[CUDA Fortran] .gpu error

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

Hello,

I have a weird error when compiling CUDA Fortran.

I receive

/tmp/pgcudafor6tecUW6Srfqs.gpu(29): error: expected an identifier
/tmp/pgcudafor6tecUW6Srfqs.gpu(138): error: expected a "("
/tmp/pgcudafor6tecUW6Srfqs.gpu(138): error: expression must have integral or enum type
/tmp/pgcudafor6tecUW6Srfqs.gpu(142): error: expected an expression
/tmp/pgcudafor6tecUW6Srfqs.gpu(142): error: expected an expression
5 errors detected in the compilation of "/tmp/pgnvdpuecTHvkmQAw.nv0".
PGF90-F-0000-Internal compiler error. pgnvd job exited with nonzero status code

all vars are of type float

when compiling the following code :

delta=psiroe*(abs(u)+abs(v)+c)
vl=qn
switch=(.5+sign(.5,abs(vl)-delta))
ambda1=switch*(abs(vl))+(1.-switch)* &
((vl**2+delta**2)/(2.*delta+cut0x))

or

delta=psiroe*(abs(u)+abs(v)+c)
vl=qn
switch = 1.
if(abs(vl) .lt. delta) switch = 0.
ambda1=switch*(abs(vl))+(1.-switch)* &
((vl**2+delta**2)/(2.*delta+cut0x))

but when go round the problems and doing

delta=psiroe*(abs(u)+abs(v)+c)
vl=qn
ambda1=(.5+sign(.5,abs(vl)-delta))*(abs(vl))+(1.-(.5+sign(.5,abs(vl)-delta)))* &
                ((vl**2+delta**2)/(2.*delta+cut0x))

it compiles without even a warning.

I looked at the generated .gpu and all seems ok in all cases.

Do someone know the reason of this ?

#2

Hi mbpl,

Do someone know the reason of this ?

These types of errors mean that the compiler is generating incorrect GPU code. If you can send an example code that exhibits the error to PGI customer service (trs@pgroup.com) we would appreciate it.

Note that we have just released version 10.2 which includes several fixes for CUDA Fortran and it’s possible that your error has already been corrected.

Thanks,
Mat

#3

Hi Mat,

Thanks for your response.

I will ask to know if I can send you a piece of code.

Regards,

Matthieu