- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

hello,

I am trying the MKL Library, in particular I am finding some problems with the multiplication.

This is a simple example in C# where a multiply two squared Complex matrix, the result is very strange because the first row have very high values and the second row zero values.

Some one can help me to understand the issue?

public static int LAPACK_ROW_MAJOR = 101; public static int LAPACK_COL_MAJOR = 102; //TRANSPOSE public static int CBLAS_NO_TRANS = 111; public static int CBLAS_TRANS = 112; public static int CBLAS_CONJ_TRANS = 113; /** CBLAS native cblas_cgemm declaration */ [DllImport("mkl_rt.dll", CallingConvention = CallingConvention.Cdecl, ExactSpelling = true, SetLastError = false)] internal static extern void cblas_cgemm( int matrix_layout, int TransA, int TransB, int M, int N, int K, Complex alpha, [In] Complex[] A, int lda, [In] Complex[] B, int ldb, Complex beta, [In, Out] Complex[] C, int ldc ); public static void MulTest() { Complex[] a = { 1, 2, 3, 4 }; Complex[] b = { 5, 6, 7, 8 }; int m = 2; int n = 2; int k = 2; Complex[] c = new Complex[m * n]; Complex alpha = 1; Complex beta = 0; int lda = k; int ldb = n; int ldc = n; MKLWrapper.cblas_cgemm(LAPACK_ROW_MAJOR, CBLAS_NO_TRANS, CBLAS_NO_TRANS, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); }

Thank You very much

Gianluca

Link Copied

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

I'm unfamiliar with the C# implementation of the Complex data type being used here, but it looks like the intent is to create 2x2 Complex matrices A and B, yet only 4 values are used to initialize each. It should require a real and imaginary component for each element of each matrix. In this case are the imaginary parts of each entry of A and B initialized to 0? If this is the case then sgemm should suffice.

After calling cgemm I'm not sure how you're reading off the elements of C, or what is meant by "very high values" but the "very high values" might correspond to the real components of each entry in C (ie, calling sgemm on real matrices A,B) and the "zeros" might correspond to the imaginary components of each entry of C, which would indeed be zero everywhere if A and B were initialized with imaginary part equal to 0.

Could you please include your entire source? It would helpful if I could see specifically how C is being read after cgemm is called, and implementation of the Complex type.

Anthony

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

The matrices are both 2x2 . I also tried with more complex matrix, with real and imaginary values different than zero, but I had the same strange results. Then I have created a simple example to submit to you.

As you understand the inizialization code set only the real part while the imaginary is set to zero.

The results values in this case correspond only to the real part. The result, with x64 dlls in this example is:

-148096, 0

-219636, 0

0, 0

0, 0

While with x86 every value is zero.

I am working with MKL 11.3.1 version and the dlls that I found this in "C:\Program Files (x86)\IntelSWTools\compilers_and_libraries_2016.0.110\windows\redist\" folder, after that "Intel Paralled Studio" was installed.

The example I commited is the complete version, as you can see is extremely simple. The result is in c vector.

Gianluca

Gianluca

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

Thank you, in this case the System.Numerics implementation of Complex takes the form Complex(Double, Double), whereas cgemm takes as input single precision complex types. I was able to reproduce your results and found that replacing the call from cgemm to a call to zgemm gives the expected results. Hope this helps.

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

Thank you very much

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

I'm sorry to say that it works partially, because with x64 platform is ok, but with x86 returns all zero values. Any idea?

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

Hi Gianluca G,

I happened to review several C# issue, so saw your questions here.

As the cblas_zgemm and cblas_cgemm ask address of alpha and beta, not value.

void cblas_zgemm (const CBLAS_LAYOUT Layout, const CBLAS_TRANSPOSE transa, const

CBLAS_TRANSPOSE transb, const MKL_INT m, const MKL_INT n, const MKL_INT k,** const void *alpha**, const void *a, const MKL_INT lda, const void *b, const MKL_INT ldb,** const void *beta,** void *c, const MKL_INT ldc);

So you may change

[DllImport("mkl_rt.dll", CallingConvention = CallingConvention.Cdecl, ExactSpelling = true, SetLastError = false)]

internal static extern void cblas_zgemm(

int matrix_layout,

int TransA,

int TransB,

int M,

int N,

int K,

** ref Complex alpha,**

[In] Complex[] A,

int lda,

[In] Complex[] B,

int ldb,

**ref Complex beta,**

[In, Out] Complex[] C,

int ldc

);

MKLWrapper.cblas_zgemm( LAPACK_ROW_MAJOR, CBLAS_NO_TRANS, CBLAS_NO_TRANS, m, n, k,** ref alpha**, a, lda, b, ldb, **ref bet**a, c, ldc)

and try it out.

Regards,

Ying

- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content

Hi Gianluca,

It should be able to solve the problem of x86 returns all zero values.

Best Regards,

Ying

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page