possible bug: gpu_time intrinsic

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

According to the cuda fortran user guide (section 3.6.4), the GPU_TIME intrinsic should be callable from device code (it appears to correspond to the cuda-C clock() function). However, when I try to call it I get a compiler error:

PGF90-S-0155-Illegal call from device code to gpu_time (smvgear-dev.cuf: 2216)
0 inform, 0 warnings, 1 severes, 0 fatal for trace
make: *** [smvgear-dev.o] Error 2
make: Target `smvgear-gpu’ not remade because of errors.

The compiler seems to think that gpu_time is a host-only subroutine, but the user guide suggests just the opposite. Is this an error in the compiler or in the manual? Does the gpu_time intrinsic correspond to the clock() function in cuda-C? If not, then is there something that does?

#2

Hi Robert,

We’re still in the process of adding these intrinsics and GPU_TIME is not there yet. While I don’t have an eta, I would expect it to be available in the near future. Our engineers just got the warp vote intrinsics into 10.6 so I expect them to finish up the rest of them here shortly.

  • Mat