[
https://issues.apache.org/jira/browse/SYSTEMML-2187?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16403907#comment-16403907
]
Matthias Boehm edited comment on SYSTEMML-2187 at 3/18/18 8:13 AM:
-------------------------------------------------------------------
After closer inspection, I was able to reproduce the reported issue of
incorrect results with Intel MKL for single-precision conv2d_bias_add (which
calls the MKL-specific {{dnnConvolutionCreateForwardBias_F32}} and
{{dnnExecute_F32}}). Interestingly, this issue does not show up with OpenBLAS
(single/double precision), or MKL double precision (which calls
{{dnnExecute_F64}}).
Most importantly, this issue only shows up on a Xeon Gold 6138 (Skylake with
AVX 512), while it works fine on older architectures without AVX 512 such as a
Xeon E5-2440.
was (Author: mboehm7):
After closer inspection, I was able to reproduce the reported issue of
incorrect results with Intel MKL for single-precision conv2d_bias_add (which
calls the MKL-specific {{dnnExecute_F32}}). Interestingly, this issue does not
show up with OpenBLAS (single/double precision), or MKL double precision (which
calls {{dnnExecute_F64}}).
Most importantly, this issue only shows up on a Xeon Gold 6138 (Skylake with
AVX 512), while it works fine on older architectures without AVX 512 such as a
Xeon E5-2440.
> Double check new single-precision native conv2d on different platforms
> -----------------------------------------------------------------------
>
> Key: SYSTEMML-2187
> URL: https://issues.apache.org/jira/browse/SYSTEMML-2187
> Project: SystemML
> Issue Type: Sub-task
> Reporter: Matthias Boehm
> Assignee: Matthias Boehm
> Priority: Major
> Fix For: SystemML 1.1
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
