Files
ispc/examples_ptx/radixSort/radixSort.ispc
Evghenii 659573338c +1
2014-01-28 16:43:00 +01:00

315 lines
8.1 KiB
Plaintext

#define NUMBITS 8
#define NUMDIGITS (1<<NUMBITS)
task
void countPass(
const uniform int keysAll[],
const uniform int bit,
const uniform int numElements,
uniform int countsAll[],
uniform int countsGlobal[])
{
const uniform int blockIdx = taskIndex;
const uniform int numBlocks = taskCount;
const uniform int blockDim = (numElements + numBlocks - 1) / numBlocks;
const uniform int mask = (1 << NUMBITS) - 1;
const uniform int * uniform keys = keysAll + blockIdx*blockDim;
uniform int * uniform counts = countsAll + blockIdx*NUMDIGITS;
const uniform int nloc = min(numElements - blockIdx*blockDim, blockDim);
#if 0
uniform int shcounts[NUMDIGITS];
#else
#define shcounts counts
#endif
foreach (digit = 0 ... NUMDIGITS)
shcounts[digit] = 0;
foreach (i = 0 ... nloc)
{
const int key = mask & ((unsigned int)keys[i] >> bit);
uniform int skey;
if (reduce_equal(key, &skey) == true)
shcounts[skey] += reduce_add(1);
else
{
#ifdef __NVPTX__
atomic_add_global(&shcounts[key], 1);
#else
atomic_add_local(&shcounts[key], 1);
#endif
}
}
foreach (digit = 0 ... NUMDIGITS)
{
counts[digit] = shcounts[digit];
atomic_add_global(&countsGlobal[digit], shcounts[digit]);
}
}
task
void sortPass(
uniform int keysAll[],
uniform int sorted[],
uniform int bit,
uniform int numElements,
uniform int digitOffsetsAll[])
{
const uniform int blockIdx = taskIndex;
const uniform int numBlocks = taskCount;
const uniform int blockDim = (numElements + numBlocks - 1) / numBlocks;
uniform int localCounts[NUMDIGITS];
const uniform int keyIndex = blockIdx * blockDim;
uniform int * uniform keys = keysAll + keyIndex;
uniform int * uniform digitOffsets = digitOffsetsAll + blockIdx*NUMDIGITS;
const uniform int nloc = min(numElements - keyIndex, blockDim);
const int unitScan = exclusive_scan_add(1);
const uniform int mask = (1 << NUMBITS) - 1;
foreach (i = 0 ... NUMDIGITS)
localCounts[i] = 0;
foreach (i = 0 ... nloc)
{
const int key = mask & ((unsigned int)keys[i] >> bit);
int rel;
if (reduce_equal(key) == true)
{
rel = localCounts[key] + unitScan;
localCounts[key] = rel+1;
}
else
{
#ifdef __NVPTX__
rel = atomic_add_global(&localCounts[key],1);
#else
rel = atomic_add_local(&localCounts[key],1);
#endif
}
const int scatter = rel + digitOffsets[key];
sorted [scatter] = keys[i];
}
}
task
void partialScanLocal(
uniform int numBlocks,
uniform int excScanAll[],
uniform int countsAll[],
uniform int partialSumAll[])
{
const uniform int blockIdx = taskIndex;
const uniform int blockDim = (numBlocks+taskCount-1)/taskCount;
const uniform int bbeg = blockIdx * blockDim;
const uniform int bend = min(bbeg + blockDim, numBlocks);
uniform int (* uniform countsBlock)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])countsAll;
uniform int (* uniform excScanBlock)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])excScanAll;
uniform int (* uniform partialSum)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])partialSumAll;
foreach (digit = 0 ... NUMDIGITS)
{
int prev = bbeg == 0 ? excScanBlock[0][digit] : 0;
for (uniform int block = bbeg; block < bend; block++)
{
const int y = countsBlock[block][digit];
excScanBlock[block][digit] = prev;
prev += y;
}
partialSum[blockIdx][digit] = excScanBlock[bend-1][digit] + countsBlock[bend-1][digit];
}
}
task
void partialScanGlobal(
const uniform int numBlocks,
uniform int partialSumAll[],
uniform int prefixSumAll[])
{
uniform int (* uniform partialSum)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])partialSumAll;
uniform int (* uniform prefixSum)[NUMDIGITS] = (uniform int (*)[NUMDIGITS]) prefixSumAll;
const uniform int digit = taskIndex;
int carry = 0;
foreach (block = 0 ... numBlocks)
{
const int value = partialSum[block][digit];
