topic two comments on this topic: in Intel® Integrated Performance Primitives

Element-wise Matrix Multiplication

Bryan — Thu, 21 Nov 2013 15:43:29 GMT

I'm new to IPP and am trying to find the function (if it exists) that supports element-wise matrix multiplication. I've found some matrix operations, but none of them seem to support element-wise operations. Can someone please point me in the right direction?

Thanks!

Do you mean:

SergeyKostrov — Fri, 22 Nov 2013 13:43:00 GMT

Do you mean [ 2 2 2 ] [ 4 4 4 ] [ 8 8 8 ] [ 2 2 2 ] x [ 4 4 4 ] => [ 8 8 8 ] [ 2 2 2 ] [ 4 4 4 ] [ 8 8 8 ] or you're trying to find out if IPP supports a classic transposed based matrix multiplication?

Here is a comment to IDZ

SergeyKostrov — Fri, 22 Nov 2013 13:50:07 GMT

Here is a comment to IDZ developers: Why does the editor remove spaces in posts? It is simply impossible in some cases to format examples.

Good question Sergey, I

Bryan — Fri, 22 Nov 2013 17:57:23 GMT

Good question Sergey, I should have been more explicit.

Using Matlab notation:
[a1,a2 ; a3,a4] .* [b1,b2 ; b3,b4] = [a1*b1,a2*b2 ; a3*b3,a4*b4]

Thanks!

Please take a look at a set

SergeyKostrov — Fri, 22 Nov 2013 21:43:15 GMT

Please take a look at a set of DSP functions from ipps.h header file: ... IPPAPI( IppStatus, ippsMul_32fc_A11, (const Ipp32fc a[],const Ipp32fc b[],Ipp32fc r[],Ipp32s n)) IPPAPI( IppStatus, ippsMul_32fc_A21, (const Ipp32fc a[],const Ipp32fc b[],Ipp32fc r[],Ipp32s n)) IPPAPI( IppStatus, ippsMul_32fc_A24, (const Ipp32fc a[],const Ipp32fc b[],Ipp32fc r[],Ipp32s n)) IPPAPI( IppStatus, ippsMul_64fc_A26, (const Ipp64fc a[],const Ipp64fc b[],Ipp64fc r[],Ipp32s n)) IPPAPI( IppStatus, ippsMul_64fc_A50, (const Ipp64fc a[],const Ipp64fc b[],Ipp64fc r[],Ipp32s n)) IPPAPI( IppStatus, ippsMul_64fc_A53, (const Ipp64fc a[],const Ipp64fc b[],Ipp64fc r[],Ipp32s n)) ...

In overall, a set of IPP

SergeyKostrov — Fri, 22 Nov 2013 21:49:27 GMT

In overall, a set of IPP functions ippsMul_xxfc_Axx could be used. This is because both matrices with the same sizes are represented in memory as "vectors" if they are located in contiguous memory blocks. Note 1: In case of array-of-arrays representations of matrices these IPP functions can't be used. However, multiplication could be done as row by row. Note 2: Take a look at IPP Vector Math library. Note 3: You could also look at MKL library and let me know if you're interested in that and I'll point you in a right direction.

two comments on this topic:

Gennady_F_Intel — Sun, 24 Nov 2013 15:55:47 GMT

two comments on this topic:

1) all -- ipp mxm operations are well optimized for very small matrixes ( <= 6x6)

2) in the case if need to compute larger problems, I would recommend you to look at MKL's implementaion of m x m operations.