Can I run KokkosKernels spgemm with float or int32 type?

[Lilian92]

I am running under ./example/buildlib/perf_test.

[srajama1]

@Lilian92 Can you give me some more details ? What is the build script you used ? The code is fully templated, so yes float is supported. All the ids are ints typically. Are you trying to multiply where the matrix entries themselves are ints ?

[Lilian92]

I am sorry for the lack of details in the question.

Myself:
I am a PhD. student at Gatech working with Prof. Ümit V. Çatalyürek.

What is the build script you used ?
./example/buildlib/compileKokkosKernelsSimple.sh

The code is fully templated, so yes float is supported. All the ids are ints typically. Are you trying to multiply where the matrix entries themselves are ints ?
Yes, I am trying to multiply with matrix having entries/value as int32 or float type.

I have tried to change:

kokkos-kernels/example/buildlib/compileKokkosKernelsSimple.sh

Line 2 in d86db11

KOKKOSKERNELS_SCALARS=double #the scalar types to instantiate =double,float...

To:
KOKKOSKERNELS_SCALARS=int

Then change:

kokkos-kernels/perf_test/sparse/KokkosSparse_spgemm.cpp

Line 45 in d86db11

#if defined(KOKKOSKERNELS_INST_DOUBLE) && \

To:
#if defined(KOKKOSKERNELS_INST_INT) && \

kokkos-kernels/perf_test/sparse/KokkosSparse_spgemm.cpp

Line 54 in d86db11

#define SCALAR_TYPE double

To:
#define SCALAR_TYPE int

kokkos-kernels/perf_test/sparse/KokkosSparse_spgemm.cpp

Line 344 in d86db11

#if !defined(KOKKOSKERNELS_INST_DOUBLE)

To:
#if !defined(KOKKOSKERNELS_INST_INT)

But during running, I will get error from:

kokkos-kernels/perf_test/sparse/KokkosSparse_spgemm.cpp

Line 345 in d86db11

std::cout << " not defined KOKKOSKERNELS_INST_DOUBLE" << std::endl;

[srajama1]

@Lilian92 Thanks for the context.

This is technically a problem with that test file which assumes double. The actual source code does not have the assumption. Can you try SCALAR to be float first ? I will see what is needed for scalar to be int.

[Lilian92]

Thank you so much for your fast reply!

I have tried with float. With float, it won't pass the compiling for spgemm.

[Lilian92]

Thanks for checking things out!

[srajama1]

Can you cut and paste the error ?

I will check.

[Lilian92]

Below is the log from '"make build-test -j". Thanks!

https://www.dropbox.com/s/hx1c3izhj4wh4bd/log?dl=0

[srajama1]

@brian-kelley As you are deep into SpGEMM, this might be a good test.

[brian-kelley]

@srajama1 I had a branch from December where I was cleaning up the sparse perf tests, and using sensible scalar, ordinal, size_t, layout types based on what is ETI'd. So I'll finish that up and make sure scalar=float works everywhere.

@Lilian92 If we get float to work, would you still want to try scalar=int? It won't be supported by the build system, but by including the right headers and plugging int into the template parameters it should be possible.

[brian-kelley]

@Lilian92 I'm sorry this took so long, it turns out we had several issues with building the performance tests with scalar_t = float (as you saw). But I have a branch that can build the SPGEMM driver with scalar_t = float, and there's also a very easy way (1-line change) to get scalar_t=int.

For now, these instructions will involve checking out my branch, but within a few days this should be in the main develop branch.

Instructions for float:

# get my branch of kokkos-kernels
git clone https://www.github.com/brian-kelley/kokkos-kernels
cd kokkos-kernels
git checkout SparsePerfCleanup
cd ..
# build 
mkdir build
cd build
# use main Kokkos repository (develop branch)
# yes, kokkos goes in the build/ directory
git clone https://www.github.com/kokkos/kokkos
cd kokkos
git checkout develop
cd ..
# configure a basic build (serial only) using float
# run the script with --help to see the options (you'll probably want to run with CUDA or OpenMP)
../kokkos-kernels/cm_generate_makefile.bash --release --with-scalars=float
# build (where $N is the number of CPU cores)
make -j $N
# and then perf_test/sparse/test_spgemm will be using floats and should work

Instructions for int:

git clone https://www.github.com/brian-kelley/kokkos-kernels
cd kokkos-kernels
git checkout SparsePerfCleanup
# need to change one line to get scalar=int
cd perf_test/sparse
### using any editor, change line 51 of KokkosSparse_spgemm.cpp to "typedef int scalar_t;"
cd ../../..
mkdir build
cd build
git clone https://www.github.com/kokkos/kokkos
cd kokkos
git checkout develop
cd ..
# configure a basic build (serial only), the scalars option doesn't matter since it's changed for the SPGEMM perf_test
# run the script with --help to see the options
../kokkos-kernels/cm_generate_makefile.bash --release
# build (where $N is the number of CPU cores)
make -j $N
# and then perf_test/sparse/test_spgemm will be using ints and should work

Note that for the int version, you'll have to have integer-formatted MatrixMarket input files. The header should look like this (maybe 'general' is different, but 'integer' should be there).

%%MatrixMarket matrix coordinate integer general

[brian-kelley]

@Lilian92 Apologies, I just realized that even if you compile with --scalars=float, double is still instantiated so it is used as the default. We'll figure out a permanent solution that lets users select the types at runtime somehow, but until then just follow the instructions for int but use "typedef float scalar_t;" instead.

[Lilian92]

Cool, thank you so much!

[srajama1]

@Lilian92 Are you able to successfully run this case ?

[Lilian92]

Yes, thank you for asking!

…

On Mon, Mar 2, 2020, 4:55 PM Siva Rajamanickam ***@***.***> wrote: @Lilian92 <https://github.com/Lilian92> Are you able to successfully run this case ? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#583?email_source=notifications&email_token=AENMP7CTINQJOK7WTAH4ZRLRFQTL7A5CNFSM4KMYKDX2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOENRENVA#issuecomment-593643220>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AENMP7C6PCJJFFXGSWWAPFTRFQTL7ANCNFSM4KMYKDXQ> .