Sadly, no performance improvements were to be found with avoiding the long[]
allocation (commit 9ea7eebf0164d1676f633b441bd952eaa20b0760), but maybe fixing
estimateRule() as well will show some small improvement.
https://github.com/apache/incubator-joshua/commit/9ea7eebf0164d1676f633b441bd952eaa20b0760
matt
> On Sep 13, 2016, at 11:54 PM, Matt Post <[email protected]> wrote:
>
> To be more specific, you should do this.
>
> cd $JOSHUA
> git pull
> rm -rf ext/kenlm/build
> ./jni/build_kenlm.sh
>
> The changes came about due to an insight and subsequent elbow grease from
> Kellen, and some technical debugging assistance from the great KenLM himself,
> who happened to be in the room here at MT Marathon. Here are some results on
> the timing tests I run every night on some moderate-sized models. The times
> here are total seconds decoding (not counting model loading). The zh-en test
> set has 1,357 sentences, while the es-en one has 3,000.
>
> model old runtime new runtime speedup sents/s
> ----- ----------- ----------- ------- -------
> es-en-phrase 422 276 1.52 7.1 -> 10.8
> es-en-phrase-packed 421 274 1.53 7.1 -> 10.8
> zh-en-hiero 258 193 1.33 5.2 -> 7.0
>
> The commits are ca0411acea805cc14086d91a43782f47d9ad4557 (old) and
> a81e51f8ef7ad3b97d891f466410f5cfe6af0d7b (new).
>
> This is just in time for the language packs we're building, and the 6.1
> release! And it's only Tuesday of the week. Nice work all around, folks.
>
> matt
>
> PS. (Kellen, be sure to collect your £ from Ken tomorrow)
>
>
>> On Sep 13, 2016, at 6:00 PM, kellen sunderland <[email protected]>
>> wrote:
>>
>> Hello everyone,
>>
>> Just wanted to give a heads up that as of this commit
>> https://github.com/apache/incubator-joshua/commit/90fff5ab1de3da23c0f64f90e69ce0da2392fd49
>> the abi for libkenlm.so has changed. That means you may have to recompile
>> it or it could crash when causing probRule.
>>
>> -Kellen
>
