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

added workable .cu files for stencil & mandelbrot

Browse Source
This commit is contained in:
Evghenii
2013-11-08 10:00:49 +01:00
parent cb7cbec0d5
commit 426afc7377
10 changed files with 645 additions and 160 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
stdlib.ispc
View File

@@ -63,47 +63,56 @@
// CUDA Specific primitives
//
#define CUDABLOCKSIZE 128
#define WARPSIZE2 5
#define WARPSIZE (1<<WARPSIZE2)
/***************/
__declspec(safe,cost0)
static inline uniform int warpSize()
{
return WARPSIZE; //__warpsize();
}
/***************/
__declspec(safe,cost0)
static inline uniform int laneIndex()
{
return __tid_x() & (WARPSIZE-1) ; //& (warpSize()-1);
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockIndex0()
{
return __ctaid_x();
return (__ctaid_x() * (CUDABLOCKSIZE >> WARPSIZE2)) + (__tid_x() >> WARPSIZE2);
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockIndex1()
{
return __ctaid_y();
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockIndex2()
{
return __ctaid_y();
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockCount0()
{
return __nctaid_x();
return __nctaid_x() * (CUDABLOCKSIZE >> WARPSIZE2);
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockCount1()
{
return __nctaid_y();
}
/***************/
__declspec(safe,cost0)
static inline uniform int blockCount2()
{
return __nctaid_z();
}
__declspec(safe,cost0)
static inline uniform int warpSize()
{
return __warpsize();
}
__declspec(safe,cost0)
static inline uniform int laneIndex()
{
return __tid_x() & (warpSize()-1);
}
///////////////////////////////////////////////////////////////////////////
// Low level primitives