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

+1

Browse Source
This commit is contained in:
Evghenii
2014-01-30 10:26:58 +01:00
parent 2f44b81d4f
commit 4e26a1b700
2 changed files with 50 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples_ptx/mergeSort/Makefile_cpu
View File

@@ -2,7 +2,7 @@
EXAMPLE=mergeSort
CPP_SRC=mergeSort.cpp
ISPC_SRC=mergeSort.ispc
ISPC_IA_TARGETS=avx1-i32x16
ISPC_IA_TARGETS=avx1-i32x8
ISPC_ARM_TARGETS=neon
#ISPC_FLAGS=-DDEBUG -g
CXXFLAGS=-g

76
examples_ptx/mergeSort/mergeSort.ispc
View File

@@ -161,43 +161,54 @@ int binarySearchExclusive1(
////////////////////////////////////////////////////////////////////////////////
task
void mergeSortGangKernel(
uniform int batchSize,
uniform Key_t dstKey[],
uniform Val_t dstVal[],
uniform Key_t srcKey[],
uniform Val_t srcVal[])
{
const uniform int blockIdx = taskIndex;
const uniform int blockDim = (batchSize + taskCount - 1)/taskCount;
const uniform int blockBeg = blockIdx * blockDim;
const uniform int blockEnd = min(blockBeg + blockDim, batchSize);
uniform Key_t s_key[2*programCount];
uniform Val_t s_val[2*programCount];
const uniform int base = taskIndex * (programCount*2);
s_key[programIndex + 0] = srcKey[base + programIndex + 0];
s_val[programIndex + 0] = srcVal[base + programIndex + 0];
s_key[programIndex + programCount] = srcKey[base + programIndex + programCount];
s_val[programIndex + programCount] = srcVal[base + programIndex + programCount];
for (uniform int stride = 1; stride < 2*programCount; stride <<= 1)
for (uniform int block = blockBeg; block < blockEnd; block++)
{
const int lPos = programIndex & (stride - 1);
uniform Key_t *baseKey = s_key + 2 * (programIndex - lPos);
uniform Val_t *baseVal = s_val + 2 * (programIndex - lPos);
const uniform int base = block * (programCount*2);
s_key[programIndex + 0] = srcKey[base + programIndex + 0];
s_val[programIndex + 0] = srcVal[base + programIndex + 0];
s_key[programIndex + programCount] = srcKey[base + programIndex + programCount];
s_val[programIndex + programCount] = srcVal[base + programIndex + programCount];
Key_t keyA = baseKey[lPos + 0];
Val_t valA = baseVal[lPos + 0];
Key_t keyB = baseKey[lPos + stride];
Val_t valB = baseVal[lPos + stride];
int posA = binarySearchExclusive(keyA, baseKey + stride, stride, stride) + lPos;
int posB = binarySearchInclusive(keyB, baseKey + 0, stride, stride) + lPos;
#if 1
for (uniform int stride = 1; stride < 2*programCount; stride <<= 1)
{
const int lPos = programIndex & (stride - 1);
uniform Key_t *baseKey = s_key + 2 * (programIndex - lPos);
uniform Val_t *baseVal = s_val + 2 * (programIndex - lPos);
baseKey[posA] = keyA;
baseVal[posA] = valA;
baseKey[posB] = keyB;
baseVal[posB] = valB;
Key_t keyA = baseKey[lPos + 0];
Val_t valA = baseVal[lPos + 0];
Key_t keyB = baseKey[lPos + stride];
Val_t valB = baseVal[lPos + stride];
int posA = binarySearchExclusive(keyA, baseKey + stride, stride, stride) + lPos;
int posB = binarySearchInclusive(keyB, baseKey + 0, stride, stride) + lPos;
baseKey[posA] = keyA;
baseVal[posA] = valA;
baseKey[posB] = keyB;
baseVal[posB] = valB;
}
#endif
dstKey[base + programIndex + 0] = s_key[programIndex + 0];
dstVal[base + programIndex + 0] = s_val[programIndex + 0];
dstKey[base + programIndex + programCount] = s_key[programIndex + programCount];
dstVal[base + programIndex + programCount] = s_val[programIndex + programCount];
}
dstKey[base + programIndex + 0] = s_key[programIndex + 0];
dstVal[base + programIndex + 0] = s_val[programIndex + 0];
dstKey[base + programIndex + programCount] = s_key[programIndex + programCount];
dstVal[base + programIndex + programCount] = s_val[programIndex + programCount];
}
static inline
@@ -208,7 +219,11 @@ void mergeSortGang(
uniform Val_t srcVal[],
uniform int batchSize)
{
launch [batchSize] mergeSortGangKernel(dstKey, dstVal, srcKey, srcVal);
uniform int nTasks = num_cores()*4;
#ifdef __NVPTX__
nTasks = batchSize/4;
#endif
launch [nTasks] mergeSortGangKernel(batchSize, dstKey, dstVal, srcKey, srcVal);
sync;
}
@@ -536,8 +551,8 @@ void mergeElementaryIntervalsKernel(
valB = srcVal[segmentBase + stride + startSrcB + programIndex];
}
// Compute destination addresses for merge data
int dstPosA, dstPosB;
// Compute destination addresses for merge data
if (programIndex < lenSrcA)
dstPosA = binarySearchExclusive1(keyA, keyB, lenSrcB, SAMPLE_STRIDE) + programIndex;
if (programIndex < lenSrcB)
@@ -560,11 +575,13 @@ void mergeElementaryIntervalsKernel(
// store merge data
if (dstA >= 0)
{
// int dstA = segmentBase + startSrcA + programIndex;
dstKey[dstA] = keyA;
dstVal[dstA] = valA;
}
if (dstB >= 0)
{
// int dstB = segmentBase + stride + startSrcB + programIndex;
dstKey[dstB] = keyB;
dstVal[dstB] = valB;
}
@@ -600,6 +617,9 @@ void mergeElementaryIntervals(
stride,
N);
#else
#ifdef __NVPTX__
nTasks = mergePairs/(4*programCount);
#endif
launch [nTasks] mergeElementaryIntervalsKernel(
mergePairs,
dstKey,
@@ -684,6 +704,7 @@ void mergeSort(
assert(N % (programCount*2) == 0);
mergeSortGang(iKey, iVal, srcKey, srcVal, N/(2*programCount));
#if 1
for (uniform int stride = 2*programCount; stride < N; stride <<= 1)
{
const uniform int lastSegmentElements = N % (2 * stride);
@@ -717,4 +738,5 @@ void mergeSort(
oVal = tmpVal;
}
}
#endif
}