Issue building ginkgo - rounding modifier required for instruction 'cvt'

[qir-at-tum]

I'm having an issue building the most recent release (Ginkgo 1.10.0) as it crashes when compiling cuda/CMakeFiles/ginkgo_cuda.dir/solver/lower_trs_kernels.cu.o for CUDA. I'm using:

CUDA 11.6.55,
Ampere architecture
Cmake 3.23.0
OpenMPI 4.0.2
GCC 9.3.0
Ubuntu 20.04 LTS.

The error is as follows: ptxas /tmp/tmpxft_0017ad67_00000000-8_lower_trs_kernels.compute_80.ptx, line 4490; error : Rounding modifier required for instruction 'cvt' ptxas fatal : Ptx assembly aborted due to errors

EDIT:
my cmake was using the following:
cmake -G "Unix Makefiles" .. -DGINKGO_FORCE_GPU_AWARE_MPI=ON -DGINKGO_CUDA_ARCHITECTURES=Ampere
EDIT2:
This issue does not happen on release 1.9.0.

Is anyone aware what is going on here?

[yhmtsai]

@qir-at-tum thanks for your report.
I guess it is something from bfloat16.
could you turn off it by -DGINKGO_ENABLE_BFLOAT16=OFF

[qir-at-tum]

Hi, yes, it is something from bfloat16. I was able to successfully build ginkgo after disabling it. Will the lack of bfloat16 affect anything negatively?

[yhmtsai]

no, if you do not need bfloat16, there will no negative impact.
Maybe cuda 11.6 does not fully support all ptx for bfloat16 yet.
We will investigate it further.