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

+checkpoint

Browse Source
This commit is contained in:
Evghenii
2014-01-29 13:47:39 +01:00
parent f6379dea82
commit d4b46b1295
2 changed files with 3 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
examples_ptx/radixSort/radixSort.cu
View File

@@ -93,9 +93,9 @@ void sortPass(
const int mask = (1 << NUMBITS) - 1; const int mask = (1 << NUMBITS) - 1;
/* copy digit offset from Gmem to Lmem */ /* copy digit offset from Gmem to Lmem */
#if 0 #if 1
__shared__ int digitOffsets_sh[NUMDIGITS*4]; __shared__ int digitOffsets_sh[NUMDIGITS*4];
int *digitOffsets = digitOffsets_sh + warpIdx*NUMDIGITS; volatile int *digitOffsets = digitOffsets_sh + warpIdx*NUMDIGITS;
for (int digit = programIndex; digit < NUMDIGITS; digit += programCount) for (int digit = programIndex; digit < NUMDIGITS; digit += programCount)
digitOffsets[digit] = digitOffsetsAll[blkIdx*NUMDIGITS + digit]; digitOffsets[digit] = digitOffsetsAll[blkIdx*NUMDIGITS + digit];
#else #else
@@ -103,11 +103,11 @@ void sortPass(
#endif #endif
int scatter;
for (int i = programIndex; i < nloc; i += programCount) for (int i = programIndex; i < nloc; i += programCount)
if (i < nloc) if (i < nloc)
{ {
const int key = mask & ((unsigned int)keys[i] >> bit); const int key = mask & ((unsigned int)keys[i] >> bit);
int scatter;
/* not a vector friendly loop */ /* not a vector friendly loop */
#pragma unroll 1 /* needed, otherwise compiler unroll and optimizes the result :S */ #pragma unroll 1 /* needed, otherwise compiler unroll and optimizes the result :S */
for (int iv = 0; iv < programCount; iv++) for (int iv = 0; iv < programCount; iv++)

2
examples_ptx/radixSort/radixSort.ispc
View File

@@ -284,7 +284,6 @@ export void radixSort(
/* computing offsets for each digit */ /* computing offsets for each digit */
radixExclusiveScan(numBlocks, excScan, counts, partialSum, prefixSum); radixExclusiveScan(numBlocks, excScan, counts, partialSum, prefixSum);
#if 1
/* sorting */ /* sorting */
launch [numBlocks] launch [numBlocks]
sortPass( sortPass(
@@ -294,7 +293,6 @@ export void radixSort(
numElements, numElements,
excScan); excScan);
sync; sync;
#endif
} }
} }