Implement Block SpGEMM

[mzuzek]

TODO:

• [ ] Tests:
  • [x] simple explicit matrices with known multiplication result;
  • [x] implement block version of random matrix generator;
  • [ ] address regression test for #402: implement or drop.
• [ ] Interface: can we conveniently avoid separate bspgemm methods with proper templating over spgemm ?
• [ ] Address review feedback and tune performance;

Implementation

• [x] Serial/Debug algorightm implementation (SPGEMM_DEBUG);
• [ ] Default algorithms Serial/Debug algorightm implementation (SPGEMM_KK_MEMORY and SPGEMM_KK_MEMSPEED);
  • [x] adjust hash accumulator for blocks;
  • [x] cover CPU variants: MultiCoreTag and MultiCoreTag4 (SPGEMM_KK_LP);
  • [ ] adjust memory estimation and algorithm/execution policy selection for block version;
  • [x] CPU (reduced to suggested_vector_size, suggested_team_size, and team_row_chunk_size);
  • [ ] GPU - detailed review of all estimations and allocations (+refactoring if possible).
  • [ ] GPU implementation;
  • [ ] Can we get unified accumulator (like HashmapAccumulator for non-block version);
• [ ] Connect TPLs for performance reference:
  • [x] MKL for CPU benchmarking;
  • [ ] CuSPARSE for GPU benchmarking;
• [ ] Speed numeric method: SPGEMM_KK_SPEED;