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

runs on GPU but need further tuning

Browse Source
This commit is contained in:
Evghenii
2014-03-05 11:49:15 +01:00
parent f086b7ff9b
commit 644118cd17
3 changed files with 45 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
examples/portable/inplaceTraspose/inplaceTranspose.ispc
View File

@@ -31,6 +31,7 @@ int __sj(const int i, const uniform int j, const uniform int m, const uniform in
}
#if 0
static inline
void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
{
@@ -83,6 +84,7 @@ void transpose_serial(uniform T A[], const uniform int m, const uniform int n)
delete joverb;
delete tmp;
}
#endif
static uniform T * uniform tmpAll = NULL;
static uniform int * uniform joverb = NULL;
@@ -90,9 +92,9 @@ static uniform int * uniform iovera = NULL;
static uniform int a,b,c;
static
void transpose_init(const uniform int m, const uniform int n, const uniform int taskCount)
void transpose_init(const uniform int m, const uniform int n, const uniform int nTask)
{
const uniform int tmpSize = max(m,n) * programCount * taskCount;
const uniform int tmpSize = max(m,n) * programCount * nTask;
tmpAll = uniform new uniform T [tmpSize];
joverb = uniform new uniform int[n];
iovera = uniform new uniform int[m];
@@ -178,7 +180,7 @@ void transpose(uniform T A[], const uniform int m, const uniform int n)
#if 0
transpose_serial(A, m, n);
#else
const uniform int nTask = num_cores()*4;
const uniform int nTask = 32*8; //num_cores()*4;
transpose_init(m,n,nTask);
launch [nTask] transpose_step1(A, m, n);
@@ -192,5 +194,6 @@ void transpose(uniform T A[], const uniform int m, const uniform int n)
transpose_finalize();
#endif
sync;
}