- Marcar como novo
- Marcador
- Subscrever
- Silenciar
- Subscrever fonte RSS
- Destacar
- Imprimir
- Denunciar conteúdo inapropriado
Is anyone can solve the problem that mkl functions didn't get consistant result?
we use 10.3.3, It looks any function like DGEMM, DSYEV get different results
The different is big since for 1000*1000 matrix, the max difference for single element is
the matrix is 1.0e-14.
We also try 16 boundary alignment, but it doesn't work.
Link copiado
4 Respostas
- Marcar como novo
- Marcador
- Subscrever
- Silenciar
- Subscrever fonte RSS
- Destacar
- Imprimir
- Denunciar conteúdo inapropriado
We've generally seen that if you are running your program on the same computer, with the same number of threads and your memory is aligned on 16-byte boundaries then you'll get consistent results from DGEMM from one run to the next.
We'd need to know a little more about your program and the system on which you are running.
We'd need to know a little more about your program and the system on which you are running.
- Marcar como novo
- Marcador
- Subscrever
- Silenciar
- Subscrever fonte RSS
- Destacar
- Imprimir
- Denunciar conteúdo inapropriado
If I set each computer as mkl_sequential and use 16 boundary alignment.
but send whole computation to cluster, if the final result will be same?
- Marcar como novo
- Marcador
- Subscrever
- Silenciar
- Subscrever fonte RSS
- Destacar
- Imprimir
- Denunciar conteúdo inapropriado
Quoting Emily Zhu
If I set each computer as mkl_sequential and use 16 boundary alignment.
but send whole computation to cluster, if the final result will be same?
- Marcar como novo
- Marcador
- Subscrever
- Silenciar
- Subscrever fonte RSS
- Destacar
- Imprimir
- Denunciar conteúdo inapropriado
Another question is that alought DGEMM is repeatable, if the result will shift with real result of matrix
multiplication. And how much it will shift?

Responder
Opções do tópico
- Subscrever fonte RSS
- Marcar tópico como novo
- Marcar tópico como lido
- Flutuar este Tópico para o utilizador atual
- Marcador
- Subscrever
- Página amigável para impressora