RadixSort& RadixSort::Sort(const uint32* input, uint32 nb, eRadixHint hint)
{
// Checkings
if(!input || !nb || nb&0x80000000) return *this;
// Stats
mTotalCalls++;
// Resize lists if needed
CheckResize(nb);
// Allocate histograms & offsets on the stack
uint32 Histogram[256*4];
uint32* Link[256];
// Create histograms (counters). Counters for all passes are created in one run.
// Pros: read input buffer once instead of four times
// Cons: mHistogram is 4Kb instead of 1Kb
// We must take care of signed/unsigned values for temporal coherence.... I just
// have 2 code paths even if just a single opcode changes. Self-modifying code, someone?
if(hint==RADIX_UNSIGNED) { CREATE_HISTOGRAMS(uint32, input); }
else { CREATE_HISTOGRAMS(int32, input); }
// Radix sort, j is the pass number (0=LSB, 3=MSB)
for(uint32 j=0;j<4;j++)
{
CHECK_PASS_VALIDITY(j);
// Sometimes the fourth (negative) pass is skipped because all numbers are negative and the MSB is 0xFF (for example). This is
// not a problem, numbers are correctly sorted anyway.
if(PerformPass)
{
// Should we care about negative values?
if(j!=3 || hint==RADIX_UNSIGNED)
{
// Here we deal with positive values only
// Create offsets
Link[0] = m_Ranks2;
for(uint32 i=1;i<256;i++) Link[i] = Link[i-1] + CurCount[i-1];
}
else
{
// This is a special case to correctly handle negative integers. They're sorted in the right order but at the wrong place.
#ifdef KYLE_HUBERT_VERSION
// From Kyle Hubert:
Link[128] = m_Ranks2;
for(uint32 i=129;i<256;i++) Link[i] = Link[i-1] + CurCount[i-1];
Link[0] = Link[255] + CurCount[255];
for(uint32 i=1;i<128;i++) Link[i] = Link[i-1] + CurCount[i-1];
#else
// Compute #negative values involved if needed
uint32 NbNegativeValues = 0;
if(hint==RADIX_SIGNED)
{
// An efficient way to compute the number of negatives values we'll have to deal with is simply to sum the 128
// last values of the last histogram. Last histogram because that's the one for the Most Significant Byte,
// responsible for the sign. 128 last values because the 128 first ones are related to positive numbers.
const uint32* h3 = &Histogram[H3_OFFSET];
for(uint32 i=128;i<256;i++) NbNegativeValues += h3[i]; // 768 for last histogram, 128 for negative part
}
// Create biased offsets, in order for negative numbers to be sorted as well
Link[0] = &m_Ranks2[NbNegativeValues]; // First positive number takes place after the negative ones
for(uint32 i=1;i<128;i++) Link[i] = Link[i-1] + CurCount[i-1]; // 1 to 128 for positive numbers
// Fixing the wrong place for negative values
Link[128] = m_Ranks2;
for(uint32 i=129;i<256;i++) Link[i] = Link[i-1] + CurCount[i-1];
#endif
}
// Perform Radix Sort
const unsigned char* InputBytes = (const unsigned char*)input;
InputBytes += BYTES_INC;
if(INVALID_RANKS)
{
for(uint32 i=0;i<nb;i++) *Link[InputBytes[i<<2]]++ = i;
VALIDATE_RANKS;
}
else
{
const uint32* Indices = m_Ranks;
const uint32* IndicesEnd = &m_Ranks[nb];
while(Indices!=IndicesEnd)
{
const uint32 id = *Indices++;
*Link[InputBytes[id<<2]]++ = id;
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
uint32* Tmp = m_Ranks;
m_Ranks = m_Ranks2;
m_Ranks2 = Tmp;
}
}
return *this;
}
RadixSort& RadixSort::Sort(const udword* input, udword nb, bool signedvalues)
{
// Checkings
if(!input || !nb) return *this;
// Stats
mTotalCalls++;
// Resize lists if needed
CHECK_RESIZE(nb);
#ifdef RADIX_LOCAL_RAM
// Allocate histograms & offsets on the stack
udword mHistogram[256*4];
udword mOffset[256];
#endif
// Create histograms (counters). Counters for all passes are created in one run.