const int scan = exclusive_scan_add(value);
prefixSum[block][digit] = scan + carry;
carry += broadcast(scan+value, programCount-1);
}
}
task
void completeScanGlobal(
uniform int numBlocks,
uniform int excScanAll[],
uniform int carryValueAll[])
{
const uniform int blockIdx = taskIndex;
const uniform int blockDim = (numBlocks+taskCount-1)/taskCount;
const uniform int bbeg = blockIdx * blockDim;
const uniform int bend = min(bbeg + blockDim, numBlocks);
uniform int (* uniform excScanBlock)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])excScanAll;
uniform int (* uniform carryValue)[NUMDIGITS] = (uniform int (*)[NUMDIGITS])carryValueAll;
foreach (digit = 0 ... NUMDIGITS)
{
const int carry = carryValue[blockIdx][digit];
for (uniform int block = bbeg; block < bend; block++)
excScanBlock[block][digit] += carry;
}
}
static
inline void radixExclusiveScan(
const uniform int numBlocks,
uniform int excScanPtr[],
uniform int countsPtr[],
uniform int partialSum[],
uniform int prefixSum[])
{
const uniform int scale = 2;
launch [numBlocks/scale] partialScanLocal(numBlocks, excScanPtr, countsPtr, partialSum);
sync;
launch [NUMDIGITS] partialScanGlobal(numBlocks/scale, partialSum, prefixSum);
sync;
launch [numBlocks/scale] completeScanGlobal(numBlocks, excScanPtr, prefixSum);
sync;
}
static uniform int * uniform memoryPool = NULL;
static uniform int numBlocks;
static uniform int nCountsGlobal;
static uniform int nExcScan;
static uniform int nCountsBlock;
static uniform int nPartialSum;
static uniform int nPrefixSum;
static uniform int * uniform countsGlobal;
static uniform int * uniform excScan;
static uniform int * uniform counts;
static uniform int * uniform partialSum;
static uniform int * uniform prefixSum;
static uniform int numElementsBuf = 0;
static uniform int * uniform bufKeys;
export void radixSort_alloc(const uniform int n)
{
assert(memoryPool == NULL);
numBlocks = num_cores()*4;
nCountsGlobal = NUMDIGITS;
nExcScan = NUMDIGITS*numBlocks;
nCountsBlock = NUMDIGITS*numBlocks;
nPartialSum = NUMDIGITS*numBlocks;
nPrefixSum = NUMDIGITS*numBlocks;
const uniform int nalloc =
nCountsGlobal +
nExcScan +
nCountsBlock +
nPartialSum +
nPrefixSum;
memoryPool = uniform new uniform int[nalloc];
countsGlobal = memoryPool;
excScan = countsGlobal + nCountsGlobal;
counts = excScan + nExcScan;
partialSum = counts + nCountsBlock;
prefixSum = partialSum + nPartialSum;
}
static
void radixSort_freeBufKeys()
{
if (numElementsBuf > 0)
{
delete bufKeys;
numElementsBuf = 0;
}
}
export void radixSort_free()
{
assert(memoryPool != NULL);
delete memoryPool;
memoryPool = NULL;
radixSort_freeBufKeys;
}
export void radixSort(
const uniform int numElements,
uniform int keys[])
{
#ifdef __NVPTX__
assert((numBlocks & 3) == 0); /* task granularity on Kepler is 4 */
#endif
if (numElementsBuf < numElements)
radixSort_freeBufKeys();
if (numElementsBuf == 0)
{
numElementsBuf = numElements;
bufKeys = uniform new uniform int[numElementsBuf];
}
const uniform int blockDim = (numElements + numBlocks - 1) / numBlocks;
for (uniform int bit = 0; bit < 32; bit += NUMBITS)
{
/* initialize histogram for each digit */
foreach (digit = 0 ... NUMDIGITS)
countsGlobal[digit] = 0;
/* compute histogram for each digit */
launch [numBlocks] countPass(keys, bit, numElements, counts, countsGlobal);
sync;
/* exclusive scan on global histogram */
int carry = 0;
excScan[0] = 0;
foreach (digit = 0 ... NUMDIGITS)
{
const int value = countsGlobal[digit];
const int scan = exclusive_scan_add(value);
excScan[digit] = scan + carry;
carry += broadcast(scan+value, programCount-1);
}
/* computing offsets for each digit */
radixExclusiveScan(numBlocks, excScan, counts, partialSum, prefixSum);
/* sorting */
launch [numBlocks]
sortPass(
keys,
bufKeys,
bit,
numElements,
excScan);
sync;
uniform int * uniform tmp = keys;
keys = bufKeys;
bufKeys = tmp;
}
}