MAINT: import time: avoid repeated textwrap function dispatch instantiation #14095

hmaarrfk · 2019-07-24T02:37:33Z

Avoid the use of textwrap to deindent static code. This shaves about 6-8ms ms of import time (at the expense of slightly less readable code depending on who you ask -- but definitely if you ask @eric-wieser)

After: 99.1±2ms
Before: 106±2ms (master)

hmaarrfk · 2019-07-24T02:38:24Z

numpy/ma/core.py

-        """)
+    long_std="""\
+masked_%(name)s(data =
+    %(data)s,


formatting definitely doesn't look right. give me a second.

eric-wieser · 2019-07-24T02:44:16Z

I'd be curious to know how much of this is the cost of importing textwrap, and how much is the cost of invoking dedent. I assume the stats you give above do not include the benefit of removing an import-time dependency on re?

hmaarrfk · 2019-07-24T02:56:22Z

The stats I gave were obtained from running

asv continuous -b time_numpy  master import_time_textwrap -E conda:3.7

with this branch including only this one commit. As such, the code is still improting re elsewhere.

A runtime dependency on re is very difficult to remove. Let me try to find a concrete example.

eric-wieser · 2019-07-24T02:57:25Z

Let me try to find a concrete example.

No need, just wanted to confirm that this timing is not caused by removing it.

In the interest of breaking down the import cost here, can you leave behind the import textwrap but not actually use it, and see how that import time compares?

hmaarrfk · 2019-07-24T02:58:28Z

yeah, i just looked at the textwrap source, it seems pretty trivial, and shouldn't be a contributor, let me test.

hmaarrfk · 2019-07-24T03:03:51Z

It seems like it is the fact that we are using dedent that is causing the slowdown

$ asv continuous -b time_numpy  master import_time_textwrap -E conda:3.7
· `wheel_cache_size` has been renamed to `build_cache_size`. Update your `asv.conf.json` accordingly.
· No executable found for python 3.6
· Creating environments
· Discovering benchmarks
·· Uninstalling from conda-py3.7-six
·· Installing 690132df <import_time_textwrap> into conda-py3.7-six.
· Running 6 total benchmarks (2 commits * 1 environments * 3 benchmarks)
[  0.00%] · For numpy commit ea965e4c <master> (round 1/2):
[  0.00%] ·· Building for conda-py3.7-six.......
[  0.00%] ·· Benchmarking conda-py3.7-six
[  8.33%] ··· Running (bench_import.Import.time_numpy--)...
[ 25.00%] · For numpy commit 690132df <import_time_textwrap> (round 1/2):
[ 25.00%] ·· Building for conda-py3.7-six.
[ 25.00%] ·· Benchmarking conda-py3.7-six
[ 33.33%] ··· Running (bench_import.Import.time_numpy--)...
[ 50.00%] · For numpy commit 690132df <import_time_textwrap> (round 2/2):
[ 50.00%] ·· Benchmarking conda-py3.7-six
[ 58.33%] ··· bench_import.Import.time_numpy                                                       101±1ms
[ 66.67%] ··· bench_import.Import.time_numpy_inspect                                               104±1ms
[ 75.00%] ··· bench_linalg.Lstsq.time_numpy_linalg_lstsq_a__b_float64                          2.69±0.08ms
[ 75.00%] · For numpy commit ea965e4c <master> (round 2/2):
[ 75.00%] ·· Building for conda-py3.7-six.
[ 75.00%] ·· Benchmarking conda-py3.7-six
[ 83.33%] ··· bench_import.Import.time_numpy                                                       107±2ms
[ 91.67%] ··· bench_import.Import.time_numpy_inspect                                               109±2ms
[100.00%] ··· bench_linalg.Lstsq.time_numpy_linalg_lstsq_a__b_float64                          2.64±0.04ms

BENCHMARKS NOT SIGNIFICANTLY CHANGED.
(numpy) mark@mark-xps ~/g/n/benchmarks import_time_textwrap|+1⚑ 1
$ git log -n 1
commit 690132dfad9ddc775fd4607581050af048c9f84f (HEAD -> import_time_textwrap, origin/import_time_textwrap)
Author: Mark Harfouche <[email protected]>
Date:   Tue Jul 23 22:59:41 2019 -0400

    Import textwrap again just to test

numpy/ma/core.py

eric-wieser · 2019-07-24T03:23:28Z

numpy/core/overrides.py

+    relevant_args = dispatcher(*args, **kwargs)
+    return implement_array_function(
+        implementation, {name}, relevant_args, args, kwargs)
+""".format(name=implementation.__name__)


I suspect this is where all the cost is - we end up dedenting the same string for every single numpy function, which is obviously expensive. Dedenting it just once wouldn't be nearly as bad, although at that point chances are the string definition lives outside the function, at which point there's nothing to even dedent.

