<div dir="ltr"><div>Yeah, the example code is below:</div><div><br></div><div> int N = 1<<20;<br><br> float *x = new float[N];<br> float *y = new float[N];<br><br> for (int i = 0; i < N; i++) {<br> x[i] = 1.0f;<br> y[i] = 2.0f;<br> }<br><br> float *z = new float[N];<br> int i;<br> #pragma omp target map(x, y, z)<br> #pragma omp parallel for<br> for (i=0; i < N; i++) {<br> z[i] = x[i] + y[i];<br> }</div><div><br></div><div>I just grab a piece of code from <a href="https://www.openmp.org/wp-content/uploads/openmp-examples-4.5.0.pdf">https://www.openmp.org/wp-content/uploads/openmp-examples-4.5.0.pdf</a> for testing. I also tested other examples in that document, but none worked. Initially, I was working on a piece of code from a legacy project.</div><div><br></div><div>BTW, when I compiled the example code, I got some warnings:</div><div><br></div><div>clang-11: warning: Unknown CUDA version 10.2. Assuming the latest supported version 10.1 [-Wunknown-cuda-version]<br>clang-11: warning: Unknown CUDA version 10.2. Assuming the latest supported version 10.1 [-Wunknown-cuda-version]<br>clang-11: warning: No library 'libomptarget-nvptx-sm_35.bc' found in the default clang lib directory or in LIBRARY_PATH. Expect degraded performance due to no inlining of runtime functions on target devices. [-Wopenmp-target]</div><div><br></div><div>I am not sure if the error is caused by the newer CUDA version (I thought 10.2 should be compatible with 10.1)</div><div><br></div><div>Thanks!</div><div><br></div><div>Gang Zhao<br></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">Alexey.Bataev <<a href="mailto:a.bataev@outlook.com">a.bataev@outlook.com</a>> 于2020年3月5日周四 下午12:59写道：<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>
<p>Could you provide an example how do you map the data in the
target region?<br>
</p>
<pre cols="72">-------------
Best regards,
Alexey Bataev</pre>
<div>05.03.2020 1:45 PM, G Zhao пишет:<br>
</div>
<blockquote type="cite">
<div dir="ltr">
<div>Thanks! I didn't notice this. The code is from a legacy
project and I just checked all the target regions. It did use
STL vectors. I just replaced all those vectors with arrays.
Now it can successfully compile. But When I run it, there is
another error:</div>
<div><br>
</div>
<div>Libomptarget fatal error 1: failure of target construct
while offloading is mandatory</div>
<div><br>
</div>
<div>I tried it on a simple vector add example, and got the same
error.</div>
<div><br>
</div>
<div>Below is the debug information with LIBOMPTARGET_DEBUG=1:</div>
<div><br>
</div>
<div>Libomptarget --> Loading RTLs...<br>
Libomptarget --> Loading library '<a href="http://libomptarget.rtl.ppc64.so" target="_blank">libomptarget.rtl.ppc64.so</a>'...<br>
Libomptarget --> Unable to load library '<a href="http://libomptarget.rtl.ppc64.so" target="_blank">libomptarget.rtl.ppc64.so</a>': <a href="http://libomptarget.rtl.ppc64.so" target="_blank">libomptarget.rtl.ppc64.so</a>: cannot
open shared object file: No such file or directory!<br>
Libomptarget --> Loading library '<a href="http://libomptarget.rtl.x86_64.so" target="_blank">libomptarget.rtl.x86_64.so</a>'...<br>
Libomptarget --> Successfully loaded library '<a href="http://libomptarget.rtl.x86_64.so" target="_blank">libomptarget.rtl.x86_64.so</a>'!<br>
Libomptarget --> Registering RTL <a href="http://libomptarget.rtl.x86_64.so" target="_blank">libomptarget.rtl.x86_64.so</a>
supporting 4 devices!<br>
Libomptarget --> Loading library '<a href="http://libomptarget.rtl.cuda.so" target="_blank">libomptarget.rtl.cuda.so</a>'...<br>
Target CUDA RTL --> Start initializing CUDA<br>
Libomptarget --> Successfully loaded library '<a href="http://libomptarget.rtl.cuda.so" target="_blank">libomptarget.rtl.cuda.so</a>'!<br>
Libomptarget --> Registering RTL <a href="http://libomptarget.rtl.cuda.so" target="_blank">libomptarget.rtl.cuda.so</a>
supporting 1 devices!<br>
Libomptarget --> Loading library '<a href="http://libomptarget.rtl.aarch64.so" target="_blank">libomptarget.rtl.aarch64.so</a>'...<br>
Libomptarget --> Unable to load library '<a href="http://libomptarget.rtl.aarch64.so" target="_blank">libomptarget.rtl.aarch64.so</a>': <a href="http://libomptarget.rtl.aarch64.so" target="_blank">libomptarget.rtl.aarch64.so</a>:
cannot open shared object file: No such file or directory!<br>
Libomptarget --> RTLs loaded!<br>
Libomptarget --> Image 0x000000000041ad20 is NOT compatible
with RTL <a href="http://libomptarget.rtl.x86_64.so" target="_blank">libomptarget.rtl.x86_64.so</a>!<br>
Libomptarget --> Image 0x000000000041ad20 is compatible
with RTL <a href="http://libomptarget.rtl.cuda.so" target="_blank">libomptarget.rtl.cuda.so</a>!<br>
Libomptarget --> RTL 0x00000000015b3c40 has index 0!<br>
Libomptarget --> Registering image 0x000000000041ad20 with
RTL <a href="http://libomptarget.rtl.cuda.so" target="_blank">libomptarget.rtl.cuda.so</a>!<br>
Libomptarget --> Done registering entries!<br>
Libomptarget --> Call to omp_get_num_devices returning 1<br>
Libomptarget --> Default TARGET OFFLOAD policy is now
mandatory (devices were found)<br>
Libomptarget --> Checking whether device 0 is ready.<br>
Libomptarget --> Is the device 0 (local ID 0) initialized?
