who use cublasCtrsv i have some problems with result vector

I’m use this function in my programm two times (for L and then for U matrix).
When i fill real and image part of cuComplex vector by random counts, on exit i have wrong vector.
If i generate cuComplex vector with zero image part, the result vector will true (compare with matlab).

Who used this function? Please help with my problem.
I will thankful for help.

Which version of Cublas do you use and on which system?

Can you specify exactly the parameters (n, N,T/C , diagonal) in order to reproduce it

I use cuda 3.2 on gtx 280, Win XP

cublasCtrsv (‘L’, ‘N’, ‘U’, N, A, N, X, 1)

A - cuComplex matrix (square)
X - cuComplex vector

What dimension N and lda do you use?

How do you know that your results are wrong?

I use different dimentions starting with two (dim == lda)
All measures on GPU i compare with matlab.
Only the first value of result vector was true in all measures (becouse L-matrix have unit diag)
Other values differed from 0.5 to 5.0 regard to real value.

I guess that Matlab is going to use double precision. And comparing with Ctrsv in single precision might lead to discrepancy. Singfle precision on solver can quickly accumulate accuracy errors.

Could you try with Ztrsv (double complex)?

Also could you try using Ctrsm and Ztrsm with m=1 which is equivalent to Ctrsv/Ztrsv