> On 06/08/2012 07:24 AM, Hauke Heibel wrote:
>>
>> Hi Gael,
>>
>> I think the most efficient solution would be to use atomics. I am not
>> sure whether the code below is allowed and fixes the problem
>>
>> #pragma omp atomic
>> static std::ptrdiff_t m_l1CacheSize =
>> manage_caching_sizes_helper(queryL1CacheSize(),8 * 1024);
>>
>> #pragma omp atomic
>> static std::ptrdiff_t m_l2CacheSize =
>> manage_caching_sizes_helper(queryTopLevelCacheSize(),1*1024*1024);
>>
>> Another alternative might be #pragma omp critical.
>>
>> Yet another idea might be to look at 'atomic_do_once' (see [1] and
>> [2]) from Intel's TBB. The code from [2] probably gets much simpler
>> for the use case at hand because the cache size itself could be stored
>> as an atomic.
>>
>> Regards,
>> Hauke
>>
>> [1]
>>
http://stackoverflow.com/questions/9344739/thread-safe-lazy-creation-with-tbb
>> [2]
>>
https://akazarov.web.cern.ch/akazarov/cmt/releases/nightly/tbb/src/tbb/tbb_misc.h
>>
>> On Fri, Jun 8, 2012 at 3:12 PM, Gael Guennebaud
>> <
gael.guennebaud@xxxxxxxxx> wrote:
>>>
>>> Hi,
>>>
>>> I'm looking for some help on understanding how the following piece of
>>> code in Eigen can lead to a race condition when the function
>>> manage_caching_sizes is called from different threads:
>>>
>>>
>>> inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a,
>>> std::ptrdiff_t b)
>>> {
>>> return a<=0 ? b : a;
>>> }
>>>
>>> inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1=0,
>>> std::ptrdiff_t* l2=0)
>>> {
>>> static std::ptrdiff_t m_l1CacheSize =
>>> manage_caching_sizes_helper(queryL1CacheSize(),8 * 1024);
>>> static std::ptrdiff_t m_l2CacheSize =
>>> manage_caching_sizes_helper(queryTopLevelCacheSize(),1*1024*1024);
>>> ...
>>> }
>>>
>>>
>>> During it's first call, this function computes and store the cache sizes.
>>>
>>> I would like to find a solution that avoids making the static variable
>>> "thread private" as suggested there:
>>>
>>>
http://stackoverflow.com/questions/8828466/using-openmp-and-eigen-causes-infinite-loop-deadlock/10540025
>>> The problem with this approach, is that the cache-sizes are recomputed
>>> for every thread.
>>>
>>>
>>> thanks,
>>>
>>> Gaël
>>>
>>>
>>
>>
>
>
>