Skip to content

README.md

Latest commit

 

History

History
42 lines (28 loc) · 2.37 KB

README.md

File metadata and controls

42 lines (28 loc) · 2.37 KB
Raw

A Machine Learning Engine/Library optimized for Zen 31 cache lines.
Full pipeline allowing for MLP and CNN neural networks.
All code can be compiled with just standard glibc
Counting *.h and *c files, there are 817 lines of code

AI was used for pseudo code generation, research, and light debugging2.
Research verified with sources such as academic papers and unclear topics were decided upon the most agreed method.

gcc (GCC) 16.1.1 20260430
cmake version 4.3.4

Features:

Stride-row padding to 64-bytes to fit within Zen 3

OpenMP for parallelization

Uses linear stride arithmatic over contiguous arrays to prevent multi-dimensional pointer-chasing

Inner matrix rows are forced to sit within CPU L1 cache

  • ikj instead of ijk: (tensor_matmul_2d)

    • ikj ensures that both C and B are sequentially stepping through rows and GCC can auto-vectorize and generate AVX vector instructions(Zen 3 is only AVX-2(256) but AVX-512(512) can be achieved Zen 4)

  • Out Of Place Transpose: (transpose_OOP)

    • An OOP transpose keeps all data sequential and padded as transposing in place across rows and columns would destroy padded cache boundaries and be complicated to implement

  • Shallow Memory Mapping: (flatten_view)

    • flatten_view points to existing pre-allocated memory buffers to prevent heavy memcpy or new matrix allocation

  • GEMM-Driven Spatial Convolutions: (im2col and col2im)

    • Instead of using deeply nested sliding window loops im2col and col2im transforms overlapping spatial receptive patches into linear columns to prevent non-contiguous memory reading patterns inside a GEMM matrix

  • Memory Resuse:

    • Every Tensor4D struct is instantiated with both its data and gradient pointer side-by-side so the backwards pass and easily write into them and prevents continuous scratchpad matix allocation and frequent garbage collection

Basic pipeline of neural network:
https://app.notion.com/p/Machine-Learning-Engine-389620b9962b80c486f0edee8021b5ad

unsigned_char_subtract within tensor_virus_hacks.c is a joke function and shouldn't be here but I can do whatever I want.

Footnotes

  1. Changing -march=znver3 to -march=znver4 in CMakeLists.txt works and GCC may vectorize SIMD differently ↩

  2. Light debugging as fixing loop order or accidentally writing over data instead of gradients ↩