different results even if seed is fixed

[bab2min]

Depending on the environment(32bit or 64bit / SSE2, AVX or AVX2) in which tomotopy are installed, different results will be produced with the same seed.
It is possibly related to #60.

[bab2min]

There are three possible causes.

1. Random Number Engine
   

   tomotopy/src/TopicModel/TopicModel.hpp

   Lines 13 to 28 in 400f060

   	#if _WIN32 || _WIN64
   	#if _WIN64
   	typedef std::mt19937_64 RandGen;
   	#else
   	typedef std::mt19937 RandGen;
   	#endif
   	#endif
   	#if __GNUC__
   	#if __x86_64__ || __ppc64__
   	typedef std::mt19937_64 RandGen;
   	#else
   	typedef std::mt19937 RandGen;
   	#endif
   	#endif

2. Prefix Sum
   

   tomotopy/src/Utils/sample.hpp

   Lines 123 to 140 in 400f060

   	inline void prefixSum(float* arr, size_t K)
   	{
   	size_t Kf = (K >> 2) << 2;
   	if (Kf) prefix_sum_SSE(arr, Kf);
   	else Kf = 1;
   	for (size_t i = Kf; i < K; ++i)
   	{
   	arr[i] += arr[i - 1];
   	}
   	}
   	#else
   	inline void prefixSum(float* arr, size_t K)
   	{
   	for (size_t i = 1; i < K; ++i)
   	{
   	arr[i] += arr[i - 1];
   	}
   	}

3. Eigen's redux function(sum())
   The summation order can be varied by SIMD options.

(1) and (2) will can be fixed soon, but (3) seems not easy to fix.

[mayankchatteron1]

@bab2min could you tell the timeline by which it will be finished?
even setting -
np.random.seed(123)
random.seed(123)

on same avx2 machine giving different results on a different run.

[bab2min]

Hi, @mayankchatteron1
Since version 0.8.2, you can obtain the same result in the same machine when the seed is fixed with only workers=1.
If the result is different even though the seed is fixed on the same machine, please check that you set workers=1 first.
If you get different results even if workers=1, this is maybe another bug, so please report your source code.
Thank you!

[mayankchatteron1]

@bab2min Thanks it working now. Thank you again for the quick resolution.
But curious to know why different workers giving different results, is it due to some kind of model parallelism?

[bab2min]

@mayankchatteron1 Yes.
In a multithreaded environment, it is impossible to completely control the execution order by software.
In order to control it, each inference step needs to be synchronized, which reduces parallelization efficiency and makes it virtually no difference from running on a single thread.

To avoid confusion, I'll update package to print a warning message when seed is fixed and workers > 1.