# unsupported device type **URL:** \ **Category:** Legacy PGI Compilers (archived)\ **Created:** [July 25, 2010, 9:19am UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874 "2010-07-25T09:19:40Z")\ **Posts on this page:** 7\ **Page:** 1
**Author:** ![mhsuan](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@mhsuan](https://forums.developer.nvidia.com/u/mhsuan)\ **Post date:** [July 25, 2010, 9:19am UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/1 "2010-07-25T09:19:40Z")
I have downloaded from the internet a simple MPI+PGI CUDA source code as follows. * * * ! 08/07/2009 I.N. Kozin, STFC Daresbury Lab ! igor.kozin @ [stfc.ac.uk](http://stfc.ac.uk) ! ! simple direct matrix-matrix multiplication ! c(n3,n1) = a(n3,n2) x b(n2,n1) ! using MPI, PGI accelerator compiler and device allocation ! ! user selectable parameters: ! work in single or double precision ! integer, parameter :: WP ! do or don’t do device allocation ! logical, parameter :: device\_allocation ! ! how to compile for GPU: ! mpif90 -fast -o matrix-mult-mpi.exe matrix-mult-mpi.f90 -ta=nvidia -Minfo ! ! how to compile for CPU: ! make sure device\_allocation = .FALSE. ! mpif90 -fast -o matrix-mult-mpi.exe matrix-mult-mpi.f90 -Minfo ! ! how to run: ! mpiexec -n 4 -machinefile hosts ./matrix-mult-mpi.exe n1 n2 n3 iter module precision ! select single or double precision !integer, parameter :: WP = KIND(1.0) integer, parameter :: WP = KIND(1.0D0) end module precision program main use accel\_lib use mpi use precision implicit none !logical, parameter :: device\_allocation = .FALSE. logical, parameter :: device\_allocation = .TRUE. integer, parameter :: nn = 1000 integer :: i,j,k,n1,n2,n3,iter,n1t real(WP), allocatable :: a(:,:), b(:,:) real(WP), allocatable, target :: ct(:,:) real(WP), pointer :: c(:,:) real mflop, msize double precision :: tstart, tend, walltime integer :: narg,iargc integer :: ierr, size, rank, n1size, n1start, n1end integer :: ibuffer(4) character\*10 :: arg integer stat(MPI\_STATUS\_SIZE), columntype integer :: mytag = 1 ! accel integer(acc\_device\_kind) devicetype integer devicenum, num\_devices n1=nn; n2=nn; n3=nn; iter=2 write(6,\*) “step 1” call MPI\_INIT(ierr) call MPI\_COMM\_SIZE( MPI\_COMM\_WORLD, size, ierr) call MPI\_COMM\_RANK( MPI\_COMM\_WORLD, rank, ierr) write(6,\*) “step 2” !=============================================== if(rank == 0) then ! input processing narg=iargc() ! check command line input if (narg == 0) then write(6,\*) “./matrix-mult-mpi.exe n1 n2 n3 iter” stop end if if (narg \> 0) then call getarg(1,arg) read(arg,‘(i10)’) n1 end if if (narg \> 1) then call getarg(2,arg) read(arg,‘(i10)’) n2 end if if (narg \> 2) then call getarg(3,arg) read(arg,‘(i10)’) n3 end if if (narg \> 3) then call getarg(4,arg) read(arg,‘(i10)’) iter end if endif write(6,\*) “step 3” !=============================================== ! initialise GPU ! set the accelerator device if(device\_allocation) then ! devicetype = 0 ! none ! devicetype = 2 ! host devicetype = 3 ! nvidia ! write(6,‘(a12,i10)’) ’ devicetype=',devicetype ! get number of devices ! num\_devices = acc\_get\_num\_devices(devicetype) num\_devices = 1 write(6,\*) “step 4” if(rank == 0) write(6,_) ‘number of devices per host=’, num\_devices devicenum = mod(rank,num\_devices) write(6,_) ’ rank=‘, rank, ’ devicenum=’, devicenum ! set accel device call acc\_set\_device\_num(devicenum,devicetype) ! call acc\_set\_device(3) write(6,\*) “step 5” call acc\_init(devicetype) endif … (not shown due to irrelevance) * * * The log of compiling and linking is * * * PVF Build Log Begin rebuild: clean project Deleting intermediate and output files for project ‘MPI\_CUDA\_Test’, configuration ‘Debug’ Compiling Project … C:\Program Files\Microsoft HPC Pack 2008 SDK\Include\mpi.f90 c:\program files\pgi\win32\10.4\bin\pgfortran.exe -Hx,123,8 -Hx,123,0x40000 -Hx,0,0x40000000 -Mx,0,0x40000000 -Hx,0,0x20000000 -g -Bstatic -Mbackslash -Mmpi=msmpi -Mcuda -Mfree -I"c:\program files\pgi\win32\10.4\include" -I"C:\Program Files\PGI\Microsoft Open Tools 9\include" -I"C:\Program Files\PGI\Microsoft Open Tools 9\PlatformSDK\include" -Ktrap=fp -ta=nvidia,fastmath,wait -Minform=warn -Mprof=func,msmpi,dwarf -module “Win32\Debug” -Minfo=accel,ccff,ftn,intensity,loop,lre,opt,par,vect -o “Win32\Debug\mpi.obj” -c “C:\Program Files\Microsoft HPC Pack 2008 SDK\Include\mpi.f90” ## Command exit code: 0 \ matrix-mult-mpi.f90 c:\program files\pgi\win32\10.4\bin\pgfortran.exe -Hx,123,8 -Hx,123,0x40000 -Hx,0,0x40000000 -Mx,0,0x40000000 -Hx,0,0x20000000 -g -Bstatic -Mbackslash -Mmpi=msmpi -Mcuda -Mfree -I"c:\program files\pgi\win32\10.4\include" -I"C:\Program Files\PGI\Microsoft Open Tools 9\include" -I"C:\Program Files\PGI\Microsoft Open Tools 9\PlatformSDK\include" -Ktrap=fp -ta=nvidia,fastmath,wait -Minform=warn -Mprof=func,msmpi,dwarf -module “Win32\Debug” -Minfo=accel,ccff,ftn,intensity,loop,lre,opt,par,vect -o “Win32\Debug\matrix-mult-mpi.obj” -c “D:\MPI\_CUDA\_Test\matrix-mult-mpi.f90” Command exit code: 0 ## Command output: [main: 197, Possible copy in and copy out of c in call to mmult mmult: 227, Generating copyin(b(:n2,:n1)) Generating copyin(a(:n3,:n2)) Generating copy(c(:n3,:n1)) Generating compute capability 1.3 kernel 229, Loop is parallelizable 231, Loop carried dependence of ‘c’ prevents parallelization Loop carried backward dependence of ‘c’ prevents vectorization 233, Loop is parallelizable Accelerator kernel generated 229, !$acc do parallel, vector(16) 231, !$acc do seq Cached references to size [16x16] block of ‘a’ Cached references to size [16x16] block of ‘b’ 233, !$acc do parallel, vector(16) Using register for ‘c’ ] \ Linking… c:\program files\pgi\win32\10.4\bin\pgfortran.exe -Wl,/libpath:“c:\program files\pgi\win32\10.4\lib” -Wl,/libpath:“C:\Program Files\PGI\Microsoft Open Tools 9\lib” -Wl,/libpath:“C:\Program Files\PGI\Microsoft Open Tools 9\PlatformSDK\lib” -Yl,“C:\Program Files\PGI\Microsoft Open Tools 9\bin” -g -Bstatic -Mmpi=msmpi -Mcuda -ta=nvidia,fastmath,wait -Mprof=func,msmpi,dwarf -o “D:\MPI\_CUDA\_Test\Win32\Debug\MPI\_CUDA\_Test.exe” -Wl,/libpath:“C:\Program Files\Microsoft HPC Pack 2008 SDK\Lib\i386” “Win32\Debug\matrix-mult-mpi.obj” “Win32\Debug\mpi.obj” Command exit code: 0 Command output: [c:\program files\pgi/win32/10.4/lib\msmpi.obj : warning LNK4042: object specified more than once; extras ignored] MPI\_CUDA\_Test build succeeded. * * * The output of executing pgaccelinfo is * * * CUDA Driver Version 3010 Device Number: 0 Device Name: Tesla C2050 Device Revision Number: 2.0 Global Memory Size: 2817720320 Number of Multiprocessors: 14 Number of Cores: 448 Concurrent Copy and Execution: Yes Total Constant Memory: 65536 Total Shared Memory per Block: 49152 Registers per Block: 32768 Warp Size: 32 Maximum Threads per Block: 1024 Maximum Block Dimensions: 1024, 1024, 64 Maximum Grid Dimensions: 65535 x 65535 x 1 Maximum Memory Pitch: 2147483647B Texture Alignment 512B Clock Rate: 1147 MHz Current free memory 2701131776 Upload time (4MB) 16 microseconds Download time 15 microseconds Upload bandwidth 262144 MB/sec Download bandwidth 279620 MB/sec * * * Running the executable generates the following output * * * $ ./MPI\_CUDA\_Test.exe 800 200 200 2 Usage error: call to acc\_set\_device with unsupported device type: 4715104 step 1 step 2 step 3 step 4 number of devices per host= 1 rank= 0 devicenum= 0 * * * The explicit setting “devicetype = 3” seems not seen by and passed to the routine “acc\_set\_device\_num”. What’s wrong with it? On the other hand, if I uncomment the line “num\_devices = acc\_get\_num\_devices(devicetype)”, the linking output indicates an unresolved external symbol \_ACC\_GET\_NUM\_DEVICE. Again, why is this routine not recognized? Thank you very much. M-H Chung ---
