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

fixed

Browse Source
This commit is contained in:
Evghenii
2014-01-31 20:13:10 +01:00
parent 11bc35eb6c
commit 98a275a382
1 changed files with 3 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
examples_ptx/mergeSort/mergeSort.ispc
View File

@@ -266,21 +266,17 @@ void generateSampleRanksKernel(
if (i < segmentSamplesA) if (i < segmentSamplesA)
{ {
ranksA[i] = i * SAMPLE_STRIDE; ranksA[i] = i * SAMPLE_STRIDE;
#if 0
ranksB[i] = binarySearchExclusive( ranksB[i] = binarySearchExclusive(
srcKey[i * SAMPLE_STRIDE], srcKey + stride, srcKey[i * SAMPLE_STRIDE], srcKey + stride,
segmentElementsB, nextPowerOfTwo(segmentElementsB)); segmentElementsB, nextPowerOfTwo(segmentElementsB));
#endif
} }
if (i < segmentSamplesB) if (i < segmentSamplesB)
{ {
ranksB[(stride / SAMPLE_STRIDE) + i] = i * SAMPLE_STRIDE; ranksB[(stride / SAMPLE_STRIDE) + i] = i * SAMPLE_STRIDE;
#if 0
ranksA[(stride / SAMPLE_STRIDE) + i] = binarySearchInclusive( ranksA[(stride / SAMPLE_STRIDE) + i] = binarySearchInclusive(
srcKey[stride + i * SAMPLE_STRIDE], srcKey + 0, srcKey[stride + i * SAMPLE_STRIDE], srcKey + 0,
segmentElementsA, nextPowerOfTwo(segmentElementsA)); segmentElementsA, nextPowerOfTwo(segmentElementsA));
#endif
} }
} }
} }
@@ -579,7 +575,7 @@ void mergeSort(
// cpu: 28 gpu: 74 M/s // cpu: 28 gpu: 74 M/s
{ {
// cpu: 356 gpu: 534 M/s // cpu: 356 gpu: 534 M/s
// mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount)); mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount));
#if 1 #if 1
for (uniform int stride = 2*programCount; stride < N; stride <<= 1) for (uniform int stride = 2*programCount; stride < N; stride <<= 1)
@@ -597,13 +593,13 @@ void mergeSort(
// cpu: 120 gpu: 457 M/s // cpu: 120 gpu: 457 M/s
//Merge ranks and indices //Merge ranks and indices
// mergeRanksAndIndices(limitsA, limitsB, ranksA, ranksB, stride, N); mergeRanksAndIndices(limitsA, limitsB, ranksA, ranksB, stride, N);
} }
#endif #endif
// cpu: 287 gpu: 194 M/s // cpu: 287 gpu: 194 M/s
//Merge elementary intervals //Merge elementary intervals
// mergeElementaryIntervals(oKey, oVal, iKey, iVal, limitsA, limitsB, stride, N); mergeElementaryIntervals(oKey, oVal, iKey, iVal, limitsA, limitsB, stride, N);
} }
if (lastSegmentElements <= stride) if (lastSegmentElements <= stride)