[cfe-dev] gpucc breaks cuda 7.0.28/7_CUDALibraries/simpleCUFFT

[Peter Steinbach via cfe-dev]

Hi Artem,

the described "bug" is gone with llvm/clang trunk. Are there any news on 
texture memory support with cuda clang yet?

Thanks a bunch -
P

On 17.05.2016 00:28, Artem Belevich wrote:
> Peter,
>
> I can't reproduce the problem with recent clang. I've tried compiling
> the same smaple code with cuda-7.0 and 7.5.
> clang version 3.9.0 (trunk 268962) (llvm/trunk 268980)
>
> Could you tell me what was clang version you used? If you still see the
> problem, please file a clang  bug on llvm.org/bugs <http://llvm.org/bugs>.
>
> --Artem
>
>
> % clang++ -I../../common/inc --cuda-gpu-arch=sm_35 simpleCUFFT.cu
> -L/usr/local/cuda-7.5/lib64 -lcufft -lcudart -o simpleCUFFT-clang
> % LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64 ./simpleCUFFT-clang
> [simpleCUFFT] is starting...
> GPU Device 0: "Tesla K40c" with compute capability 3.5
>
> Transforming signal cufftExecC2C
> Launching ComplexPointwiseMulAndScale<<< >>>
> Transforming signal back cufftExecC2C
>
>
> On Tue, Apr 5, 2016 at 2:12 PM, Jingyue Wu <jingyue at google.com
> <mailto:jingyue at google.com>> wrote:
>
>     Would you mind upload your simpleCUFFT.cu code? It looks related to
>     device code generation because building was successful.
>
>     On Tue, Apr 5, 2016 at 1:56 PM, Peter Steinbach <steinbac at mpi-cbg.de
>     <mailto:steinbac at mpi-cbg.de>> wrote:
>
>         Hi guys,
>
>         first of all, please accept my apologies for contacting you by
>         mail. I was a bit lost, which mailing list to choose from as
>         pointed to by
>         http://llvm.org/docs/CompileCudaWithLLVM.html
>         and the subsequent
>         http://llvm.org/docs/#mailing-lists
>         Feel free to deflect this request to the relevant mailing list
>         or bug tracker.
>
>         In any case, I am very interested in using GPUCC in favor of
>         NVCC for a multitude of reasons (C++1X, compilation speed, ...).
>         I started to "port" my favorite samples from the nvidia SDK.
>         With clang 3.8, samples-7.0.28/7_CUDALibraries/simpleCUFFT as
>         compiled with clang produces an error at runtime! Here is what I
>         see with a K20c:
>
>         $ clang++ --cuda-path=/sw/apps/cuda/7.0.28   -I../../common/inc
>         -m64    --cuda-gpu-arch=sm_35 --cuda-gpu-arch=sm_35 -o
>         simpleCUFFT.o -c simpleCUFFT.cu
>         $ clang++ --cuda-path=/sw/apps/cuda/7.0.28
>         -L/sw/apps/cuda/7.0.28/lib64 -lcudart -ldl -lrt -pthread  -m64
>              -o simpleCUFFT.llvm simpleCUFFT.o  -lcufft
>         $ ./simpleCUFFT.llvm
>         [simpleCUFFT] is starting...
>         GPU Device 0: "Tesla K20c" with compute capability 3.5
>
>         Transforming signal cufftExecC2C
>         Launching ComplexPointwiseMulAndScale<<< >>>
>         simpleCUFFT.cu(132) : getLastCudaError() CUDA error : Kernel
>         execution failed [ ComplexPointwiseMulAndScale ] : (8) invalid
>         device function.
>
>         The same source code works just fine with nvcc 7.0.
>         Any help would be appreciated.
>
>         Best,
>         Peter
>
>         PS. From random comments, I had the feeling that you are looking
>         at the SHOC benchmarks with gpucc. If so, please comment on:
>         https://github.com/vetter/shoc/issues/48
>         I don't wanna do work that is either pointless (support for
>         textures) or was already done. ;)
>         --
>         Peter Steinbach, Dr. rer. nat.
>         HPC Developer, Scientific Computing Facility
>
>         Max Planck Institute of Molecular Cell Biology and Genetics
>         Pfotenhauerstr. 108
>         01307 Dresden
>         Germany
>
>
>         phone +49 351 210 2882 <tel:%2B49%20351%20210%202882>
>         fax +49 351 210 1689 <tel:%2B49%20351%20210%201689>
>         www.mpi-cbg.de <http://www.mpi-cbg.de>
>
>
>
>

-- 
Peter Steinbach, Dr. rer. nat.
HPC Developer, Scientific Computing Facility

Max Planck Institute of Molecular Cell Biology and Genetics
Pfotenhauerstr. 108
01307 Dresden
Germany


phone +49 351 210 2882
fax   +49 351 210 1689
www.mpi-cbg.de