Mirrors/bx
1
0
Fork 0
mirror of https://github.com/bkaradzic/bx.git synced 2026-02-18 04:53:06 +01:00
Code Issues Packages Projects Releases Wiki Activity

Added radix sort just for keys.

Browse Source
This commit is contained in:
Branimir Karadžić
2015-10-15 16:35:21 -07:00
parent 6a66432486
commit d84f2b2307
1 changed files with 124 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

126
include/bx/radixsort.h
View File

@@ -14,8 +14,69 @@ namespace bx
#define BX_RADIXSORT_HISTOGRAM_SIZE (1<<BX_RADIXSORT_BITS)
#define BX_RADIXSORT_BIT_MASK (BX_RADIXSORT_HISTOGRAM_SIZE-1)
inline void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, uint32_t _size)
{
uint32_t* __restrict keys = _keys;
uint32_t* __restrict tempKeys = _tempKeys;
uint32_t histogram[BX_RADIXSORT_HISTOGRAM_SIZE];
uint16_t shift = 0;
uint32_t pass = 0;
for (; pass < 3; ++pass)
{
memset(histogram, 0, sizeof(uint32_t)*BX_RADIXSORT_HISTOGRAM_SIZE);
bool sorted = true;
{
uint32_t key = keys[0];
uint32_t prevKey = key;
for (uint32_t ii = 0; ii < _size; ++ii, prevKey = key)
{
key = keys[ii];
uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
++histogram[index];
sorted &= prevKey <= key;
}
}
if (sorted)
{
goto done;
}
uint32_t offset = 0;
for (uint32_t ii = 0; ii < BX_RADIXSORT_HISTOGRAM_SIZE; ++ii)
{
uint32_t count = histogram[ii];
histogram[ii] = offset;
offset += count;
}
for (uint32_t ii = 0; ii < _size; ++ii)
{
uint32_t key = keys[ii];
uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
uint32_t dest = histogram[index]++;
tempKeys[dest] = key;
}
uint32_t* swapKeys = tempKeys;
tempKeys = keys;
keys = swapKeys;
shift += BX_RADIXSORT_BITS;
}
done:
if (0 != (pass&1) )
{
// Odd number of passes needs to do copy to the destination.
memcpy(_keys, _tempKeys, _size*sizeof(uint32_t) );
}
}
template <typename Ty>
void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
inline void radixSort32(uint32_t* __restrict _keys, uint32_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
{
uint32_t* __restrict keys = _keys;
uint32_t* __restrict tempKeys = _tempKeys;
@@ -87,8 +148,69 @@ done:
}
}
inline void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, uint32_t _size)
{
uint64_t* __restrict keys = _keys;
uint64_t* __restrict tempKeys = _tempKeys;
uint32_t histogram[BX_RADIXSORT_HISTOGRAM_SIZE];
uint16_t shift = 0;
uint32_t pass = 0;
for (; pass < 6; ++pass)
{
memset(histogram, 0, sizeof(uint32_t)*BX_RADIXSORT_HISTOGRAM_SIZE);
bool sorted = true;
{
uint64_t key = keys[0];
uint64_t prevKey = key;
for (uint32_t ii = 0; ii < _size; ++ii, prevKey = key)
{
key = keys[ii];
uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
++histogram[index];
sorted &= prevKey <= key;
}
}
if (sorted)
{
goto done;
}
uint32_t offset = 0;
for (uint32_t ii = 0; ii < BX_RADIXSORT_HISTOGRAM_SIZE; ++ii)
{
uint32_t count = histogram[ii];
histogram[ii] = offset;
offset += count;
}
for (uint32_t ii = 0; ii < _size; ++ii)
{
uint64_t key = keys[ii];
uint16_t index = (key>>shift)&BX_RADIXSORT_BIT_MASK;
uint32_t dest = histogram[index]++;
tempKeys[dest] = key;
}
uint64_t* swapKeys = tempKeys;
tempKeys = keys;
keys = swapKeys;
shift += BX_RADIXSORT_BITS;
}
done:
if (0 != (pass&1) )
{
// Odd number of passes needs to do copy to the destination.
memcpy(_keys, _tempKeys, _size*sizeof(uint64_t) );
}
}
template <typename Ty>
void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
inline void radixSort64(uint64_t* __restrict _keys, uint64_t* __restrict _tempKeys, Ty* __restrict _values, Ty* __restrict _tempValues, uint32_t _size)
{
uint64_t* __restrict keys = _keys;
uint64_t* __restrict tempKeys = _tempKeys;