Skip to main content

EES Standard (Energy-Efficient Software)

Skigen is built under the EES standard — a set of engineering constraints that ensure every cycle spent is a cycle of useful computation.

Principles​

1. Zero Interpreter Tax​

Machine learning code written in interpreted languages pays a constant overhead for object dispatch, garbage collection, and the GIL. Skigen eliminates this by compiling directly to vectorized machine code.

2. Template-Based Scalar Flexibility​

All estimators are templatized on Scalar (default: double). Switching to float doubles SIMD throughput on the same hardware:

Skigen::StandardScaler<float> scaler;   // 2× SIMD density
Skigen::StandardScaler<double> scaler;  // Full precision

3. Zero-Copy Inputs​

All read-only inputs accept Eigen::Ref<const MatrixType>, which supports:

  • Direct matrix references (no copy)
  • Sub-block views
  • Memory-mapped data

4. Expression Templates Over Raw Loops​

Eigen's expression templates fuse operations and vectorize automatically. Skigen never uses manual element-wise loops — all computation flows through Eigen's optimized kernel pipeline.

5. In-Place Operations​

Every transformer provides _inplace variants that modify data directly, eliminating temporary allocations:

scaler.transform_inplace(X);           // No allocation
Eigen::MatrixXd Z = scaler.transform(X); // Allocates result

6. Static Polymorphism​

CRTP replaces virtual dispatch. The compiler resolves all calls at compile time — no vtable, no indirect jumps, no branch mispredictions from polymorphism.

Measurable Goals​

  • Throughput: Operations should achieve ≥ 80% of Eigen's raw kernel throughput.
  • Memory: transform_inplace should use zero additional heap memory beyond the input.
  • Float/Double ratio: float throughput should be ≈ 2× double throughput on AVX2 hardware.