Add sparse matrix to mp #860
Conversation
…uted_vector rma memory access
benchmarks/gbench/mp/gemv.cpp
Outdated
|
|
||
| namespace mp = dr::mp; | ||
|
|
||
| #ifdef STANDALONE_BENCHMARK |
There was a problem hiding this comment.
I don't see in any of CMakeLists.txt building gemv.cpp with STANDALONE_BENCHMARK option. Code under this ifdef is never compiled. Please remove if you don't need it or add a target to cmakelists to make it being compiled and take care it is compiled during CI (to catch compile time errors in CI)
There was a problem hiding this comment.
Fair point, I removed it, since it is not strictly necessary for benchmarking
benchmarks/gbench/mp/gemv.cpp
Outdated
| } | ||
| } // namespace | ||
| static auto getMatrix() { | ||
| std::size_t n = std::max(1., std::sqrt(default_vector_size / 100000)) * 50000; |
There was a problem hiding this comment.
default_vector_size is to be more-less equal to size of data being allocated. However when I run all benchamrks on dnp02 host the only tests which fail with bad_alloc are Gemv_DR.
Please add explanation to the code what are 100000 and 50000 constants for.
Please make the benchamrk allocating more less the same ammount of data like other benchmars so command:\
ONEAPI_DEVICE_SELECTOR='level_zero:gpu' I_MPI_OFFLOAD=1 \
I_MPI_OFFLOAD_CELL_LIST=0-11 I_MPI_HYDRA_BOOTSTRAP=ssh \
mpiexec -n 2 -ppn 2 build/benchmarks/gbench/mp/mp-bench --vector-size 1000000000 \
--reps 50 --v=3 --benchmark_filter=GemvEq_DR/ --sycl
will not fail on kdse-pre-dnp-02
There was a problem hiding this comment.
I corrected sizes of the matrix, so that data size is actually similar (not exactly the same, but very close).
The line with 100000 and 50000 was commented out, and added comment that that matrix size is suitable for weak testing with dr-bench.
The command should also be working on kdse-pre-dnp-02
There was a problem hiding this comment.
What's important, it may sometimes freeze - it is due to gtest deciding on some hosts to stop iterating state, and not on others. I am not sure why it decides that, but in that case the test needs to be re-run
No description provided.