In the interest of preserving the visual indent, I might be inclined to move this string definition to a global _wrapped_func_source variable in the file, which still avoids paying for textwrap

In the interest of preserving the visual indent, I might be inclined to move this string definition to a global _wrapped_func_source variable in the file, which still avoids paying for textwrap

Is this an opinion, or a requirement to get your approval on this PR?

But I do suspect you are right in the source of the problem. I just don't personally see any issue with deindenting strings manually.

Is this an opinion

It's an opinion that probably gates my immediate approval until another maintainer weighs ins.

I'm over it, lets hunt more milliseconds, and work to deprecate the whole testing module. Much more interesting ;)

I have a problem with dedenting strings manually, they are ugly, and ugly is unpleasant to maintain.

hmaarrfk · 2019-07-24T03:25:00Z

@eric-wieser, you know, my original attempt at ripping out re used a ton of lru_cache. I felt like I would never code like that unless I was really trying to shave off milliseconds, and even then, it would end up being in vain since somebody else would end up importing re and the lru_cache would likely make things slower than the original case.

I think in this case, manual deindentenation is an appropriate compromise.

This reverts commit fa60c30.

hmaarrfk · 2019-07-25T03:33:57Z

@eric-wieser. My milliseconds need this more than my intention to discuss coding style.

Benchmarks show similar improvements, again, hard to say when you are using the computer. Seeing as textwrap is built on top of re, it it likely only adds 1 ms instead of the 6 ms it added before.

Thanks for hunting down the true source of the issue. I won't be affraid to use textwrap in my other projects now.

eric-wieser

I could have been persuaded to take the original with some more thought, but this shortcuts that step. Might be worth a comment explaining it's location.

Can you confirm this still gives the savings you needed?

hmaarrfk · 2019-07-25T03:38:34Z

Yeah, it is still about a 4-6 ms in saving. Pretty good.

seberg · 2019-07-26T00:43:51Z

I agree, simply moving out the string formatting seems fine for that import time boost, thanks @hmaarrfk. I will assume that even if we do things such as vendoring dedent this will still make sense. So putting it in.

seberg · 2019-07-26T00:47:18Z

numpy/core/overrides.py

@@ -109,6 +109,18 @@ def decorator(func):
    return decorator


+


Opst, empty line too much (don't worry about it though), just noticed that I did not nitpick ;).

hmaarrfk commented Jul 24, 2019

View reviewed changes

MNT: import time: avoid textwrap in core library

fa60c30

hmaarrfk force-pushed the import_time_textwrap branch from bf780b8 to fa60c30 Compare July 24, 2019 02:40

hmaarrfk mentioned this pull request Jul 24, 2019

WIP, MAINT: Improve import time #14083

Closed

eric-wieser reviewed Jul 24, 2019

View reviewed changes

numpy/ma/core.py Outdated Show resolved Hide resolved

eric-wieser reviewed Jul 24, 2019

View reviewed changes

Revert "MNT: import time: avoid textwrap in core library"

875c044

This reverts commit fa60c30.

hmaarrfk force-pushed the import_time_textwrap branch from e688181 to 71f8cbe Compare July 25, 2019 03:31

hmaarrfk changed the title ~~MNT: import time: avoid textwrap in core library~~ MNT: import time: avoid repeated textwrap function dispatch instatiation Jul 25, 2019

eric-wieser approved these changes Jul 25, 2019

View reviewed changes

Define a variable for the _wrapped_func_source

3f06982

hmaarrfk force-pushed the import_time_textwrap branch from 71f8cbe to 3f06982 Compare July 25, 2019 03:38

seberg changed the title ~~MNT: import time: avoid repeated textwrap function dispatch instatiation~~ MAINT: import time: avoid repeated textwrap function dispatch instantiation Jul 26, 2019

seberg merged commit 702c357 into numpy:master Jul 26, 2019

seberg reviewed Jul 26, 2019

View reviewed changes

Uh oh!

MAINT: import time: avoid repeated textwrap function dispatch instantiation #14095

MAINT: import time: avoid repeated textwrap function dispatch instantiation #14095

Uh oh!

Conversation

hmaarrfk commented Jul 24, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

eric-wieser commented Jul 24, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hmaarrfk commented Jul 24, 2019

Uh oh!

eric-wieser commented Jul 24, 2019

Uh oh!

hmaarrfk commented Jul 24, 2019

Uh oh!

hmaarrfk commented Jul 24, 2019

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

hmaarrfk commented Jul 24, 2019

Uh oh!

hmaarrfk commented Jul 25, 2019

Uh oh!

eric-wieser left a comment

Choose a reason for hiding this comment

Uh oh!

hmaarrfk commented Jul 25, 2019

Uh oh!

seberg commented Jul 26, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

eric-wieser commented Jul 24, 2019 •

edited

Loading