another issue with spmvm and views

Issue #314 resolved
Jonas Thies
created an issue

See also #310, #312, the following test currently fails on our test system:

mpirun -np 2 ./phist-1.3.0-kernels-test-Release --gtest_filter=DSparseMatTest_sprandn_25_4.sparseMat_times_mvec_random_with_input_view_0_0

only occurs with CUDA and with -np 2 (hybrid CPU/GPU, that is)

Comments (5)

  1. Jonas Thies reporter

    OK, I figured out it has nothing to do with CUDA but with the SELL chunk height C, I use C=32 if there is a GPU process. I can also get the problem when configuring phist with

    -DPHIST_SELL_C=32

    and running mpirun -np 2 env GHOST_TYPE=work ./phist-1.3.0-kernels-test-Debug

  2. Jonas Thies reporter

    it also segfaults for one MPI process now:

    [ RUN ] DSparseMatTestWithUnalignedViews_sprandn_25_4.sparseMat_times_mvec_random_with_input_view_1_1 [GHOST] PE0 INFO at ghost_sell_spmv_selector() <sell_spmv.cpp:581>: Try chunkheight=32, blocksz=1, impl=AVX2, aligned [GHOST] PE0 PERFWARNING at ghost_sell_spmv_selector() <sell_spmv.cpp:606>: Using potentially non-optimal kernel: C=32 blocksz=1 align=1 impl=AVX Segmentation fault (core dumped)

  3. Log in to comment