`TBBMultiThreader::SetNumberOfWorkUnits` ignores GlobalMaximumNumberOfThreads, OK?

[N-Dekker]

The SetNumberOfWorkUnits member function of MultiThreaderBase does at

ITK/Modules/Core/Common/src/itkMultiThreaderBase.cxx

Lines 259 to 262 in c14ff0e

	MultiThreaderBase::SetNumberOfWorkUnits(ThreadIdType numberOfWorkUnits)
	{
	m_NumberOfWorkUnits = std::clamp<ThreadIdType>(numberOfWorkUnits, 1, m_PimplGlobals->m_GlobalMaximumNumberOfThreads);
	}

m_NumberOfWorkUnits = std::clamp<ThreadIdType>(numberOfWorkUnits, 1, m_PimplGlobals->m_GlobalMaximumNumberOfThreads);

However, SetNumberOfWorkUnits member function of TBBMultiThreader ignores GlobalMaximumNumberOfThreads, as it just does, at

ITK/Modules/Core/Common/src/itkTBBMultiThreader.cxx

Lines 105 to 108 in c14ff0e

	TBBMultiThreader::SetNumberOfWorkUnits(ThreadIdType numberOfWorkUnits)
	{
	m_NumberOfWorkUnits = std::max(1u, numberOfWorkUnits);
	}

  m_NumberOfWorkUnits = std::max(1u, numberOfWorkUnits);

Is that behavior difference intended? If so, why would TBBMultiThreader behave different than the other threaders, with respect to SetNumberOfWorkUnits?

[dzenanz]

Each thread can process multiple work units. The logic behind it is that sometimes different work units (typically blocks of the image) require different amount of work to complete. So having N:1 ratio of work units to actual threads allows load balancing.

[N-Dekker]

Thanks @dzenanz

But then, why does itk::MultiThreaderBase do m_NumberOfWorkUnits = std::clamp(numberOfWorkUnits, 1, m_GlobalMaximumNumberOfThreads)? More specifically, it appears that for itk::PoolMultiThreader the number of work units is limited to ITK_MAX_THREADS, by using a fixed array, at

ITK/Modules/Core/Common/include/itkPoolMultiThreader.h

Line 125 in 4260088

InternalWorkUnitInfo m_ThreadInfoArray[ITK_MAX_THREADS]{};

    InternalWorkUnitInfo m_ThreadInfoArray[ITK_MAX_THREADS]{};

Shouldn't that be a resizable container (like std::vector<InternalWorkUnitInfo>) instead?

[dzenanz]

I don't remember any longer, but maybe the image region splitters etc never exceed that number of work units? Dynamic multi threading can exceed that number, I believe.

[blowekamp]

EDIT: Never mind, I need to look at these classes closer...