Matrix Vector multiplication

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Volodimir
Posts: 10
Joined: Fri Jun 29, 2018 2:52 pm

Matrix Vector multiplication

Post by Volodimir » Fri Jun 29, 2018 3:05 pm

Hello,

I am new to MAGMA, have a couple of questions regarding availability of functions:

1. in CUDA there is a

Code: Select all

cublas<t>dgmm()
function allowing to perform element-wise matrix-matrix multiplication
2. Likewise, there is a function

Code: Select all

 cublasGetMatrix
allowing to extract subarray from array.

Question - how I can implement these using MAGMA?

Thanks,

Volodimir

mgates3
Posts: 842
Joined: Fri Jan 06, 2012 2:13 pm

Re: Matrix Vector multiplication

Post by mgates3 » Fri Jun 29, 2018 3:29 pm

1) MAGMA doesn't have an element-wise multiply. However, you can still call the cublas function while using MAGMA.

2) MAGMA has:
magma_*setmatrix
magma_*getmatrix
magma_*setmatrix_async
magma_*getmatrix_async
where * is one of s, d, c, z, or i for single, double, single-complex, double-complex, or integer, respectively. See
http://icl.cs.utk.edu/projectsfiles/mag ... _comm.html
for all communication-related routines.

Unlike cublasGetMatrix, the MAGMA routines do not require passing sizeof the elements.

-mark

Volodimir
Posts: 10
Joined: Fri Jun 29, 2018 2:52 pm

Re: Matrix Vector multiplication

Post by Volodimir » Fri Jun 29, 2018 3:53 pm

Mark,

thanks for your feedback.

I am having difficulties in calling cublas functions while using MAGMA, as I am using complex arrays, and there is no straightforward conversion between MAGMA complex types magmaFloatComplex (or magmaDoubleComplex) and cuComplex (or cuDoubleComplex) in CUDA. Or did I miss something?

Thanks,
Volodimir

mgates3
Posts: 842
Joined: Fri Jan 06, 2012 2:13 pm

Re: Matrix Vector multiplication

Post by mgates3 » Fri Jun 29, 2018 4:19 pm

magmaFloatComplex is a typedef of cuFloatComplex. It shouldn't need a conversion. See include/magma_types.h:

Code: Select all

    typedef cuDoubleComplex magmaDoubleComplex;
    typedef cuFloatComplex  magmaFloatComplex;
Nonetheless, if you need a conversion, just do a cast using a pointer. E.g., to convert between std::complex and magmaComplex:

Code: Select all

  std::complex<float> alpha;
  std::complex<float> *A;
  magmaFloatComplex alpha2 = *((magmaFloatComplex*) &alpha);
  magmaFloatComplex *A2 = (magmaFloatComplex*) A;
-mark

Post Reply