// Pros: read input buffer once instead of four times
// Cons: mHistogram is 4Kb instead of 1Kb
// We must take care of signed/unsigned values for temporal coherence.... I just
// have 2 code paths even if just a single opcode changes. Self-modifying code, someone?
if(!signedvalues) { CREATE_HISTOGRAMS(udword, input); }
else { CREATE_HISTOGRAMS(sdword, input); }
// Compute #negative values involved if needed
udword NbNegativeValues = 0;
if(signedvalues)
{
// An efficient way to compute the number of negatives values we'll have to deal with is simply to sum the 128
// last values of the last histogram. Last histogram because that's the one for the Most Significant Byte,
// responsible for the sign. 128 last values because the 128 first ones are related to positive numbers.
udword* h3= &mHistogram[768];
for(udword i=128;i<256;i++) NbNegativeValues += h3[i]; // 768 for last histogram, 128 for negative part
}
// Radix sort, j is the pass number (0=LSB, 3=MSB)
for(udword j=0;j<4;j++)
{
CHECK_PASS_VALIDITY(j);
// Sometimes the fourth (negative) pass is skipped because all numbers are negative and the MSB is 0xFF (for example). This is
// not a problem, numbers are correctly sorted anyway.
if(PerformPass)
{
// Should we care about negative values?
if(j!=3 || !signedvalues)
{
// Here we deal with positive values only
// Create offsets
mOffset[0] = 0;
for(udword i=1;i<256;i++) mOffset[i] = mOffset[i-1] + CurCount[i-1];
}
else
{
// This is a special case to correctly handle negative integers. They're sorted in the right order but at the wrong place.
// Create biased offsets, in order for negative numbers to be sorted as well
mOffset[0] = NbNegativeValues; // First positive number takes place after the negative ones
for(udword i=1;i<128;i++) mOffset[i] = mOffset[i-1] + CurCount[i-1]; // 1 to 128 for positive numbers
// Fixing the wrong place for negative values
mOffset[128] = 0;
for(udword i=129;i<256;i++) mOffset[i] = mOffset[i-1] + CurCount[i-1];
}
// Perform Radix Sort
ubyte* InputBytes = (ubyte*)input;
udword* Indices = mIndices;
udword* IndicesEnd = &mIndices[nb];
InputBytes += j;
while(Indices!=IndicesEnd)
{
udword id = *Indices++;
mIndices2[mOffset[InputBytes[id<<2]]++] = id;
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mIndices after the swap.
udword* Tmp = mIndices; mIndices = mIndices2; mIndices2 = Tmp;
}
}
return *this;
}
RadixSort& RadixSort::Sort(const NxF32* input2, NxU32 nb)
{
// Checkings
if(!input2 || !nb || nb&0x80000000) return *this;
// Stats
mTotalCalls++;
NxU32* input = (NxU32*)input2;
// Resize lists if needed
CheckResize(nb);
// Allocate histograms & offsets on the stack
NxU32 mHistogram[256*4];
NxU32* mLink[256];
// Create histograms (counters). Counters for all passes are created in one run.
// Pros: read input buffer once instead of four times
// Cons: mHistogram is 4Kb instead of 1Kb
// Floating-point values are always supposed to be signed values, so there's only one code path there.
// Please note the floating point comparison needed for temporal coherence! Although the resulting asm code
// is dreadful, this is surprisingly not such a performance hit - well, I suppose that's a big one on first
// generation Pentiums....We can't make comparison on integer representations because, as Chris said, it just
// wouldn't work with mixed positive/negative values....
