https://github.com/openmachine-ai/huggingfive

HuggingFive 🖐️ is a collection of ML functions and libraries written in RISC-V assembly and C.
https://github.com/openmachine-ai/huggingfive

ai assembly machine-learning ml risc-v riscv-assembly riscv32

Last synced: 2 months ago
JSON representation

HuggingFive 🖐️ is a collection of ML functions and libraries written in RISC-V assembly and C.

• Host: GitHub
• URL: https://github.com/openmachine-ai/huggingfive
• Owner: OpenMachine-ai
• License: mit
• Created: 2023-07-22T09:22:24.000Z (over 2 years ago)
• Default Branch: main
• Last Pushed: 2023-07-29T20:00:05.000Z (over 2 years ago)
• Last Synced: 2025-04-13T23:37:00.135Z (7 months ago)
• Topics: ai, assembly, machine-learning, ml, risc-v, riscv-assembly, riscv32
• Homepage:
• Size: 76.2 KB
• Stars: 8
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# HuggingFive :raised_hand_with_fingers_splayed:

HuggingFive :raised_hand_with_fingers_splayed: is a collection of ML functions and libraries written in RISC-V assembly and C.  This includes neural network layers, activation functions, as well as entire neural networks. Think of it as a low-level HuggingFace for RISC-V assembly code.  The table below includes performance numbers for benchmarking.  The hope is to eventually roll all these handwritten tricks into the existing compiler toolchains (or have AI generate even better assembly code). 

  

    Description

    Author

    RV config

    Data types

    Performance numbers for an exemplary config

    Notes

   

    Config

    MACs

    Ops

    Register utilization

    Memory size (B)   

   

    

    Conv2D 1x1

    OpenMachine

    RV32IF

    FP32

    C=32, F=32, R=6

    C F R2 = 36,864

    57,953

    8/31 x-regs, 21/32 f-regs

    

    

   

    

    Conv2D 3x3

    OpenMachine

    RV32IF

    FP32

    C=3, F=8, R=12, stride=1

    9 C F R2 = 31,104

    TBD

    TBD

    

    

   

    

    Depthwise Conv2D 3x3

    OpenMachine

    RV32IF

    FP32

    C=4, R=6, stride=1

     9 C R2 = 1,296

    TBD

    TBD

    

    

   

    

    

    

    

    

    

    

    

    

    

    

  

- C : input channels

- F : output channels (or filters), only used if F is not the same as C

- R : square root of input resolution (e.g. R=6 for image resolution of 6x6 pixels)

- Q : square root of output resolution, only used if Q is not the same as R

- MACs : number of fused multiply-accumulate operations required by the neural-network layer (can be used as a lower-bound for total number of ops; this number ignores possible savings from zero-padding for conv-layers)  

## Contribute

Please add your functions and routines to HuggingFive :raised_hand_with_fingers_splayed:: Add a link to your code in the 

table and submit a PR, which will get approved promptly because there are no rules here.

More details coming soon ...