Refactor the bitset data structure about weave

[mratsim]

Alternatively if we can replace hae steal request as pointers (#7), we could use a set instead of a bitset.

The following is probably the most efficient set datastructure found bounded sets that needs fast random picking and has been used for Montecarlo Tree Search: https://github.com/mratsim/golem-prime/blob/09a81ebf9f4c0789e830f391cdfde1e796ea0765/src/datatypes.nim#L106-L119 and https://github.com/mratsim/golem-prime/blob/09a81ebf9f4c0789e830f391cdfde1e796ea0765/src/core/c_empty_points.nim

• An array of indices
• An array of values
• A length

The array of values + len can be appended to and values can be deleted in any order (we replace the deleted value with the last value to keep it constant time we don't copy everything). The array of indices will map the "real" index to the actual index in the array of values.

This trades speed of space.

Bitset of 256 victims would require 32 bytes, such an array would require at the minimum 2x uint8 + length so 513 bytes. This requires steal request as pointers as otherwise this will have a huge impact on the size of the MPSC channels, the size required would grow by 481 bytes * numWorkers * numWorkers (as each channel needs to be sized for numWorkers steal requests, and each workers has its own)

[mratsim]

3 interesting approaches for a compressed bitset random victim selection:

• bithacks: http://graphics.stanford.edu/~seander/bithacks.html#SelectPosFromMSBRank
• faster uncompressing: https://lemire.me/blog/2018/02/21/iterating-over-set-bits-quickly/
• Adapt the following algorithm:
  • https://lemire.me/blog/2018/01/31/picking-distinct-numbers-at-random-benchmarking-a-brilliant-algorithm-javascript-edition/
  • https://blog.acolyer.org/2018/01/30/a-sample-of-brilliance/

Uncompressing will cause stack overflow with high core counts
The bithack require quite a lot of instructions even though it benefits from instruction level parallelism.

[mratsim]

closed by #15