**Author:** ![MatColgrove](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/matcolgrove/32/14047_2.png) [@MatColgrove](https://forums.developer.nvidia.com/u/MatColgrove)\ **Post date:** [July 26, 2010, 6:51pm UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/2 "2010-07-26T18:51:21Z")
Hi M-H Chung, Mixing the PGI Accelerator model with CUDA Fortran isn’t supported in the 10.4 compiler on Win32. Try removing the “-Mcuda” flag. - Mat ---
**Author:** ![mhsuan](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@mhsuan](https://forums.developer.nvidia.com/u/mhsuan)\ **Post date:** [July 27, 2010, 12:35am UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/3 "2010-07-27T00:35:56Z")
Removing -Mcuda still generates similar errors: * * * $ ./MPI\_CUDA\_Test.exe 800 200 200 2 step 1 step 2 step 3 step 4 number of devices per host= 1 rank= 0 devicenum= 0 Usage error: call to acc\_set\_device with unsupported device type: 4715104 * * * BTW, the error message moves to the firstline in the terminal out text if the code is run under the cygwin bash shell. * * * $ ./MPI\_CUDA\_Test.exe 800 200 200 2 Usage error: call to acc\_set\_device with unsupported device type: 4715104 step 1 step 2 step 3 step 4 number of devices per host= 1 rank= 0 devicenum= 0 * * * Thank you very much. M-H Chung ---
**Author:** ![MatColgrove](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/matcolgrove/32/14047_2.png) [@MatColgrove](https://forums.developer.nvidia.com/u/MatColgrove)\ **Post date:** [July 27, 2010, 10:54pm UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/4 "2010-07-27T22:54:18Z")
Hi M-H Chung, I tried without success to recreate your issue here, so I’m not sure what’s wrong. > Usage error: call to acc\_set\_device with unsupported device type: 4715104 This seems to imply that value of ‘devicetype’ contains the number 4715104, which is most likely a garbage value. It should be ‘3’. Can you double check that ‘devicetype’ does get set? Also, can you print out ‘devicetype’ right before the call to ‘acc\_set\_device\_num’? ```plaintext print *, 'Device Type: ', devicetype, ' Dev Num:', devicenum call acc_set_device_num(devicenum,devicetype) ``` - Mat ---
**Author:** ![mhsuan](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@mhsuan](https://forums.developer.nvidia.com/u/mhsuan)\ **Post date:** [July 28, 2010, 2:17am UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/5 "2010-07-28T02:17:07Z")
Hi Mat, The terminal output after adding the two lines you suggested is ```nohighlight D:\MPI_CUDA_Test\Win32\Debug>.\MPI_CUDA_Test.exe 800 200 200 2 step 1 step 2 step 3 step 4 number of devices per host= 1 rank= 0 devicenum= 0 Device Type: 3 Dev Num: 0 Usage error: call to acc_set_device with unsupported device type: 4715132 ``` It’s so strange! M-H ---
**Author:** ![MatColgrove](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/matcolgrove/32/14047_2.png) [@MatColgrove](https://forums.developer.nvidia.com/u/MatColgrove)\ **Post date:** [July 28, 2010, 4:33pm UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/6 "2010-07-28T16:33:38Z")
Hi M-H, Can you try updating your compiler to version 10.6 or wait til next week when 10.8 is out with CUDA 3.1 support? We just added support for Fermi in 10.4 and I wondering if there was a compatibility issue. - Mat ---
**Author:** ![mhsuan](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@mhsuan](https://forums.developer.nvidia.com/u/mhsuan)\ **Post date:** [July 30, 2010, 3:26am UTC](https://forums.developer.nvidia.com/t/unsupported-device-type/131874/7 "2010-07-30T03:26:31Z")
Hi Mat, Exactly the same error message occurred when I ran the code compiled by v10.6. Any other possible reasons? Thanks a lot. M-H