CUDA: complex scalar * double sparse matrix * double vector

Question

I need a function that performs operation y = y + s*A*x on CUDA. Where y is complex vector ( cuDoubleComplex , for instance), x is double vector, A is double sparse matrix (csr format) and s is complex scalar.

Question : Is there any library to perform that operation?

I would like to skip transformation everything to complex, unless you convince me that it can be done efficiently.

PS

Unfortunately, it can't be done by cusparse function cusparseCsrmvEx()

Answer 1

Is there any library to perform that operation?

Almost certainly not. The type promotion you need to do pretty much rules that out. What you could to is something like:

1. cusparseDcsrmv to yield z0 = A*x
2. custom kernel or perhaps thrust operation to yield z1 = complex(z0)
3. cublas Zaxpy to calculate y = y + s * z1

Even better would be to write a custom kernel to fuse (2) and (3) together. Your choice.