<html>
<head>
<base href="https://bugs.llvm.org/">
</head>
<body><table border="1" cellspacing="0" cellpadding="8">
<tr>
<th>Bug ID</th>
<td><a class="bz_bug_link
bz_status_NEW "
title="NEW - OpenMP target regions cause CUBLAS_STATUS_INTERNAL_ERROR in libcublas"
href="https://bugs.llvm.org/show_bug.cgi?id=42643">42643</a>
</td>
</tr>
<tr>
<th>Summary</th>
<td>OpenMP target regions cause CUBLAS_STATUS_INTERNAL_ERROR in libcublas
</td>
</tr>
<tr>
<th>Product</th>
<td>OpenMP
</td>
</tr>
<tr>
<th>Version</th>
<td>unspecified
</td>
</tr>
<tr>
<th>Hardware</th>
<td>PC
</td>
</tr>
<tr>
<th>OS</th>
<td>Linux
</td>
</tr>
<tr>
<th>Status</th>
<td>NEW
</td>
</tr>
<tr>
<th>Severity</th>
<td>normal
</td>
</tr>
<tr>
<th>Priority</th>
<td>P
</td>
</tr>
<tr>
<th>Component</th>
<td>Clang Compiler Support
</td>
</tr>
<tr>
<th>Assignee</th>
<td>unassignedclangbugs@nondot.org
</td>
</tr>
<tr>
<th>Reporter</th>
<td>csdaley@lbl.gov
</td>
</tr>
<tr>
<th>CC</th>
<td>llvm-bugs@lists.llvm.org
</td>
</tr></table>
<p>
<div>
<pre>Created <span class=""><a href="attachment.cgi?id=22253" name="attach_22253" title="Source files and build scripts to reproduce the error">attachment 22253</a> <a href="attachment.cgi?id=22253&action=edit" title="Source files and build scripts to reproduce the error">[details]</a></span>
Source files and build scripts to reproduce the error
Hello all,
We are trying to use libcublas and OpenMP target offloading in a small
mini-app. When we compile with Clang compiler we find that our second call to
cublasDgemm returns CUBLAS_STATUS_INTERNAL_ERROR - see the top of run-clang.sh
and run-clang.out. We print "** cublasDgemm status-2: 14" when there is
CUBLAS_STATUS_INTERNAL_ERROR return value. The numerical results in this small
mini-app are OK, but eventually cause problems in our full application.
The following alternate compilers work:
* The current IBM XL compiler on OLCF Summit supercomputer - see run-xl.sh and
run-xl.out.
* The PGI compiler, where we replace the OpenMP target offload calls with
OpenACC.
The following 2 versions work with Clang compiler.
* Replace OpenMP target offload code with CUDA code - see
CUBLAS_testcase_cuda.cu. This is the second experiment that is run in
run-clang.sh.
* Replace cublasDgemm calls in libcublas with cblas_dgemm calls in MKL
This indicates that the error is a result of mixing both OpenMP target offload
and libcublas with the Clang compiler. We are using LLVM/Clang from June 06,
but have also reproduced using a version of LLVM/Clang from Feb 20. Our
mini-app is quite complicated, however, it is the simplest mini-app we could
create which still caused CUBLAS_STATUS_INTERNAL_ERROR return value. Please see
run-clang.sh and CUBLAS_testcase.c.
Thanks,
Chris and Fazlay</pre>
</div>
</p>
<hr>
<span>You are receiving this mail because:</span>
<ul>
<li>You are on the CC list for the bug.</li>
</ul>
</body>
</html>