#201 sMVN tuning dies

Developer
closed-fixed
5
2012-03-16
2012-03-14
No

In my 32-bit archdef install on etl-opt8, I got a seg fault in the command:
make res/sMVNK.sum pre=s

Am presently running to see if this is repeatable.

Since I see these seg faults sporadically from user accounts, I need to be sure and run this with valgrind to see if we can figure out what is going on.

Discussion

  • R. Clint Whaley

    R. Clint Whaley - 2012-03-14

    the the make command just runs:
    ./xmvnksearch -p s
    so should be easy to valgrind.

     
  • R. Clint Whaley

    R. Clint Whaley - 2012-03-14

    seg fault coming from timer:
    make smvnktime M=3000 N=2000 lda=3008 mvnrout="ATL_sgemvN_8x4_sse.c" align="-Fa 4 -Fa -8 -Fx 16 -Fx -32 -Fy 4 -Fy -8" tflags="-f res/smv2N_3000x2000_3008-16_8x4_0_a0x0x16_4107 -# 3 -F 0" mu=8 nu=4

     
  • R. Clint Whaley

    R. Clint Whaley - 2012-03-14

    This bug was caused by ambiguity of the alignment commands. For alignX2A, X and A modulo veclen must be identical, but timer was calling them saying thay had to be 4-bye aligned, but not allowing them to be 8 byte aligned. This means that A could be 4-byte aligned while A is 12-byte, for instance, causing a problem.

    I fixed this by having the timer always set single cases for 12-byte min align, and now allowing 16, which will force them the same, thus avoiding this problem.

    However, we will have the same ambiguity possible for AVX's 32-byte vectors and double, which I have not yet fixed. So, AVX may still have this problem, and I will need to think long and hard on how to fix this going forward.

     
  • R. Clint Whaley

    R. Clint Whaley - 2012-03-14

    With the proposed fix in, search now dies with:
    /home/whaley/TEST/ATLAS3.9.70.0/obj32/bin/ATLrun.sh /home/whaley/TEST/ATLAS3.9.70.0/obj32/tune/blas/gemv xsmvntime -m 3000 -n 2000 -l 3008 \ -C -1 -Fa 12 -Fa -16 -Fy 12 -Fy -16 -Fx 16 -Fx -32 -f res/smv2N_3000x2000_3008-16_8x4_800_a0x0x16_4107 -# 3 -F 0
    xsmvntime: /home/whaley/TEST/ATLAS3.9.70.0/obj32/..//tune/blas/gemv/mvktime.c:427: mvtime: Assertion `((size_t)a)%FAa == 0' failed.
    GEMV: M=3000, N=2000, lda=3008, AF=[12,16,12], AM=[16,32,16], beta=1.000000e+00, alpha=1.000000e+00:
    /home/whaley/TEST/ATLAS3.9.70.0/obj32/bin/ATLrun.sh: line 4: 12792 Aborted $atldir/$*
    make: *** [smvntime] Error 134

    Will need to debug.

     
  • R. Clint Whaley

    R. Clint Whaley - 2012-03-16
    • status: open --> closed-fixed
     
  • R. Clint Whaley

    R. Clint Whaley - 2012-03-16

    As far as I can tell, this bug is fixed in 3.9.70. I'll reopen if it reoccurs.

     

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:





No, thanks