Can't compile cuda samples

[magistri]

Hi there

I have troubles compiling the cuda samples in version 10.2 on a fresh ubuntu 18.04.
here's the issue:

In file included from cudaNvSci.cpp:12:0: cudaNvSci.h:14:10: fatal error: nvscibuf.h: No such file or directory #include <nvscibuf.h> ^~~~~~~~~~~~ compilation terminated. Makefile:394: recipe for target 'cudaNvSci.o' failed make[1]: *** [cudaNvSci.o] Error 1 make[1]: Leaving directory '/home/federico/NVIDIA_CUDA-10.2_Samples/0_Simple/cudaNvSci' Makefile:51: recipe for target '0_Simple/cudaNvSci/Makefile.ph_build' failed make: *** [0_Simple/cudaNvSci/Makefile.ph_build] Error 2

any idea?

[Helenll]

Same problem! cannot find this .h file

[Evanslooten]

I am also experiencing this issue on a fresh ubuntu 18.04 install. Using Nvidia-driver-440 from apt and installing with the run file.

[Helenll]

it seems the default installation just dose not include nvscibuf.h, while the sample needs it?

[Naowak]

Hello, I encountered the same problem. It seems that you can continue the build with
make -k.

[mdoijade]

@magistri @Helenll @Evanslooten you can continue the build with make -k while using master Makefile this will allow make to skip this error due to build failure of cudaNvSci sample.

As documented in the CUDA 10.2 release notes, the NvSci interop support is a new BETA feature to CUDA 10.2 and hence it's broad support is limited at this current time. The cudaNvSci sample that is contained in the CUDA 10.2 toolkit requires a new NvStreams development package that has, to date, not yet been released. NvStreams is a product of NVIDIA Automotive SW and is already supported in DRIVE OS aarch64 Linux on DRIVE OS 5.1.12.0 . The x86_64 support package for Nvstreams is new and is due to be released in late January of 2020 as a part of the DRIVE OS 5.1.12.4 release. If you are interested to obtain the upcoming release of Nvstreams support for DRIVE OS on x86_64 please register for an account at https://developer.nvidia.com/DRIVE-AGX-program The initial release of the BETA Nvstreams x86_64 library package will be handled via https://partners.nvidia.com/ which is a credentialed access portal automotive partners for obtaining all related DRIVE OS software.

[kamrankausar]

make: Target 'all' not remade because of errors.

[twobob]

The initial release of the BETA Nvstreams x86_64 library package will at blah blah and you don't have it

Then don't shove/require/embed it in the basic build samples?

[Vfitstory]

I specifically logged in to second twobob.

[qhaas]

Seeing the following 'not founds' in EL8, CUDA 11 samples, CUDA 11 SDK + cudnn8, and proprietary driver 450.51.05, which is oddly a beta driver in the stable repo:

$ make -j$(nproc) 2>&1 | sort -u | grep -i 'not found'
>>> WARNING - libnvscibuf.so not found, please install libnvscibuf.so <<<
>>> WARNING - libnvscisync.so not found, please install libnvscisync.so <<<
>>> WARNING - nvscibuf.h not found, please install nvscibuf.h <<<
>>> WARNING - nvscisync.h not found, please install nvscisync.h <<<

yum provides and locate do not return results for said files. As mentioned in the above comment, sounds like this is expected, but I echo the opinion of others that some of these shouldn't be in the default build target...

[twobob]

Can't understand why this issue was closed tbh

[sentientmachine]

I got the same error and found a fix.

I'm running Gentoo Linux x86_64, PC with dual (two) Geforce RTX 2080Ti using root> sh NVIDIA-Linux-x86_64-418.43.run

make: Target ... not remade because of errors.

Problem being I downloaded a 2 month-old version of the driver from nvidia. So go to nvidia.com, go to geforce drivers choose your exact GPU make/model, select x86_64 Bit ask for the latest one, and whisks you to a new page with: NVIDIA-Linux-x86_64-455.45.01.run note the version upgrade to: 455.45.01 Download that and run, and and the installer reports a success.

[sephethus]

Every single answer with Linux has a new problem. Installing the latest 455.45 drivers lead to an error that an X server is running. Now I have to solve that problem. Will it ever end?

EDIT: This solution above did not solve the problem by the way.

[sentientmachine]

@sephethus
About your error of an X server is running. I got a similar sounding error because naturally the motherboard Video card fights over Both 2080Ti GPU's over "who is going to paint pixels on your monitor?" Google-kung-fu directed me to replace /usr/X11/xorg.conf with the prior one named: xorg.conf.nvidia-xconfig-original and after reboot the nvidia-smi is reporting in and helloworld cuda/gpu code is working.
Will it ever end? No, but your skills eventually rise to the point where you can keep the lights on with income > expenses. This cuda code and python3.7 torch code is a landfill fire of spaghetti, I share your pain. You have to roll a +20 Persistence to even begin to begin.

[FilippoRomeo]

bf16TensorCoreGemm.cu(176): error: more than one constructor applies to convert from "int" to "__nv_bfloat16":
            function "__nv_bfloat16::__nv_bfloat16(float)"
            function "__nv_bfloat16::__nv_bfloat16(double)"

bf16TensorCoreGemm.cu(182): error: more than one constructor applies to convert from "int" to "__nv_bfloat16":
            function "__nv_bfloat16::__nv_bfloat16(float)"
            function "__nv_bfloat16::__nv_bfloat16(double)"

2 errors detected in the compilation of "bf16TensorCoreGemm.cu".
Makefile:331: recipe for target 'bf16TensorCoreGemm.o' failed
make[1]: *** [bf16TensorCoreGemm.o] Error 1
make[1]: Leaving directory '/home/agx/NVIDIA_CUDA-11.0_Samples/0_Simple/bf16TensorCoreGemm'
Makefile:51: recipe for target '0_Simple/bf16TensorCoreGemm/Makefile.ph_build' failed
make: *** [0_Simple/bf16TensorCoreGemm/Makefile.ph_build] Error 2

same problem here, I am using arm

[Ru7w1k]

Arm will require an explicit cast.

Please replace
a[i*K_GLOBAL+j] = (__nv_bfloat16)(rand() % 3); to a[i*K_GLOBAL+j] = (__nv_bfloat16)(float)(rand() % 3);
and
b[i*K_GLOBAL+j] = (__nv_bfloat16)(rand() % 3); to b[i*K_GLOBAL+j] = (__nv_bfloat16)(float)(rand() % 3);

[alexvsad62]

sad@Sysadmin:~/tmp/cuda-samples/Samples/0_Introduction/asyncAPI> make TARGET_ARCH=x86_64
/usr/local/cuda/bin/nvcc -ccbin g++ -I../../../Common -m64 --threads 0 --std=c++11 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 -gencode arch=compute_80,code=sm_80 -gencode arch=compute_86,code=sm_86 -gencode arch=compute_89,code=sm_89 -gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90 -o asyncAPI.o -c asyncAPI.cu
g++: No such file or directory
nvcc fatal : Failed to preprocess host compiler properties.
make: *** [Makefile:328: asyncAPI.o] Ошибка 1

Which file is absent? What to do?