Need help on cgesvd

[hmli]

testing_cgesvd completes successfully for some sizes but fails for some other.

Code: Select all

/magma-1.2.1/testing> ./testing_cgesvd -M 256 -N 1800
device 0: GeForce GT 640, 901.5 MHz clock, 2047.7 MB memory, capability 3.0
  testing_cgesvd -M 256 -N 1800

  N     CPU Time(s)    GPU Time(s)     ||R||_F / ||A||_F
==========================================================
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
CUBLAS error: memory mapping error (11) in magma_cgetmatrix_async at cset_get.cpp:145
CUBLAS error: memory mapping error (11) in magma_csetmatrix at cset_get.cpp:98
 1800       4.38           4.32         4.211870e-01


I tested it on GT640, C1060 and GTX670 and replicated the same error. I traced the error to cgebrd() function, which by itself passed tests. The magma module was built on mkl. I rebulit magma with gcc instead of icc ( the mkl lib was still built by icc though) but the error was still there. Any suggestions?

[mgates3]

This is an error that we recently found. For [zc]gesvd, if you allocate the workspace of size
(M+N)*nb + 2*min(M,N) + 2*min(M,N)**2
it should work. There's a more involved fix to allow using the smaller workspace of size
(M+N)*nb + 2*min(M,N)
with reduced performance.

Note: for [ds]gesvd, in both cases the workspace requires 3*min(M,N) instead of 2*min(M,N), since it doesn't take rwork separately.

-mark

[hmli]

Yes, That is the problem; now it works.
Thank you, mark!