Plan & Journal

[larryxiao]

SIMD Wrapper for ARM NEON, Intel AVX512 & KNC

• Pre GSoC ( - 26 April)
  • [X] Study the HPX, libflatarray codebase
• Community Bonding Period (27 April - 24 May)
  • [ ] Discuss with mentor about the architectures to support (ARM NEON, Intel AVX512, Intel IMCI, CUDA, or KNC), and regard each as sub-project.
  • [ ] Sub-project 1: Support for ARM NEON · Issue #2 · larryxiao/libflatarray
  • [ ] Sub-project 2: Support for Intel AVX512 · Issue #3 · larryxiao/libflatarray
  • [ ] Sub-project 3: Support for Intel IMCI · Issue #4 · larryxiao/libflatarray
• Week 1, 2 (25 May - 7 June)
  • [ ] Write detailed implementation plans about sub-projects, and refine with help from mentor.
  • [ ] Implement key functions in the project, get approval from mentor.
• Week 3, 4 (8 June - 21 June)
  • [ ] Implement sub-projects 1-3, 2 weeks per sub-project.
  • [ ] Submit midterm evaluation (26 June)
• Week 5, 6 (22 June - 5 July)
  • [ ] Testing for sub-projects 1-3, ensure correctness
• Week 7, 8 (6 July - 19 July)
  • [ ] Evaluate performance, then optimize
• Week 9, 10 (20 July - 2 August))
  • [ ] Implement more sub-projects, depends on plan.
• Week 11, 12 (3 August - 21 August)
  • [ ] Testing, improve documentation
  • [ ] Submit final evaluation (28 August)
• (Remain one week available for allocation)

  Plan

First implement a prototype for a new vectorization architecture, eg. float 4. Then test the prototype, make sure it's correct. Then expand into more data types, like float 8, float 16, and double.

[larryxiao]

Pre GSoC ( - 26 April)

• Study: HPX, libflatarray · Issue #1 · larryxiao/libflatarray
• Written concept code with ARM NEON

Community Bonding Period (27 April - 24 May)

• Write code for ARM NEON float_4, compile (cross compiled), not tested yet
• Setup toolchain for AVX 512, run sample code with Intel SDE

Week 1 May 25 - May 31

• [In progress] Study: related project, and proposals · Issue #5 · larryxiao/libflatarray

Week 2 Jun 1 - Jun 7

• [done] Study: related project, and proposals · Issue #5 · larryxiao/libflatarray

Week 3 Jun 8 - Jun 14

• [AVX512] compile avx512 code with Intel compiler

Week 4 Jun 15 - Jun 21

• set up testing environment
• IMCI prototype code

Week 5 Jun 22 - Jun 28

• merge with bitbucket upstream, convert git to mercurial
• compile on jetson: compile flag, include path
• implement new functions: load/store aligned, store_nt, gather scatter
• mid term evaluation

Week 6 Jun 29 - July 5

• test NEON on jetson

Week 7 July 6 - July 12

• week off (graduation)

Week 8 July 13 - July 19

• NEON: non-temporal store, aligned load/store, gather, scatter
• implementation and test

Week 9 July 20 - July 26

[larryxiao]

Question

• [x] In short_vec there's only support for floating point data types. How does libflatarray support integer?
• [ ] What does libflatarray/detail/sqrt_reference.hpp do?