Clang driver for OpenMP target offloading

Project Infrastructure LLVM Dev List Archives
1

Hi,

I am having some trouble to understand how the clang driver works for OpenMP offloading.

For example, if I run the following command (with the -v flag):

clang -v -g -fopenmp offloading_success.c -o offloading_success

I can see the invocations that clang does and if I run those command one by one I can successfully compile my program.

Now, if I do the same but with the offloading option, as in the following command:

clang -v -g -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda offloading_success.c -o offloading_success

the command works, compiles correctly, and when I run the program, it successfully run on the device.
However, if I show the invocations with “-v” and try to run the commands one by one it does not work and in particular I get an error with the last command which is the linking part with “ld”.
The error i get is the following:

/usr/bin/ld: cannot open linker script file /tmp/offloading_success-2b4739.lk: No such file or directory

Where is this “.lk” file generated? Is there a missing commands in the invocations?

I attached a file with the full invocation.

Thanks!
Best,
Simone

output.txt (15.2 KB)

2

Hi Simone,

you are more likely to get answers to frontend questions asking on cfe-dev. However, your output shows that you are using clang-ykt which isn't supported upstream so this is not the right place for your questions. Nevertheless, I'll answer some because the same "problem" will arise with (upstream) trunk.

Hi,

I am having some trouble to understand how the clang driver works for
OpenMP offloading.

For example, if I run the following command (with the -v flag):

clang -v -g -fopenmp offloading_success.c -o offloading_success

I can see the invocations that clang does and if I run those command
one by one I can successfully compile my program.

Just to clarify: This will compile but will fallback to run on the host because the arguments don't specify which device to compile for.

Now, if I do the same but with the offloading option, as in the
following command:

clang -v -g -fopenmp -fopenmp-targets=nvptx64-nvidia-cuda
offloading_success.c -o offloading_success

the command works, compiles correctly, and when I run the program, it
successfully run on the device.

(which would not be the case for upstream trunk yet...)

However, if I show the invocations with "-v" and try to run the
commands one by one it does not work and in particular I get an error
with the last command which is the linking part with "ld".
The error i get is the following:

/usr/bin/ld: cannot open linker script file
/tmp/offloading_success-2b4739.lk [1]: No such file or directory

Where is this ".lk" file generated? Is there a missing commands in the
invocations?

The linker script is generated by Clang's driver (both in clang-ykt and trunk) and is deleted after its usage which is the default for temporary files. If you want to inspect the intermediate files, I suggest adding --save-temps which will dump a bunch of files to the current directory. (AFAIK this will result in more files than generated by default so that developers have more intermediate steps to look at.)

Cheers,
Jonas

3

Hi Jonas,

thanks for the answer.
For now I can use the “–save-temps” flag to keep my investigation.

Do you know what’s the timeline for integrating offloading for NVidia GPUs upstream?
I remember seeing a big patch for integrating OpenMP target generation in the clang driver and I thought it was already accepted and committed, but I might be wrong.

Thanks.
Simone

4

Hi Simone,

I'm not aware of a timeline, it's ready when all necessary patches are merged. One large missing piece is the device runtime.

Jonas