[argIndex], lanes )
#if defined(UseSSE2)
#define Extr2( argIndex ) lanes0 = LOAD128( stateAsLanes[argIndex] ), \
lanes1 = LOAD128( stateAsLanes[(argIndex)+1] ), \
lanes = UNPACKL( lanes0, lanes1 ), \
lanes0 = UNPACKH( lanes0, lanes1 ), \
STORE128u( *(V128*)&curData0[argIndex], lanes ), \
STORE128u( *(V128*)&curData1[argIndex], lanes0 )
if ( laneCount >= 16 ) {
V128 lanes0, lanes1;
Extr2( 0 );
Extr2( 2 );
Extr2( 4 );
Extr2( 6 );
Extr2( 8 );
Extr2( 10 );
Extr2( 12 );
Extr2( 14 );
if ( laneCount >= 20 ) {
Extr2( 16 );
Extr2( 18 );
for(i=20; i<laneCount; i++)
Extr( i );
}
else {
for(i=16; i<laneCount; i++)
Extr( i );
}
}
#undef Extr2
#else
if ( laneCount >= 17 ) {
Extr( 0 );
Extr( 1 );
Extr( 2 );
Extr( 3 );
Extr( 4 );
Extr( 5 );
Extr( 6 );
Extr( 7 );
Extr( 8 );
Extr( 9 );
Extr( 10 );
Extr( 11 );
Extr( 12 );
Extr( 13 );
Extr( 14 );
Extr( 15 );
Extr( 16 );
if ( laneCount >= 21 ) {
Extr( 17 );
Extr( 18 );
Extr( 19 );
Extr( 20 );
for(i=21; i<laneCount; i++)
Extr( i );
}
else {
for(i=17; i<laneCount; i++)
Extr( i );
}
}
#endif
else {
for(i=0; i<laneCount; i++)
Extr( i );
}
#undef Extr
}