DGEMM inconsistent result

[Viet-Duc]

• Library version: 2021.3
• OS info and version: CentOS 7.9.2009
• Compiler verison: icc version 2021.3.0 (gcc version 4.8.5 compatibility)
• MKL linking command:

icc -O2 -std=c11 -mkl=sequential -L$MKLROOT/lib/intel64 dgemm.c -o dgemm.x

• Reference: https://software.intel.com/content/www/us/en/develop/documentation/mkl-tutorial-c/top/multiplying-matrices-using-dgemm.html

Dear Support Team,

I am measuring performance of matrix multiplication using dgemm().

The code to reproduce the issue is attached. dgemm() was invoked as following:

dgemm("N", "N", &m, &n, &p, &alpha, A, &p, B, &n, &beta, C, &n);

The example is a simple 3x3 multiplication. In the source code, there are two ways to initialize A and B. I marked these two methods with approriate comments in the file.

• Double casting of interger: this gives correct C
• Random number between (0,1]: this gives incorrect C

To rule out the possibility of thread interference, I linked with the sequential version of the library. 

The file 'correct.txt' is the expected output for randomized input matrices.

The file 'incorrect.txt' is the result I obtained with dgemm() for randomized input matrices

I appreciate if you can point me to the origin of this inconsistency.

• dgemm.c: https://justpaste.it/7zhsb
• correct.txt: https://justpaste.it/6mizg
• incorrect.txt: https://justpaste.it/5g7bi

(For some reason, I could not attach the files using forum, I appologize for the inconvenience)

[VidyalathaB_Intel · Accepted solution]

Hi,

Thanks for reaching out to us.

We tried the provided code from our end and could reproduce the results which you are getting.

But here are some observations:

To check the results we used cblas_dgemm where you will get an option to change the matrix layout (Row Major or Col Major)

Here is the function call

cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans, m, n, p, alpha, A, p, B, n, beta, C, n);

By using the above function call, we observed it is giving the correct result (matches with your expected result)

cblas_dgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, m, n, p, alpha, A, p, B, n, beta, C, n);

whereas the above function call gives the output which matches with your incorrect result.

>>dgemm("N", "N", &m, &n, &p, &alpha, A, &p, B, &n, &beta, C, &n);

By default dgemm follows the matrix layout as column-major, hence there is a mismatch in your expected and actual results.

Please find the attached file for the code snippet we have tried.

Hope the provided information helped.

Regards,

Vidya.

[Viet-Duc]

Dear Vidya,

Thanks for the detailed answer and test. I am able to obtain correct result using cblas_dgemm as you suggested.

Regards.

[VidyalathaB_Intel]

Hi,

Thanks for accepting our solution.

As this issue is resolved, we are closing this thread. Please post a new question if you need any additional information from Intel as this thread will no longer be monitored.

Have a Good Day!

Regards,

Vidya.