aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

working. next step tuning

Browse Source
This commit is contained in:
Evghenii
2014-01-28 15:18:50 +01:00
parent 1b993e167f
commit f343e4cb0e
2 changed files with 84 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
examples_ptx/radixSort/radixSort.ispc
View File

@@ -158,7 +158,7 @@ inline void radixExclusiveScan(
uniform int partialSum[],
uniform int prefixSum[])
{
const uniform int scale = 4;
const uniform int scale = 2;
launch [numBlocks/scale] partialScanLocal(numBlocks, excScanPtr, countsPtr, partialSum);
sync;
@@ -169,25 +169,36 @@ inline void radixExclusiveScan(
sync;
}
export void radixSort(
const uniform int numElements,
uniform int keys[],
uniform int sorted[])
static uniform int * uniform memoryPool = NULL;
static uniform int numBlocks;
static uniform int nSharedCounts;
static uniform int nCountsGlobal;
static uniform int nExcScan;
static uniform int nCountsBlock;
static uniform int nPartialSum;
static uniform int nPrefixSum;
static uniform int * uniform sharedCounts;
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)
{
const uniform int numBlocks = num_cores()*4;
assert(memoryPool == NULL);
numBlocks = num_cores()*4;
nSharedCounts = NUMDIGITS*numBlocks;
nCountsGlobal = NUMDIGITS;
nExcScan = NUMDIGITS*numBlocks;
nCountsBlock = NUMDIGITS*numBlocks;
nPartialSum = NUMDIGITS*numBlocks;
nPrefixSum = NUMDIGITS*numBlocks;
#ifdef __NVPTX__
assert((numBlocks & 3) == 0); /* task granularity on Kepler is 4 */
#endif
const uniform int blockDim = (numElements + numBlocks - 1) / numBlocks;
const uniform int nSharedCounts = NUMDIGITS*numBlocks;
const uniform int nCountsGlobal = NUMDIGITS;
const uniform int nExcScan = NUMDIGITS*numBlocks;
const uniform int nCountsBlock = NUMDIGITS*numBlocks;
const uniform int nPartialSum = NUMDIGITS*numBlocks;
const uniform int nPrefixSum = NUMDIGITS*numBlocks;
const uniform int nalloc =
nSharedCounts +
@@ -197,14 +208,52 @@ export void radixSort(
nPartialSum +
nPrefixSum;
uniform int * uniform mem_pool = uniform new uniform int[nalloc];
memoryPool = uniform new uniform int[nalloc];
uniform int * uniform sharedCounts = mem_pool;
uniform int * uniform countsGlobal = sharedCounts + nSharedCounts;
uniform int * uniform excScan = countsGlobal + nCountsGlobal;
uniform int * uniform counts = excScan + nExcScan;
uniform int * uniform partialSum = counts + nCountsBlock;
uniform int * uniform prefixSum = partialSum + nPartialSum;
sharedCounts = memoryPool;
countsGlobal = sharedCounts + nSharedCounts;
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)
{
@@ -234,7 +283,7 @@ export void radixSort(
launch [numBlocks]
sortPass(
keys,
sorted,
bufKeys,
bit,
numElements,
excScan,
@@ -242,9 +291,8 @@ export void radixSort(
sync;
uniform int * uniform tmp = keys;
keys = sorted;
sorted = tmp;
keys = bufKeys;
bufKeys = tmp;
}
delete mem_pool;
}