An Easy Introduction to CUDA Fortran

Technical Blogs & Events Technical Blog
1

Originally published at: https://developer.nvidia.com/blog/easy-introduction-cuda-fortran/

CUDA Fortran for Scientists and Engineers shows how high-performance application developers can leverage the power of GPUs using Fortran. This post is the first in a series on CUDA Fortran, which is the Fortran interface to the CUDA parallel computing platform. If you are familiar with CUDA C, then you are already well on your…

2

Hi dear all
I am a beginner to CUDA (CUDA FORTRAN) . I have installed PGI 13.9 but the problem is, when I try to debug even a very simple CUDA Fortran code I get plenty of errors such as below:
Error 1 unresolved external symbol cudaSetupArgument referenced in function mathops_saxpy_ saxp.obj
Error 2 unresolved external symbol cudaLaunch referenced in function mathops_saxpy_ saxp.obj
Error 3 unresolved external symbol __cudaRegisterFatBinary referenced in function mathops_saxpy_ saxp.obj
Error 4 unresolved external symbol __cudaRegisterFunction referenced in function mathops_saxpy_ saxp.obj
Error 5 unresolved external symbol __cudaUnregisterFatBinary referenced in function mathops_saxpy_ saxp.obj
Error 6 unresolved external symbol pgf90_dev_auto_alloc04 referenced in function MAIN_ saxp.obj
Error 7 unresolved external symbol pgf90_dev_copyin referenced in function MAIN_ saxp.obj
.
.
.
.Error 12 unresolved external symbol CUDAFOR saxp.obj

Furthermore, when I check the project properties I see that the CUDA FOR is not even enabled when I enable it and debug again I get the same errors. I will be appreciated if someone help me with this problem.
Thanks,
Reza

4

Hi,
i tried to run this script and it returned 'Max error: 2.0000'.
Where this error come from?
from the cudaDeviceProp in Fortran CUDA i got
Device Number: 0
Device name: GeForce GTX 1060 3GB
Memory Clock Rate (KHz): 4004000
Memory Bus Width (bits): 192
Peak Memory Bandwidth (GB/s): 192.19

5

this is work when I use pgf90 -Mcuda=cc60 -o saxpy saxpy.cuf to compile

6

Thanks! I have exact same problem with you. And it can be solved by your solution.

7

What we really need is to discuss why the different in "results" depending upon the compute capability (or compiler version)

8

Hi, I am experiencing the same problem as you did. It seems to me that the kernel subroutine never return any value to the host. I have tried your suggestion but still not working

9

I have a problem about the p2pBandwidth code on Page 128. Line 52-53 and Line 55-56 are the same. Is it right ?

50 do i = 0, nDevices -1
51 if (i == j) cycle
52 istat = cudaMemcpyPeer ( distArray (j )% a_d , j , &
53 distArray (i )% a_d , i , N)
54 istat = cudaEventRecord ( startEvent ,0)
55 istat = cudaMemcpyPeer ( distArray (j )% a_d , j , &
56 distArray (i )% a_d , i , N)
57 istat = cudaEventRecord ( stopEvent ,0)
58 istat = cudaEventSynchronize ( stopEvent )
59 istat = cudaEventElapsedTime ( time , &
60 startEvent , stopEvent )

I think Line 52-53 should be removed. Maybe I was wrong. If not, could you give me a explanation ?