some tining
This commit is contained in:
@@ -201,6 +201,7 @@ void mergeSortGangKernel(
|
||||
Val_t valA = baseVal[lPos + 0];
|
||||
Key_t keyB = baseKey[lPos + stride];
|
||||
Val_t valB = baseVal[lPos + stride];
|
||||
#if 1
|
||||
int posA = binarySearchExclusive(keyA, baseKey + stride, stride, stride) + lPos;
|
||||
int posB = binarySearchInclusive(keyB, baseKey + 0, stride, stride) + lPos;
|
||||
|
||||
@@ -208,6 +209,12 @@ void mergeSortGangKernel(
|
||||
baseVal[posA] = valA;
|
||||
baseKey[posB] = keyB;
|
||||
baseVal[posB] = valB;
|
||||
#else
|
||||
s_key[programIndex] = keyA;
|
||||
s_val[programIndex] = valA;
|
||||
s_key[programCount+programIndex] = keyB;
|
||||
s_val[programCount+programIndex] = valB;
|
||||
#endif
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
@@ -164,7 +164,8 @@ void mergeSortGangKernel(
|
||||
uniform Key_t dstKey[],
|
||||
uniform Val_t dstVal[],
|
||||
uniform Key_t srcKey[],
|
||||
uniform Val_t srcVal[])
|
||||
uniform Val_t srcVal[],
|
||||
uniform int arrayLength)
|
||||
{
|
||||
const uniform int blockIdx = taskIndex;
|
||||
const uniform int blockDim = (batchSize + taskCount - 1)/taskCount;
|
||||
@@ -182,17 +183,33 @@ void mergeSortGangKernel(
|
||||
s_key[programIndex + programCount] = srcKey[base + programIndex + programCount];
|
||||
s_val[programIndex + programCount] = srcVal[base + programIndex + programCount];
|
||||
|
||||
#define STEP(stride) {\
|
||||
const int lPos = programIndex & (stride - 1); \
|
||||
const int offset = 2 * (programIndex - lPos); \
|
||||
Key_t keyA = s_key[lPos + 0]; \
|
||||
Val_t valA = s_val[lPos + 0]; \
|
||||
Key_t keyB = s_key[lPos + stride]; \
|
||||
Val_t valB = s_val[lPos + stride]; \
|
||||
s_key[programIndex] = keyA; \
|
||||
s_val[programIndex] = valA; \
|
||||
s_key[programCount+programIndex] = keyB; \
|
||||
s_val[programCount+programIndex] = valB; \
|
||||
}
|
||||
|
||||
#if 1
|
||||
for (uniform int stride = 1; stride < 2*programCount; stride <<= 1)
|
||||
for (uniform int stride = 1; stride < arrayLength; stride <<= 1)
|
||||
{
|
||||
const int lPos = programIndex & (stride - 1);
|
||||
const int offset = 2 * (programIndex - lPos);
|
||||
uniform Key_t *baseKey = s_key + 2 * (programIndex - lPos);
|
||||
uniform Val_t *baseVal = s_val + 2 * (programIndex - lPos);
|
||||
|
||||
|
||||
Key_t keyA = baseKey[lPos + 0];
|
||||
Val_t valA = baseVal[lPos + 0];
|
||||
Key_t keyB = baseKey[lPos + stride];
|
||||
Val_t valB = baseVal[lPos + stride];
|
||||
|
||||
#if 1
|
||||
int posA = binarySearchExclusive(keyA, baseKey + stride, stride, stride) + lPos;
|
||||
int posB = binarySearchInclusive(keyB, baseKey + 0, stride, stride) + lPos;
|
||||
|
||||
@@ -200,6 +217,12 @@ void mergeSortGangKernel(
|
||||
baseVal[posA] = valA;
|
||||
baseKey[posB] = keyB;
|
||||
baseVal[posB] = valB;
|
||||
#else
|
||||
s_key[programIndex] = keyA;
|
||||
s_val[programIndex] = valA;
|
||||
s_key[programCount+programIndex] = keyB;
|
||||
s_val[programCount+programIndex] = valB;
|
||||
#endif
|
||||
}
|
||||
#endif
|
||||
|
||||
@@ -222,7 +245,7 @@ void mergeSortGang(
|
||||
#ifdef __NVPTX__
|
||||
nTasks = batchSize/4;
|
||||
#endif
|
||||
launch [nTasks] mergeSortGangKernel(batchSize, dstKey, dstVal, srcKey, srcVal);
|
||||
launch [nTasks] mergeSortGangKernel(batchSize, dstKey, dstVal, srcKey, srcVal, 2*programCount);
|
||||
sync;
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user