autogenerated CPU kernels never get used with Intel compiler?

Issue #322 new
Jonas Thies created an issue

With the Intel 17 compiler:

At least when called from phist, I always get ghost messages that a slow fallback kernel is used, despite the fact that I compiled the block size into ghost (tested with spmv and tsmttsm). He seems to always think the data is unaligned, which it shouldn't be in this case.

From the phist build dir I run

./Dbench_mvecT_times_mvec 1000000 8 8 100 ./Dbench_sparseMat_times_mvec BENCH3D-128-A0 8

(for instance)

with gcc 7.2 it works fine

I also noticed that the output line on how to re-cmake ghost with the missing kernels says GHOST_SPMMV= instead of GHOST_AUTOGEN...

Comments (2)

  1. Dominik Ernst

    I had redone the code that generated the missing kernels notice configure string, because the old code produced garbled strings for many different used kernel configurations. I forgot the AUTOGEN in the configuration option.

    About the non-usage of specialized kernels, I'll have to investigate. I remember that I made the alignment detection more conservative in some scenario to fix an alignment related segfault. I figured it would'nt matter much, as unaligned loads should not have a performance penalty any more, but not every kernel has unaligned versions for all configurations and would therefore need to fall back even further.

  2. Log in to comment