Grainsize calculation on TinyThreads

[franzbischoff]

RcppParallel/inst/include/RcppParallel/TinyThread.h

Line 68 in 39303ec

std::vector<IndexRange> splitInputRange(const IndexRange& range,

I was working fine with TBB in my app on Windows, defining the grainsize, but I got problems when porting to Linux and using TinyThreads. So I looked at the code.

Here I can see that, if I have a vector of size 767, and I set the grainsize to 64 I get the following:

1. Total threads is 8
2. So, 767/8 is not exact, and the algorithm chooses to do 767/(8-1) ~= 109
3. So each chunk will be of size 109. That is fine for my algorithm that needs a chunk larger than, let's say, 60.
4. There will be 767 / 109 = 7.0366 jobs. That means one job will be of size 4 that is below my grainsize of 64. And this will break my algorithm.

So, what about the last chunk to be of size 113 instead of 109 + 4? This will be in conformity with what is expected from grainsize.

I can write a pull request. For now, I'm just discussing the case.

[franzbischoff]

Fix:

  // allocate ranges
  std::vector<IndexRange> ranges;
  std::size_t begin = range.begin();
  std::size_t end = begin;
  while (begin < range.end()) {
    if ((range.end() - (begin + grainSize)) < grainSize)
      end = range.end();
    else
      end = std::min(begin + grainSize, range.end());

    ranges.push_back(IndexRange(begin, end));
    begin = end;
  }

[kevinushey]

Thanks for the bug report + investigation. I agree with your assessment; would you be willing to write a pull request?

[franzbischoff]

Sure, asap

[kevinushey]

Looks like this was fixed in #146 -- thanks again for the PR!