Tensor Engine

A tensor engine built from scratch in Rust — no PyTorch, no ndarray, no dependencies beyond my own matmul kernel.

I built this to understand what actually sits underneath frameworks like PyTorch. Not to replace them — just to understand them. Turns out a tensor engine is a surprisingly small thing when you strip away the abstractions.

---

What it does

• Generic tensor operations — add, sub, mul, relu across f32/f64
• Matrix multiply backed by an AVX2 SIMD kernel with cache blocking
• Autograd — forward and backward pass with Value::matmul
• Linear regression as a working end-to-end example
• Criterion benchmarks with results committed

---

Structure

crates/
  core/     — Tensor type, layout, slice/get API
  ops/      — Element-wise ops, relu, matrix multiply
  autograd/ — Forward + backward pass, sum(), matmul backward
  macros/   — Derive macros

---

Run it

cargo test
cargo bench

---

What it covers

• Generic Rust with trait bounds — Copy + Add<Output = T>
• AVX2 intrinsics for matrix multiply performance
• Autograd from first principles — no magic, just the chain rule
• Cargo workspace with multiple crates
• CI with GitHub Actions