added performacne data
This commit is contained in:
@@ -583,6 +583,10 @@ void mergeSort___export(
|
||||
|
||||
assert(N <= SAMPLE_STRIDE * MAX_SAMPLE_COUNT);
|
||||
assert(N % (programCount*2) == 0);
|
||||
|
||||
// k20m: 140 M/s
|
||||
{
|
||||
// k20m: 2367 M/s
|
||||
mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount));
|
||||
|
||||
#if 1
|
||||
@@ -590,14 +594,25 @@ void mergeSort___export(
|
||||
{
|
||||
const uniform int lastSegmentElements = N % (2 * stride);
|
||||
|
||||
// k20m: 271 M/s
|
||||
{
|
||||
#if 1
|
||||
// k20m: 944 M/s
|
||||
{
|
||||
// k20m: 1396 M/s
|
||||
//Find sample ranks and prepare for limiters merge
|
||||
generateSampleRanks(ranksA, ranksB, iKey, stride, N);
|
||||
|
||||
// k20m: 2379 M/s
|
||||
//Merge ranks and indices
|
||||
mergeRanksAndIndices(limitsA, limitsB, ranksA, ranksB, stride, N);
|
||||
}
|
||||
#endif
|
||||
|
||||
// k20m: 371 M/s
|
||||
//Merge elementary intervals
|
||||
mergeElementaryIntervals(nTasks, oKey, oVal, iKey, iVal, limitsA, limitsB, stride, N);
|
||||
}
|
||||
|
||||
if (lastSegmentElements <= stride)
|
||||
for (int i = programIndex; i < lastSegmentElements; i += programCount)
|
||||
@@ -619,6 +634,7 @@ void mergeSort___export(
|
||||
oVal = tmpVal;
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
}
|
||||
extern "C"
|
||||
|
||||
@@ -687,6 +687,10 @@ void mergeSort(
|
||||
|
||||
assert(N <= SAMPLE_STRIDE * MAX_SAMPLE_COUNT);
|
||||
assert(N % (programCount*2) == 0);
|
||||
|
||||
// cpu: 28 gpu: 74 M/s
|
||||
{
|
||||
// cpu: 356 gpu: 534 M/s
|
||||
mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount));
|
||||
|
||||
#if 1
|
||||
@@ -694,14 +698,25 @@ void mergeSort(
|
||||
{
|
||||
const uniform int lastSegmentElements = N % (2 * stride);
|
||||
|
||||
// cpu: 30 gpu: 112 M/s
|
||||
{
|
||||
#if 1
|
||||
// cpu: 121 gpu: 460 M/s
|
||||
{
|
||||
// cpu: 190 gpu: 600 M/s
|
||||
//Find sample ranks and prepare for limiters merge
|
||||
generateSampleRanks(ranksA, ranksB, iKey, stride, N);
|
||||
|
||||
// cpu: 120 gpu: 457 M/s
|
||||
//Merge ranks and indices
|
||||
mergeRanksAndIndices(limitsA, limitsB, ranksA, ranksB, stride, N);
|
||||
}
|
||||
#endif
|
||||
|
||||
// cpu: 287 gpu: 194 M/s
|
||||
//Merge elementary intervals
|
||||
mergeElementaryIntervals(nTasks, oKey, oVal, iKey, iVal, limitsA, limitsB, stride, N);
|
||||
}
|
||||
|
||||
if (lastSegmentElements <= stride)
|
||||
foreach (i = 0 ... lastSegmentElements)
|
||||
@@ -723,5 +738,6 @@ void mergeSort(
|
||||
oVal = tmpVal;
|
||||
}
|
||||
}
|
||||
}
|
||||
#endif
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user