cuda runtime error

[renbem]

Hi,

it took a while but now it is possible to build the package on our system. We had to do some changes in the file reconstruction_cuda2.cu to get rid of errors related to thrust. Attached there is the file containing the compilation with the error output and the corresponding changes we made to build the code successfully.

These changes were tested on the following systems:

1. System with

• x86_64 CPU
• Ubuntu 14.04.3 LTS
• Nvidia GeForce GTX 980 Ti
• Cuda 7.5

2. System with

• x86_64 CPU
• elementary OS Freya (based on Ubuntu 14.04)
• Nvidia Quadro K4000 (Compute Capability = 3)
• Cuda 7.0

The described error ocurred on both systems and could be resolved with the mentioned editings. However, during runtime it throws the error message

on system 1. (System 2 can't be used in GPU mode anyway given its insufficient compute capability). Both systems work fine when run with the useCPU flag.

Do you know what can be done to get the GPU computation work? Given that I'm not experienced with cuda I'd be grateful for some hints where and what to look for.

Best,
Michael

error.txt
changes_in_reconstruction_cuda2_cu.txt

[bkainz]

Dear Michael,

this looks like you try to run compute capability 3.5 device code on a
compute capability 5.0 device. Try to change the arch to sm_50.
We are currently working on a new version, which shouldn't have these
problems. However, it is not quite ready yet to be released.

best wishes,
Bernhard

On 27/11/2015 16:30, Michael Ebner wrote:

Hi,

it took a while but now it is possible to build the package on our
system. We had to do some changes in the file reconstruction_cuda2.cu
to get rid of errors related to thrust. Attached there is the file
containing the compilation with the error output and the corresponding
changes we made to build the code successfully.

These changes were tested on the following systems:

1. System with

• x86_64 CPU
• Ubuntu 14.04.3 LTS
• Nvidia GeForce GTX 980 Ti
• Cuda 7.5

2. System with

• x86_64 CPU
• elementary OS Freya (based on Ubuntu 14.04)
• Nvidia Quadro K4000 (Compute Capability = 3)
• Cuda 7.0

The described error ocurred on both systems and could be resolved with
the mentioned editings. However, during runtime it throws the error
message
screen shot 2015-11-27 at 16 09 54
https://cloud.githubusercontent.com/assets/11680758/11445101/596a2bd2-9521-11e5-9327-6c30fc50f4ea.png
on system 1. (System 2 can't be used in GPU mode anyway given its
insufficient compute capability). Both systems work fine when run with
the useCPU flag.

Do you know what can be done to get the GPU computation work? Given
that I'm not experienced with cuda I'd be grateful for some hints
where and what to look for.

Best,
Michael

error.txt
https://github.com/bkainz/fetalReconstruction/files/45923/error.txt
changes_in_reconstruction_cuda2_cu.txt
https://github.com/bkainz/fetalReconstruction/files/45922/changes_in_reconstruction_cuda2_cu.txt

—
Reply to this email directly or view it on GitHub
#4.

Dr. Bernhard Kainz
Department of Computing
Imperial College London
180 Queens' Gate
London SW7 2AZ
mailto:B.Kainz@imperial.ac.uk

[renbem]

Thanks for the quick reply!

I changed the line in the source/reconstructionGPU2/CMakeLists.txt to

set(CUDA_NVCC_FLAGS "${CUDA_NVCC_FLAGS};-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_35,code=sm_35;-use_fast_math;")

Now I get a different error at a later stage:
Iteration 0.
CUDA error at /home/mebner/Development/IRTK_BKainz/fetalReconstruction/source/reconstructionGPU2/reconstruction_cuda2.cu:2437 code=77(cudaErrorIllegalAddress) "cudaDeviceSynchronize()"

Best,
Michael

[bkainz]

Hi,

perhaps you have two devices in your rack, one for display and one for
computing? Try to specify the device with -d

best wishes,
Bernhard

Am 27.11.2015 um 17:19 schrieb Michael Ebner:

Thanks for the quick reply!

I changed the line in the source/reconstructionGPU2/CMakeLists.txt to

set(CUDA_NVCC_FLAGS
"${CUDA_NVCC_FLAGS};-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_35,code=sm_35;-use_fast_math;")

Now I get a different error at a later stage:
Iteration 0.
CUDA error at
/home/mebner/Development/IRTK_BKainz/fetalReconstruction/source/reconstructionGPU2/reconstruction_cuda2.cu:2437
code=77(cudaErrorIllegalAddress) "cudaDeviceSynchronize()"

Best,
Michael

—
Reply to this email directly or view it on GitHub
#4 (comment).

Department of Computing
Imperial College London
180 Queens' Gate
London SW7 2AZ
mailto:B.Kainz@imperial.ac.uk