PGI User Forum

[alechand]

Hello,
i really appreciate your attention.

About the WRITE statement in the Mol_Dyn.f90 code, i tried to use it just one time. It worked one time. In the second time, the memory errors begin to appear on all codes, even if i remove the WRITE statement, and also in this simple program picalc.f90. What i thought is that this statement caused, for example, problems in the graphic card.

About the picalc.f90, i did not change nothing. I just installed a new system (kubuntu 11.04) and a new PGI compiler.

Here i show the complete output as you asked (i use exactly the same code as you, i've copied it) :

##########################################################

alechand@pcsantos2:~/test_openacc$ more picalc.f90
program picalc
implicit none
integer, parameter :: n=1000000
integer :: i
real(kind=8) :: t, pi
pi = 0.0
!$acc parallel loop
do i=0, n-1
t = (i+0.5)/n
pi = pi + 4.0/(1.0 + t*t)
end do
!$acc end parallel loop
print *, 'pi=', pi/n
end program picalc
alechand@pcsantos2:~/test_openacc$ pgfortran -fast -Minfo=all -o MOL_DYN picalc.f90 -ta=nvidia,4.2 -V12.10
picalc:
7, Accelerator kernel generated
7, CC 1.3 : 24 registers; 32 shared, 36 constant, 0 local memory bytes
CC 2.0 : 23 registers; 0 shared, 52 constant, 0 local memory bytes
8, !$acc loop gang, vector(256) ! blockidx%x threadidx%x
10, Sum reduction generated for pi
7, Generating compute capability 1.3 binary
Generating compute capability 2.0 binary
alechand@pcsantos2:~/test_openacc$ PGI_ACC_DEBUG=1
alechand@pcsantos2:~/test_openacc$ ./MOL_DYN
call to cuMemcpyDtoH returned error 700: Launch failed
CUDA driver version: 5050

##########################################################

I dont know if i did something wrong, but the command PGI_ACC_DEBUG=1 seems to dont work.

thanks

[mkcolg]

[alechand]

Here is the output again :

######################################

alechand@pcsantos2:~/test_openacc$ more picalc.f90
program picalc
implicit none
integer, parameter :: n=1000000
integer :: i
real(kind=8) :: t, pi
pi = 0.0
!$acc parallel loop
do i=0, n-1
t = (i+0.5)/n
pi = pi + 4.0/(1.0 + t*t)
end do
!$acc end parallel loop
print *, 'pi=', pi/n
end program picalc
alechand@pcsantos2:~/test_openacc$ pgfortran -fast -Minfo=all -o MOL_DYN picalc.f90 -ta=nvidia,4.2 -V12.10
picalc:
7, Accelerator kernel generated
7, CC 1.3 : 24 registers; 32 shared, 36 constant, 0 local memory bytes
CC 2.0 : 23 registers; 0 shared, 52 constant, 0 local memory bytes
8, !$acc loop gang, vector(256) ! blockidx%x threadidx%x
10, Sum reduction generated for pi
7, Generating compute capability 1.3 binary
Generating compute capability 2.0 binary
alechand@pcsantos2:~/test_openacc$ export PGI_ACC_DEBUG=1
alechand@pcsantos2:~/test_openacc$ ./MOL_DYN
__pgi_cu_init() found 1 devices
__pgi_cu_init( file=/home/alechand/test_openacc/picalc.f90, function=picalc, line=7, startline=1, endline=14 )
__pgi_cu_init() will use device 0 (V3.0)
__pgi_cu_init() compute context created
__pgi_cu_module3( lineno=7 )
__pgi_cu_module3 module loaded at 0x928ba28
__pgi_cu_module_function( name=0x8098c4a=picalc_7_gpu, lineno=7, argname=(nil)=, argsize=8, varname=0x8098c57=b1, varsize=8, SWcachesize=0 )
Function handle is 0x93752a0
__pgi_cu_module_function( name=0x8098c38=picalc_10_gpu_red, lineno=7, argname=(nil)=, argsize=0, varname=(nil)=, varsize=0, SWcachesize=0 )
Function handle is 0x9372bf0
__pgi_cu_alloc(size=31256,lineno=7,name=)
__pgi_cu_alloc(31256) returns 0x206c0000
__pgi_cu_uploadc( "b1", size=8, offset=0, lineno=7 )
constant data b1 at address 0x205c0000 devsize=8, size=8, offset=0
First arguments are:
0 0

0x00000000 0x00000000
__pgi_cu_launch_a(func=0x93752a0, grid=3907x1x1, block=256x1x1, lineno=7)
__pgi_cu_launch_a(func=0x93752a0, params=0xbf98c97c, bytes=4, sharedbytes=2048)
First arguments are:
543948800

0x206c0000
__pgi_cu_launch_a(func=0x9372bf0, grid=1x1x1, block=256x1x1, lineno=10)
__pgi_cu_launch_a(func=0x9372bf0, params=0xbf98c97c, bytes=8, sharedbytes=2048)
First arguments are:
543948800 3907

0x206c0000 0x00000f43
__pgi_cu_downloadc( "b1", size=8, offset=0, lineno=7 )
call to cuMemcpyDtoH returned error 700: Launch failed
CUDA driver version: 5050

############################

Are you seeing something strange ?

thanks a lot

[mkcolg]

Given the debug output, it looks like you're compiling to 32-bits? Can you try compiling in 64-bits? (i.e. add "-m64" to you compile options).

The program still works fine for me in 32-bits, but I'm just wondering.

- Mat

[alechand]

But, my system is 32bits, and i installed the 32bits version of the compiler. Do you want to i install the 64 bits version of the compiler to test ?
Can i install it in a 32 bits ubuntu version ?

thanks