0<br>
Target CUDA RTL --> Init requires flags to 1<br>
Target CUDA RTL --> Getting device 0<br>
Target CUDA RTL --> Max CUDA blocks per grid 2147483647
exceeds the hard team limit 65536, capping at the hard limit<br>
Target CUDA RTL --> Using 1024 CUDA threads per block<br>
Target CUDA RTL --> Max number of CUDA blocks 65536,
threads 1024 & warp size 32<br>
Target CUDA RTL --> Default number of teams set according
to library's default 128<br>
Target CUDA RTL --> Default number of threads set according
to library's default 128<br>
Libomptarget --> Device 0 is ready to use.<br>
Target CUDA RTL --> Load data from image 0x000000000041ad20<br>
Target CUDA RTL --> Error when loading CUDA module<br>
Target CUDA RTL --> CUDA error is: device kernel image is
invalid<br>
Libomptarget --> Unable to generate entries table for
device id 0.<br>
Libomptarget --> Failed to init globals on device 0<br>
Libomptarget --> Failed to get device 0 ready<br>
Libomptarget fatal error 1: failure of target construct while
offloading is mandatory<br>
Libomptarget --> Unloading target library!<br>
Libomptarget --> Image 0x000000000041ad20 is compatible
with RTL 0x00000000015b3c40!<br>
Libomptarget --> Unregistered image 0x000000000041ad20 from
RTL 0x00000000015b3c40!<br>
Libomptarget --> Done unregistering images!<br>
Libomptarget --> Removing translation table for descriptor
0x0000000000440810<br>
Libomptarget --> Done unregistering library!<br>
Libomptarget --> Deinit target library!<br>
</div>
<div><br>
</div>
<div>Any hints about this? <br>
</div>
<div><br>
</div>
<div>Regards,</div>
<div>Gang Zhao<br>
</div>
<br>
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">Alexey Bataev <<a href="mailto:a.bataev@hotmail.com" target="_blank">a.bataev@hotmail.com</a>>
于2020年3月5日周四 上午5:44写道：<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Neither libc++, nor
libstdc++ cannot be linked with nvlink. NVidia does not
provide implementations for either libc++, or libstdc++. You
must exclude the use of the standard c++ library from target
regions.<br>
<br>
Best regards,<br>
Alexey Bataev<br>
<br>
> 5 марта 2020 г., в 00:25, G Zhao via Openmp-dev <<a href="mailto:openmp-dev@lists.llvm.org" target="_blank">openmp-dev@lists.llvm.org</a>>
написал(а):<br>
> <br>
> <br>
> Hi,<br>
> <br>
> I just compiled LLVM and enable NVPTX by
-DLLVM_TARGETS_TO_BUILD="X86;NVPTX". But when I compiled my
code using the below command:<br>
> <br>
> clang++ main.cpp -fopenmp
-fopenmp-targets=nvptx64-nvidia-cuda -o a_gpu.exe<br>
> <br>
> I got the below error:<br>
> <br>
>
/usr/lib64/gcc/x86_64-pc-linux-gnu/9.2.1/../../../../include/c++/9.2.1/bits/std_abs.h:75:3:
error: declaration conflicts with target of using
declaration already in scope<br>
> abs(float __x)<br>
> <br>
> I think the reason is I am using GCC-9. I did a bit
search and someone said using libc++ can address this. So I
compiled libcxx and libcxxabi, and used the below command to
compile my code again:<br>
> <br>
> clang++ stdlib=libc++ main.cpp -fopenmp
-fopenmp-targets=nvptx64-nvidia-cuda -o a_gpu.exe<br>
> <br>
> I got different errors:<br>
> <br>
> nvlink error : Undefined reference to
'_ZNKSt3__120__vector_base_commonILb1EE20__throw_length_errorEv'
in '/tmp/main-42e0a6.cubin'<br>
> nvlink error : Undefined reference to 'abort' in
'/tmp/main-42e0a6.cubin'<br>
> <br>
> I think the reason here is nvlink don't know we should
link libc++ with those cubin files together. But I don't
know how to solve this.<br>
> <br>
> Any one know a workaround to this?<br>
> <br>
> Thanks!<br>
> <br>
> <br>
> <br>
> <br>
> _______________________________________________<br>
> Openmp-dev mailing list<br>
> <a href="mailto:Openmp-dev@lists.llvm.org" target="_blank">Openmp-dev@lists.llvm.org</a><br>
> <a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/openmp-dev" rel="noreferrer" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/openmp-dev</a><br>
</blockquote>
</div>
</div>
</blockquote>
</div>
</blockquote></div>