Dear,

I would like to put a buffer `shared(char[])` into a shared memory ( L2 cache ) and provide a start and end index to multiple thread in order to perform computation on a piece of original buffer.

I read doc from std.concurrency, std.parallelism, core.thead (druntime) and vectorization (simd) .  But I found any example or unittest which perform such calculus.

1/ So did you have an example somewhere ?
2/ about shared keyword do we ensure that its size is lower than L2 or L3 cache using :

```
core.sys.posix.unistd: sysconf, _SC_LEVEL2_CACHE_SIZE, _SC_LEVEL3_CACHE_SIZE;
immutable size_t L2_CACHE_SIZE = cast(size_t) sysconf(_SC_LEVEL2_CACHE_SIZE);
immutable size_t L3_CACHE_SIZE = cast(size_t) sysconf(_SC_LEVEL3_CACHE_SIZE);

```
3/ variable flagged as `shared` does at mean the variable is put into L2 cache ?

thanks for your help

best regards

On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
> Dear,
>
> I would like to put a buffer `shared(char[])` into a shared memory ( L2 cache ) and provide a start and end index to multiple thread in order to perform computation on a piece of original buffer.
>
> [...]

One image reveals more than a thousand words, see my picture :-)
   -> http://pix.toile-libre.org/?img=1572906675.png

On Monday, 4 November 2019 at 21:53:45 UTC, bioinfornatics wrote:
> On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
>> Dear,
>>
>> I would like to put a buffer `shared(char[])` into a shared memory ( L2 cache ) and provide a start and end index to multiple thread in order to perform computation on a piece of original buffer.
>>
>> [...]
>
> One image reveals more than a thousand words, see my picture :-)
>    -> http://pix.toile-libre.org/?img=1572906675.png

oups got a problem with the alpha layer:
   -> http://pix.toile-libre.org/?img=1572908302.png

On Monday, 4 November 2019 at 22:19:29 UTC, bioinfornatics wrote:
> On Monday, 4 November 2019 at 21:53:45 UTC, bioinfornatics wrote:
>> On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
>>> Dear,
>>>
>>> I would like to put a buffer `shared(char[])` into a shared memory ( L2 cache ) and provide a start and end index to multiple thread in order to perform computation on a piece of original buffer.
>>>
>>> [...]
>>
>> One image reveals more than a thousand words, see my picture :-)
>>    -> http://pix.toile-libre.org/?img=1572906675.png
>
> oups got a problem with the alpha layer:
>    -> http://pix.toile-libre.org/?img=1572908302.png

here my (not working) snippet code:
   -> https://run.dlang.io/gist/75429dcb8612e00635f7d53aabc4776f

the example do not meet exactly the spec and he do not compile.

On Tuesday, 5 November 2019 at 00:04:05 UTC, bioinfornatics wrote:
> On Monday, 4 November 2019 at 22:19:29 UTC, bioinfornatics wrote:
>> On Monday, 4 November 2019 at 21:53:45 UTC, bioinfornatics wrote:
>>> On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
>>>> [...]
>>>
>>> One image reveals more than a thousand words, see my picture :-)
>>>    -> http://pix.toile-libre.org/?img=1572906675.png
>>
>> oups got a problem with the alpha layer:
>>    -> http://pix.toile-libre.org/?img=1572908302.png
>
> here my (not working) snippet code:
>    -> https://run.dlang.io/gist/75429dcb8612e00635f7d53aabc4776f
>
> the example do not meet exactly the spec and he do not compile.

arf the generated link raise a 404, here a gist https://gist.github.com/run-dlang/4ffcc0641acef5eedf52a17c4995599b

On Tuesday, 5 November 2019 at 00:07:40 UTC, bioinfornatics wrote:
> On Tuesday, 5 November 2019 at 00:04:05 UTC, bioinfornatics wrote:
>> On Monday, 4 November 2019 at 22:19:29 UTC, bioinfornatics wrote:
>>> On Monday, 4 November 2019 at 21:53:45 UTC, bioinfornatics wrote:
>>>> On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
>>>>> [...]
>>>>
>>>> One image reveals more than a thousand words, see my picture :-)
>>>>    -> http://pix.toile-libre.org/?img=1572906675.png
>>>
>>> oups got a problem with the alpha layer:
>>>    -> http://pix.toile-libre.org/?img=1572908302.png
>>
>> here my (not working) snippet code:
>>    -> https://run.dlang.io/gist/75429dcb8612e00635f7d53aabc4776f
>>
>> the example do not meet exactly the spec and he do not compile.
>
> arf the generated link raise a 404, here a gist

I got some help from `feep` on IRC and he show me that the task pool can not to be const as reduce is not.
So now I have this script:
----> https://gist.github.com/run-dlang/46c753f7233a624bb5639db615c9634a
And this error:
----> /dlang/dmd/linux/bin64/../../src/phobos/std/parallelism.d(2730): Error: function std.parallelism.TaskPool.reduce!"a + b".reduce!(MapResult!(__lambda3, Result)).reduce cannot get frame pointer to onlineapp.countAllLines

I found this thread which talk about this error, but that did not help me to fix it: https://github.com/dlang/phobos/pull/4915

Thanks for your help




> https://gist.github.com/run-dlang/4ffcc0641acef5eedf52a17c4995599b

On Monday, 4 November 2019 at 19:53:29 UTC, bioinfornatics wrote:
> 3/ variable flagged as `shared` does at mean the variable is put into L2 cache ?

First caching is controlled by CPU not programmer.
Second not all architectures share L2 between cores.
4 jaguar cores share L2 cache in consoles
https://images.anandtech.com/doci/6976/Screen%20Shot%202013-05-23%20at%2012.22.12%20AM_678x452.png

2 bulldozer cores share L2 cache in FX processors
https://s3.amazonaws.com/hs-wordpress/wp-content/uploads/2017/12/13035240/bulldozer_021.jpg

each zen core has its own L2 cache in Ryzen cpu's
https://images.anandtech.com/doci/10591/HC28.AMD.Mike%20Clark.final-page-014.jpg


Third your link to run.dlang is 404

While you cant control cache contents explicitly you can control it by changing memory access patterns. When you fetch a peace of data CPU will load 64 byte size bucket/chunk that contains your data into L1 cache. Since L1 cache size is limited CPU had to remove existing data from L1 based on its eviction policy. Evicted data goes into L2 cache. That is the way to control what is where.

By default all variables(except immutable or static) are thread local and can only be accessed by one thread. Shared keyword makes data accessible trough multiple threads