Introduce Pickle::{Read,Write}PODArray to serialize arrays of POD

Introduce Pickle::{Read,Write}PODArray to serialize arrays of POD

This lets us gain a little bit more.
DelegatedFrame_ManyQuads_100000_100000 goes from 13261 to 12019

BUG=307480

[piman, 2013-10-24 06:26:21]

danakj: please review
jam: OWNERS

This is on top of https://codereview.chromium.org/34413002/

This gives us another ~10%. Is it worth the slight complexity? The callsites
look nicer though, I think.

[Tom Hudson, 2013-10-24 09:55:03]

Usual driveby questions:

> DelegatedFrame_ManyQuads_100000_100000 goes from 13261 to 12019

On which hardware?
How typical are those input values?
How sensitive are those results to the input values?
How stable are those results?

(We've looked at cc optimizations which seemed to give nice results, but not on
the hardware we cared about or not for input values that we actually encountered
in the real world or not repeatably.)

Quibbling about results aside, the callsites do look nice. I can't believe I'm
endorsing templates.

[danakj, 2013-10-24 14:40:08]

On Thu, Oct 24, 2013 at 5:55 AM, <tomhudson@chromium.org> wrote:

> Usual driveby questions:
>
>
>  DelegatedFrame_ManyQuads_**100000_100000 goes from 13261 to 12019
>>
>
> On which hardware?
>

z620


> How typical are those input values?
>

Spaceport is 4000_4000. Using larger numbers gets us more consistent test
results.


> How sensitive are those results to the input values?
>

Sensitive to the input? I'm not sure what this means. The time taken to
serialize is dependent on the number of quads/shared quad states being
serialized.


> How stable are those results?
>

We're using 100k instead of 4k because that gives more stable results. We
get quite consistent numbers between runs. You can try out the
content_perftests yourself if you'd like (would love it if so).


>
> (We've looked at cc optimizations which seemed to give nice results, but
> not on
> the hardware we cared about or not for input values that we actually
> encountered
> in the real world or not repeatably.)
>

This is largely to address crazy scenarios like the spaceport benchmark,
not sure where this will happen outside of js benchmarks.


>
> Quibbling about results aside, the callsites do look nice. I can't believe
> I'm
> endorsing templates.
>
>
https://codereview.chromium.**org/39463002/<https://codereview.chromium.org/3...
>

To unsubscribe from this group and stop receiving emails from it, send an email
to chromium-reviews+unsubscribe@chromium.org.

[jam, 2013-10-24 16:18:38]

i'm not a base owner