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

some tuning

Browse Source
This commit is contained in:
Evghenii
2014-01-31 19:53:52 +01:00
parent bead800c13
commit eb82195ad7
2 changed files with 17 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
examples_ptx/mergeSort/mergeSort.cu
View File

@@ -233,7 +233,7 @@ void mergeSortGang(
uniform Val_t srcVal[], uniform Val_t srcVal[],
uniform int batchSize) uniform int batchSize)
{ {
uniform int nTasks = batchSize/4; uniform int nTasks = batchSize;
launch (nTasks,1,1,mergeSortGangKernel)(batchSize, dstKey, dstVal, srcKey, srcVal); launch (nTasks,1,1,mergeSortGangKernel)(batchSize, dstKey, dstVal, srcKey, srcVal);
sync; sync;
} }
@@ -306,7 +306,7 @@ void generateSampleRanks(
(N - lastSegmentElements) / (2 * SAMPLE_STRIDE); (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE); uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE);
uniform int nTasks = nBlocks/4; uniform int nTasks = nBlocks;
launch (nTasks,1,1, generateSampleRanksKernel)(nBlocks, ranksA, ranksB, srcKey, stride, N, threadCount); launch (nTasks,1,1, generateSampleRanksKernel)(nBlocks, ranksA, ranksB, srcKey, stride, N, threadCount);
sync; sync;
@@ -372,7 +372,7 @@ void mergeRanksAndIndices(
(N - lastSegmentElements) / (2 * SAMPLE_STRIDE); (N - lastSegmentElements) / (2 * SAMPLE_STRIDE);
const uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE); const uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE);
uniform int nTasks = nBlocks/4; uniform int nTasks = nBlocks;
launch (nTasks,1,1,mergeRanksAndIndicesKernel)( launch (nTasks,1,1,mergeRanksAndIndicesKernel)(
nBlocks, nBlocks,
@@ -448,20 +448,21 @@ void mergeElementaryIntervalsKernel(
} }
// Compute destination addresses for merge data // Compute destination addresses for merge data
int dstPosA = binarySearchExclusive1(keyA, keyB, lenSrcB, SAMPLE_STRIDE) + programIndex; int dstPosA, dstPosB, dstA = -1, dstB = -1;
int dstPosB = binarySearchInclusive1(keyB, keyA, lenSrcA, SAMPLE_STRIDE) + programIndex; if (any(programIndex < lenSrcA))
dstPosA = binarySearchExclusive1(keyA, keyB, lenSrcB, SAMPLE_STRIDE) + programIndex;
if (any(programIndex < lenSrcB))
dstPosB = binarySearchInclusive1(keyB, keyA, lenSrcA, SAMPLE_STRIDE) + programIndex;
int dstA = -1, dstB = -1;
if (programIndex < lenSrcA && dstPosA < lenSrcA) if (programIndex < lenSrcA && dstPosA < lenSrcA)
dstA = segmentBase + startDstA + dstPosA; dstA = segmentBase + startDstA + dstPosA;
if (programIndex < lenSrcB && dstPosB < lenSrcA)
dstB = segmentBase + startDstA + dstPosB;
dstPosA -= lenSrcA; dstPosA -= lenSrcA;
dstPosB -= lenSrcA;
if (programIndex < lenSrcA && dstPosA < lenSrcB) if (programIndex < lenSrcA && dstPosA < lenSrcB)
dstA = segmentBase + startDstB + dstPosA; dstA = segmentBase + startDstB + dstPosA;
if (programIndex < lenSrcB && dstPosB < lenSrcA)
dstB = segmentBase + startDstA + dstPosB;
dstPosB -= lenSrcA;
if (programIndex < lenSrcB && dstPosB < lenSrcB) if (programIndex < lenSrcB && dstPosB < lenSrcB)
dstB = segmentBase + startDstB + dstPosB; dstB = segmentBase + startDstB + dstPosB;
@@ -499,7 +500,7 @@ void mergeElementaryIntervals(
const uniform int mergePairs = (lastSegmentElements > stride) ? getSampleCount(N) : (N - lastSegmentElements) / SAMPLE_STRIDE; const uniform int mergePairs = (lastSegmentElements > stride) ? getSampleCount(N) : (N - lastSegmentElements) / SAMPLE_STRIDE;
nTasks = mergePairs/(4*programCount); nTasks = mergePairs/(programCount);
launch (nTasks,1,1, mergeElementaryIntervalsKernel)( launch (nTasks,1,1, mergeElementaryIntervalsKernel)(
mergePairs, mergePairs,

8
examples_ptx/mergeSort/mergeSort.ispc
View File

@@ -221,7 +221,7 @@ void mergeSortGang(
{ {
uniform int nTasks = num_cores()*4; uniform int nTasks = num_cores()*4;
#ifdef __NVPTX__ #ifdef __NVPTX__
nTasks = iDivUp(batchSize,4); nTasks = iDivUp(batchSize,1);
#endif #endif
launch [nTasks] mergeSortGangKernel(batchSize, dstKey, dstVal, srcKey, srcVal, 2*programCount); launch [nTasks] mergeSortGangKernel(batchSize, dstKey, dstVal, srcKey, srcVal, 2*programCount);
sync; sync;
@@ -297,7 +297,7 @@ void generateSampleRanks(
uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE); uniform int nBlocks = iDivUp(threadCount, SAMPLE_STRIDE);
uniform int nTasks = num_cores()*4; uniform int nTasks = num_cores()*4;
#ifdef __NVPTX__ #ifdef __NVPTX__
nTasks = iDivUp(nBlocks,4); nTasks = iDivUp(nBlocks,1);
#endif #endif
launch [nTasks] generateSampleRanksKernel(nBlocks, ranksA, ranksB, srcKey, stride, N, threadCount); launch [nTasks] generateSampleRanksKernel(nBlocks, ranksA, ranksB, srcKey, stride, N, threadCount);
@@ -367,7 +367,7 @@ void mergeRanksAndIndices(
uniform int nTasks = num_cores()*4; uniform int nTasks = num_cores()*4;
#ifdef __NVPTX__ #ifdef __NVPTX__
nTasks = iDivUp(nBlocks,4); nTasks = iDivUp(nBlocks,1);
#endif #endif
launch [nTasks] mergeRanksAndIndicesKernel( launch [nTasks] mergeRanksAndIndicesKernel(
@@ -492,7 +492,7 @@ void mergeElementaryIntervals(
uniform int nTasks = num_cores()*4; uniform int nTasks = num_cores()*4;
#ifdef __NVPTX__ #ifdef __NVPTX__
nTasks = iDivUp(mergePairs,4*programCount); nTasks = iDivUp(mergePairs,1*programCount);
#endif #endif
launch [nTasks] mergeElementaryIntervalsKernel( launch [nTasks] mergeElementaryIntervalsKernel(