{ CREATE_HISTOGRAMS(NxF32, input2); }
// Compute #negative values involved if needed
NxU32 NbNegativeValues = 0;
// An efficient way to compute the number of negatives values we'll have to deal with is simply to sum the 128
// last values of the last histogram. Last histogram because that's the one for the Most Significant Byte,
// responsible for the sign. 128 last values because the 128 first ones are related to positive numbers.
// ### is that ok on Apple ?!
NxU32* h3= &mHistogram[768];
for(NxU32 i=128;i<256;i++) NbNegativeValues += h3[i]; // 768 for last histogram, 128 for negative part
// Radix sort, j is the pass number (0=LSB, 3=MSB)
for(NxU32 j=0;j<4;j++)
{
// Should we care about negative values?
if(j!=3)
{
// Here we deal with positive values only
CHECK_PASS_VALIDITY(j);
if(PerformPass)
{
// Create offsets
mLink[0] = mRanks2;
for(NxU32 i=1;i<256;i++) mLink[i] = mLink[i-1] + CurCount[i-1];
// Perform Radix Sort
NxU8* InputBytes = (NxU8*)input;
InputBytes += BYTES_INC;
if(INVALID_RANKS)
{
for(NxU32 i=0;i<nb;i++) *mLink[InputBytes[i<<2]]++ = i;
VALIDATE_RANKS;
}
else
{
NxU32* Indices = mRanks;
NxU32* IndicesEnd = &mRanks[nb];
while(Indices!=IndicesEnd)
{
NxU32 id = *Indices++;
*mLink[InputBytes[id<<2]]++ = id;
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
NxU32* Tmp = mRanks; mRanks = mRanks2; mRanks2 = Tmp;
}
}
else
{
// This is a special case to correctly handle negative values
CHECK_PASS_VALIDITY(j);
if(PerformPass)
{
// Create biased offsets, in order for negative numbers to be sorted as well
mLink[0] = &mRanks2[NbNegativeValues]; // First positive number takes place after the negative ones
for(NxU32 i=1;i<128;i++) mLink[i] = mLink[i-1] + CurCount[i-1]; // 1 to 128 for positive numbers
// We must reverse the sorting order for negative numbers!
mLink[255] = mRanks2;
for(NxU32 i=0;i<127;i++) mLink[254-i] = mLink[255-i] + CurCount[255-i]; // Fixing the wrong order for negative values
for(NxU32 i=128;i<256;i++) mLink[i] += CurCount[i]; // Fixing the wrong place for negative values
// Perform Radix Sort
if(INVALID_RANKS)
{
for(NxU32 i=0;i<nb;i++)
{
NxU32 Radix = input[i]>>24; // Radix byte, same as above. AND is useless here (NxU32).
// ### cmp to be killed. Not good. Later.
if(Radix<128) *mLink[Radix]++ = i; // Number is positive, same as above
else *(--mLink[Radix]) = i; // Number is negative, flip the sorting order
}
VALIDATE_RANKS;
}
else
{
for(NxU32 i=0;i<nb;i++)
{
NxU32 Radix = input[mRanks[i]]>>24; // Radix byte, same as above. AND is useless here (NxU32).
// ### cmp to be killed. Not good. Later.
if(Radix<128) *mLink[Radix]++ = mRanks[i]; // Number is positive, same as above
else *(--mLink[Radix]) = mRanks[i]; // Number is negative, flip the sorting order
}
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
NxU32* Tmp = mRanks; mRanks = mRanks2; mRanks2 = Tmp;
}
else
{
// The pass is useless, yet we still have to reverse the order of current list if all values are negative.
if(UniqueVal>=128)
{
if(INVALID_RANKS)
{
// ###Possible?
for(NxU32 i=0;i<nb;i++) mRanks2[i] = nb-i-1;
VALIDATE_RANKS;
}
else
{
for(NxU32 i=0;i<nb;i++) mRanks2[i] = mRanks[nb-i-1];
}
// Swap pointers for next pass. Valid indices - the most recent ones - are in mRanks after the swap.
NxU32* Tmp = mRanks; mRanks = mRanks2; mRanks2 = Tmp;
}
}
}
