[
https://issues.apache.org/jira/browse/MXNET-323?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16463208#comment-16463208
]
Anirudh Subramanian edited comment on MXNET-323 at 5/3/18 11:56 PM:
--------------------------------------------------------------------
Since the bottleneck seems to be the binary_broadcast_kernel, I tried to fuse
unravel and dot inside the kernel. This doesn't seem to help much.
Results after the run:
{quote}('PASS: {}...', 0)
func1: 0.764457941055
func2: 2.51862192154
func3: 0.547819852829
('PASS: {}...', 1)
func1: 0.487478017807
func2: 1.92227315903
func3: 0.447080135345
('PASS: {}...', 2)
func1: 0.491612911224
func2: 1.92092394829
func3: 0.447576999664
('PASS: {}...', 3)
func1: 0.490720033646
func2: 1.92054891586
func3: 0.446537017822
{quote}
Building without the debug flag (-O3 flag) gives a better speedup for func2
which is the mx.sym.broadcast_add
{quote}('PASS: {}...', 0)
func1: 0.143939971924
func2: 0.392518043518
func3: 0.147929906845
('PASS: {}...', 1)
func1: 0.129691123962
func2: 0.389477968216
func3: 0.145953178406
('PASS: {}...', 2)
func1: 0.128004074097
func2: 0.388610124588
func3: 0.144158840179
('PASS: {}...', 3)
func1: 0.130708932877
func2: 0.292946100235
func3: 0.10039305687
{quote}
was (Author: anirudh2290):
Since the bottleneck seems to be the binary_broadcast_kernel, I tried to fuse
unravel and dot inside the kernel. This doesn't seem to help much.
Results after the run:
{quote}('PASS: {}...', 0)
func1: 0.764457941055
func2: 2.51862192154
func3: 0.547819852829
('PASS: {}...', 1)
func1: 0.487478017807
func2: 1.92227315903
func3: 0.447080135345
('PASS: {}...', 2)
func1: 0.491612911224
func2: 1.92092394829
func3: 0.447576999664
('PASS: {}...', 3)
func1: 0.490720033646
func2: 1.92054891586
func3: 0.446537017822
{quote}
Building without the debug flag (-O3 flag) gives a better speedup for func2.
{quote}('PASS: {}...', 0)
func1: 0.143939971924
func2: 0.392518043518
func3: 0.147929906845
('PASS: {}...', 1)
func1: 0.129691123962
func2: 0.389477968216
func3: 0.145953178406
('PASS: {}...', 2)
func1: 0.128004074097
func2: 0.388610124588
func3: 0.144158840179
('PASS: {}...', 3)
func1: 0.130708932877
func2: 0.292946100235
func3: 0.10039305687
{quote}
> Broadcasting ops are slow
> -------------------------
>
> Key: MXNET-323
> URL: https://issues.apache.org/jira/browse/MXNET-323
> Project: Apache MXNet
> Issue Type: Improvement
> Reporter: Lai Wei
> Assignee: Anirudh Subramanian
> Priority: Major
>
> https://github.com/apache/incubator-mxnet/issues/8219
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@mxnet.apache.org
For additional commands, e-mail: issues-h...@mxnet.apache.org
