Memory error in zgetrf_gpu

[jeremiahpalmer]

Hello!

I was running magma_zgetrf_gpu (from magma_zgesv_gpu) and found a memory error. I am calling the routine magma_zgetrf_gpu with n=m=1024.

In line 107, nb is set to 32. In line 108, s is set to 32. Then, in line 219, we have

Code: Select all

       
lapackf77_zgetrf( &rows, &nb0, work, &lddwork, ipiv+s*nb, &iinfo);


The magma_int_t array "ipiv" is supposed to be length n. The above code segment will reference ipiv out of bounds.

-Jeremiah

P.S. This is in magma_1.1. So, if it is fixed in magma_1.2, I apologize!

[Stan Tomov]

I am trying to figure out if indeed there is a problem at this place because when min(m, n) is divisible by nb, as in this case, nb0 is actually zero, so the routine in question

Code: Select all

lapackf77_zgetrf( &rows, &nb0, work, &lddwork, ipiv+s*nb, &iinfo);


must exit with success without modifying any data. We could have executed this clean up code conditionally (when nb0 > 0) but we haven't seen a problems so far. It is possible though that an LU factorization implementation is not checking end exiting if n or m is zero. Did you indeed see a problem at this place and with what LAPACK implementation?
Thanks,
Stan

[jeremiahpalmer]

I ran magma_zgesv_gpu with the size 1024. The execution exited quite ungracefully. I traced the memory error using valgrind and found that it was failing at that particular line in magma_zgetrf_gpu. I printed the integer data s and nb just before lapackf77_zgetrf was called. It printed that both s and nb were 32.
-Jeremiah

[jeremiahpalmer]

Any news on this error?

Thanks,
Jeremiah