https://github.com/xrq-phys/blis_apple

BLIS fork with kernels for Apple M1. (Perhaps) The first open-source BLAS with Apple Matrix Coprocessor support.
https://github.com/xrq-phys/blis_apple

Last synced: 3 days ago
JSON representation

BLIS fork with kernels for Apple M1. (Perhaps) The first open-source BLAS with Apple Matrix Coprocessor support.

• Host: GitHub
• URL: https://github.com/xrq-phys/blis_apple
• Owner: xrq-phys
• License: other
• Created: 2021-07-17T08:13:18.000Z (over 3 years ago)
• Default Branch: amx-dev
• Last Pushed: 2023-01-07T14:09:32.000Z (over 2 years ago)
• Last Synced: 2025-03-26T03:51:15.405Z (21 days ago)
• Language: C
• Homepage:
• Size: 44.2 MB
• Stars: 33
• Watchers: 4
• Forks: 5
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • Changelog: CHANGELOG
  • Contributing: CONTRIBUTING.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-gemm - blis_apple: BLIS optimized for Apple M1 - 3-Clause) (Libraries 🗂️ / CPU Libraries 💻)

README

        
Current MD file explains this port of BLIS to Apple's matrix coprocessor.

- Here is the original BLIS [README](README_BLIS.md).

As of Jul., 2021, the coprocessor is undocumented but not protected either. (Any user / program is allowed to invoke this coprocessor and it's supposed to be safe.) This work is based on Dougall Johnson's effort on analyzing the related instructions.

Known issues:

- Generic-strided is not supported by our microkernels for the destination matrix. Program would `assert(false)` upon encountering such a situation.

- TRSM might fail. Try commenting out function call to `bli_cntx_set_packm_kers` in `config/aaplmx/bli_cntx_init_aaplmx.c` if your need TRSM to work.

Performance:

  ![](docs/graphs/aaplmx/output_st_dgemm_asm_blis.png)

  ![](docs/graphs/aaplmx/output_st_sgemm_asm_blis.png)