Bingo! On Fri, Nov 5, 2021 at 12:54 PM Satish Balay <[email protected]> wrote:
Yeah remove [C,CPP,CXX,CXXPP,F] FLAGS
CPP defaults to '$CC -E' - so with "--with-cc='cc -mp=gpu'" - it should use "cc -mp=gpu -E"
Satish
On Fri, 5 Nov 2021, Mark Adams wrote:
How about CPP flags?
On Fri, Nov 5, 2021 at 12:29 PM Satish Balay <[email protected]> wrote:
I guess another way to deal with this is: not use CFLAGS etc..
--with-cc='cc -mp=gpu' --with-cxx='CC -mp=gpu' --with-fc='ftn -mp=gpu'
Satish
On Fri, 5 Nov 2021, Mark Adams wrote:
Yes, thanks. I emailed the NERSc person and told him where we are and that we could fix this manually, but I don't have this issue on Summit ... He did not understand for the longest time, not sure he does now, that we do not add -gpu. And I gave him your reproducer. With any luck, with the reproducer, they can figure this out at some point. MPI is not working well for my scaling studies and my app has also given up on nvhpc, so we are good.
On Fri, Nov 5, 2021 at 9:49 AM Satish Balay <[email protected]> wrote:
For now - you could manually edit petscvariables and remove -mp=gpu from it.
Its primarily required to make configure happy.
Satish
On Thu, 4 Nov 2021, Barry Smith wrote:
This comes from the persistent problem with PETSc's make
system using too many flags for compiling CUDA that have not been tested by configure. See below the -mp=gpu is provided probably from the CPPFLAGS or CXXPPFLAGS(sp) that is improperly used by the PETSc makefiles to compile CUDA code!
Using CUDA compile:
/global/common/software/nersc/cos1.3/cuda/11.3.0/bin/nvcc -o .o -I/opt/cray/pe/mpich/8.1.10/ofi/nvidia/20.7/include -I/opt/cray/pe/libsci/ 21.08.1.2/NVIDIA/20.7/x86_64/include -I/opt/cray/pe/pmi/6.0.14/include -I/opt/cray/pe/dsmml/0.2.2/dsmml//include -I/opt/cray/xpmem/2.2.40-7.0.1.0_3.1__g1d7a24d.shasta/include -g -Xcompiler -rdynamic -DLANDAU_DIM=2 -DLANDAU_MAX_SPECIES=10 -DLANDAU_MAX_Q=4 -Xcompiler -fPIC -std=c++17 -gencode arch=compute_80,code=sm_80 -Wno-deprecated-gpu-targets -c --compiler-options=-I/global/homes/m/madams/petsc/include
-I/global/homes/m/madams/petsc/arch-perlmutter-opt-nvidia-cuda/include
-I/global/common/software/nersc/cos1.3/cuda/11.3.0/include -mp=gpu -I/global/homes/m/madams/petsc/include
-I/global/homes/m/madams/petsc/arch-perlmutter-opt-nvidia-cuda/include
-I/global/common/software/nersc/cos1.3/cuda/11.3.0/include
On Nov 4, 2021, at 7:31 PM, Mark Adams <[email protected]>
wrote:
OK, configure is done. Maybe I have too many -mp=gpu
CUDAC
arch-perlmutter-opt-nvidia-cuda/obj/sys/classes/random/impls/curand/curand2.o
gcc: error: unrecognized command line option ‘-mp=gpu’; did you mean ‘-mpku’?
On Thu, Nov 4, 2021 at 5:51 PM Barry Smith <[email protected] <mailto: [email protected]>> wrote:
Need the same thing for the C++ preprocessor flag
On Nov 4, 2021, at 5:44 PM, Mark Adams <[email protected] <mailto: [email protected]>> wrote:
It gets a lot further.
On Thu, Nov 4, 2021 at 5:32 PM Mark Adams <[email protected] <mailto: [email protected]>> wrote: OK, sorry I missed the CPPFLAGS. It is running now. Thanks,
On Thu, Nov 4, 2021 at 4:43 PM Satish Balay < [email protected] <mailto:[email protected]>> wrote: Multiple e-mail threads on the same issue (:
As suggested in my earlier thread - add -mp=gpu to both CPPFLAGS and CFLAGS [or LDFLAGS]
Satish
------- Executing: cc -o /tmp/petsc-Vvs8_T/config.setCompilers/conftest -g -DLANDAU_DIM=2 -DLANDAU_MAX_SPECIES=10 -DLANDAU_MAX_Q=4 /tmp/petsc-Vvs8_T/config.setCompilers/conftest.o Possible ERROR while running linker: stderr: nvc-Warning-The -gpu option has no effect unless a language-specific option to enable GPU code generation is used (e.g.: -acc, -mp=gpu, -stdpar, -cuda)
On Thu, 4 Nov 2021, Mark Adams wrote:
It is CPPFLAGS. I seem to get the same behavior.
FWIW, I did get this response from NERSc but I don't know how to interpret it.
He seems to be saying that I don't need -mp=gpu for the device compiler (nvcc). He seems to think that I am adding -gpu.
nvcc -- The device compiler does not need any of those flags because it already knows that it's being fed cuda code. The warning you're seeing is coming from nvc (which is the host / CPU side compiler) if you're in the PrgEnv-nvidia environment. You should not need to add -mp=gpu and -cuda, please just add the -cuda flag (to your host code) not to the device code.
I will try to talk with this guy again.
Thanks,
On Thu, Nov 4, 2021 at 4:11 PM Barry Smith < [email protected] <mailto:[email protected]>> wrote:
Yes, you need to use the CPPFLAGS which maybe called
CPPCFLAGS I am not
sure
On Nov 4, 2021, at 3:23 PM, Mark Adams <[email protected] <mailto: [email protected]>> wrote:
Ah, CCFLAGS does not seem to work.
On Thu, Nov 4, 2021 at 3:07 PM Barry Smith < [email protected] <mailto:[email protected]>> wrote:
You have to pass in the flag to turn off the bitching
about -gpu to the
C preprocessor, not the C compiler.
stderr: nvc-Warning-The -gpu option has no effect unless a language-specific option to enable GPU code generation is used (e.g.: -acc, -mp=gpu, -stdpar, -cuda) Source: #include "confdefs.h" #include "conffix.h" #include <stdlib.h>
On Nov 4, 2021, at 2:49 PM, Mark Adams < [email protected] <mailto:[email protected]>> wrote:
on Perlmutter with nvhpc:
Defined make macro "CPP" to "cc --use cpp32" Preprocessing source: #include "confdefs.h" #include "conffix.h" #include <stdlib.h>
Executing: cc --use cpp32 -I/tmp/petsc-jV9U1b/config.setCompilers /tmp/petsc-jV9U1b/config.setCompilers/conftest.c Possible ERROR while running preprocessor: exit code 1 stderr: nvc-Error-Unknown switch: --use Source: <configure.log>
<configure.log>
<configure.log>
<make.log><configure.log>
participants (1)
-
Mark Adams