[
https://issues.apache.org/jira/browse/IMPALA-12449?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sebastian Pop updated IMPALA-12449:
-----------------------------------
Description:
The port to arm64 uses https://github.com/DLTcollab/sse2neon an automatic tool
to convert x86_64 intrinsics to arm64 NEON.
As some of the x86_64 vector instructions do not have a one-to-one translation
to NEON, the sse2neon translation tool falls back to a sequential emulation of
the missing instructions.
To avoid those slow paths, we need to adapt to NEON instructions all the
code-paths with x86_64 intrinsics.
was:
TestWebPage.test_webserver_hide_logs_link recently fails due to IMPALA-12053
makes the "/events" link visible in catalogd WebUI.
{noformat}
custom_cluster/test_web_pages.py:242: in test_webserver_hide_logs_link
assert found_links == expected_catalog_links, msg
E AssertionError: bad links from webui port 25020
E assert ['/', '/catal... '/memz', ...] == ['/', '/catalo...metrics', ...]
E At index 2 diff: u'/events' != '/jmx'
E Left contains more items, first extra item: '/varz'
E Full diff:
E - [u'/',
E ? -
E + ['/',
E - u'/catalog',
E ? -
E + '/catalog',
E - u'/events',
E - u'/jmx',
E ? -
E + '/jmx',
E - u'/log_level',
E ? -
E + '/log_level',
E - u'/memz',
E ? -
E + '/memz',
E - u'/metrics',
E ? -
E + '/metrics',
E - u'/operations',
E ? -
E + '/operations',
E - u'/profile_docs',
E ? -
E + '/profile_docs',
E - u'/rpcz',
E ? -
E + '/rpcz',
E - u'/threadz',
E ? -
E + '/threadz',
E - u'/varz']
E ? -
E + '/varz'] {noformat}
> SSE2NEON is slow compared to native NEON code
> ---------------------------------------------
>
> Key: IMPALA-12449
> URL: https://issues.apache.org/jira/browse/IMPALA-12449
> Project: IMPALA
> Issue Type: Bug
> Affects Versions: Impala 4.3.0
> Reporter: Sebastian Pop
> Assignee: Quanlong Huang
> Priority: Critical
> Fix For: Impala 4.3.0
>
>
> The port to arm64 uses https://github.com/DLTcollab/sse2neon an automatic
> tool to convert x86_64 intrinsics to arm64 NEON.
> As some of the x86_64 vector instructions do not have a one-to-one
> translation to NEON, the sse2neon translation tool falls back to a sequential
> emulation of the missing instructions.
> To avoid those slow paths, we need to adapt to NEON instructions all the
> code-paths with x86_64 intrinsics.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
