- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
(Cross-posting as requested from the toolkit forum.)
The Intel OpenCL drivers (2024.2.0.980 on Windows and 2021.12.6.0.19_160000 on Linux) seem to have an issue with vectorization resulting in corrupted data. This simple kernel:
__kernel void test(__global float *f, __global float *r) {
int i = get_global_id(0);
r[i] = 0.0;
if (f[i] == 1.0F) {
r[i] = 1.0F+pow(1.0F, 1.0F);
}
}
when run over a buffer of length 16 with f equal to:
1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1
results in the following values in r:
2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 3126, 0, 0, 0, 0, 3126
Note the odd 3126 values. Turning off vectorisation using
CL_CONFIG_CPU_VECTORIZER_MODE=1
Results in the correct values:
2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2
The issue seems to be caused by adding a value to the function (i.e. here 1.0+pow()), for example this gives an expected result:
__kernel void test(__global float *f, __global float *r) {
int i = get_global_id(0);
r[i] = 0.0;
if (f[i] == 1.0F) {
r[i] = pow(1.0F, 1.0F);
}
}
Code for a MVE is attached.
Link Copied
0 Replies

Reply
Topic Options
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page