kernel usage when compiling CUDA Fortran

[Tuan]

In C CUDA, we can print out the kernel usage with -ptxas-options=-v

This is a very helpful feature to tune the code. I don’t whether we can do this in CUDA Fortran. I search the manual, yet there is no clue.

Thank you,
Tuan.

[MatColgrove]

Hi Tuan,

No ETA on when it will be available or the exact form, but we are working on getting this information back to the user.

• Mat

[TheMatt]

@Tuan

In C CUDA, we can print out the kernel usage with -ptxas-options=-v

This is a very helpful feature to tune the code. I don’t whether we can do this in CUDA Fortran. I search the manual, yet there is no clue.

As Mat said, there is, as yet, no equivalent to -ptxas-options=-v with CUDA Fortran. But, you can get the same information if you use the keep options. -Mcuda=keepbin will write the equivalent of the .cubin file which contains the smem, lmem, reg information.

[Tuan]

@TheMatt

@Tuan

In C CUDA, we can print out the kernel usage with -ptxas-options=-v

This is a very helpful feature to tune the code. I don’t whether we can do this in CUDA Fortran. I search the manual, yet there is no clue.

As Mat said, there is, as yet, no equivalent to -ptxas-options=-v with CUDA Fortran. But, you can get the same information if you use the keep options. -Mcuda=keepbin will write the equivalent of the .cubin file which contains the smem, lmem, reg information.

I read one thread from NVIDIA forum that, such information in the cubin is not exactly the same as the one generated by -ptxas-options=-v. No one know how the values in the cubin are generated as it is internal.

Do you have the same experience with this.
Thanks,
Tuan.