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

runs, next check correctness

Browse Source
This commit is contained in:
Evghenii
2013-11-13 18:15:52 +01:00
parent 8b0f871c06
commit e442139c39
3 changed files with 50 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
examples_cuda/sort/sort1.ispc
View File

@@ -194,51 +194,56 @@ task void bumpup (uniform int h[], uniform int g[])
}
}
static void prefix_sum (uniform int num, uniform int h[])
static void prefix_sum (uniform int num, uniform int h[], uniform int g[])
{
uniform int * uniform g = uniform new uniform int [num+1];
uniform int i;
launch[num] addup (h, g+1);
sync;
for (g[0] = 0, i = 1; i < num; i ++) g[i] += g[i-1];
g[0] = 0;
for (i = 1; i < num; i ++)
g[i] += g[i-1];
launch[num] bumpup (h, g);
sync;
delete g;
}
export void sort_ispc (uniform int n, uniform unsigned int code[], uniform int order[], uniform int ntasks)
export void sort_ispc (uniform int n,
uniform unsigned int code[],
uniform int order[],
uniform int ntasks,
uniform int _hist[],
uniform int64 _pair[],
uniform int64 _temp[],
uniform int _g[])
{
uniform int num = ntasks < 1 ? 13*4*8 : ntasks;
uniform int num = ntasks;
uniform int span = n / num;
#if 0
uniform int hsize = 256*programCount*num;
uniform int * uniform hist = uniform new uniform int [hsize];
uniform int64 * uniform pair = uniform new uniform int64 [n];
uniform int64 * uniform temp = uniform new uniform int64 [n];
uniform int pass, i;
#if DEBUG
if (n < 100)
{
print ("input: ");
for (i = 0; i < n; i ++) print ("%, ", code[i]);
print ("\n");
}
uniform int * uniform g = uniform new uniform int [num+1];
#else
uniform int * uniform hist = _hist;
uniform int64 * uniform pair = _pair;
uniform int64 * uniform temp = _temp;
uniform int * uniform g = _g;
#endif
uniform int pass, i;
launch[num] pack (span, n, code, pair);
sync;
#if 0
for (pass = 0; pass < 4; pass ++)
{
launch[num] histogram (span, n, pair, pass, hist);
sync;
prefix_sum (num, hist);
prefix_sum (num, hist, g);
launch[num] permutation (span, n, pair, pass, hist, temp);
sync;
@@ -250,26 +255,11 @@ export void sort_ispc (uniform int n, uniform unsigned int code[], uniform int o
launch[num] unpack (span, n, pair, code, order);
sync;
#if DEBUG
for (i = 0; i < n; i ++)
{
if (i > 0 && code[i-1] > code[i])
print ("ERR at % => % > %; ", i, code[i-1], code[i]);
}
if (n < 100)
{
print ("output: ");
for (i = 0; i < n; i ++) print ("%, ", code[i]);
print ("\n");
print ("order: ");
for (i = 0; i < n; i ++) print ("%, ", order[i]);
print ("\n");
}
#endif
#endif
#if ALLOCATED
delete g;
delete hist;
delete pair;
delete temp;
#endif
}