Project

General

Profile

Feature #22638

Package a newer assembler suitable for gcc optimizations

Added by Christopher Backhouse 4 months ago. Updated 3 months ago.

Status:
Feedback
Priority:
Normal
Assignee:
-
Target version:
-
Start date:
05/28/2019
Due date:
% Done:

0%

Estimated time:
Duration:

Description

Not sure if this is the right place to file this?

On SL6 (novagpvm0?.fnal.gov), compiling with -march=native fails with errors like this

/tmp/ccTLDisd.s:43: Error: suffix or operands invalid for `vbroadcastsd'
/tmp/ccTLDisd.s:66: Error: no such instruction: `vpermpd $216,%ymm3,%ymm3'

Presumably gcc is relying on features that aren't implemented in the system version of as. It seems to me the gcc product should either package a suitable assembler, or depend on another product which does.

History

#1 Updated by Kyle Knoepfel 3 months ago

  • Status changed from New to Feedback
  • Tracker changed from Bug to Feature

We do not support code that can be built on arbitrary systems using -march=native. What applications/libraries are you trying to build, and where are you planning on deploying them?

#2 Updated by Christopher Backhouse 3 months ago

This is non-art code to run on FermiGrid or at NERSC. I realize that the binary won't necessarily be portable from the build machine (gpvm) to nodes, but my hot function will certainly benefit from vectorization, and it looks like the assembler won't support it however it's done.

#3 Updated by Kyle Knoepfel 3 months ago

Let's continue this discussion offline. We will setup a meeting.



Also available in: Atom PDF