Version 0.0.3

GradieX Documentation

Single-precision arithmetic, OpenMP simd pragmas on the hot loops, mini-batch gradient accumulation, and a derivative lookup that ties each activation and loss to its own derivative.

#Build Status

The 18 warnings are 17 instances of 'restrict' type qualifier on return type has no effect plus one unused parameter. They are benign but load-bearing in an odd way — see The __restrict__ return qualifier. tests/run_tests.sh already passes -Wno-ignored-qualifiers.

#Getting Started

Header-only, one translation unit, as in v0.0.2:

clang++ -std=c++20 -O2 main.cpp -o main
./main

To get anything out of the #pragma omp simd directives, OpenMP must be enabled:

clang++ -std=c++20 -O2 -fopenmp-simd main.cpp -o main   # Apple clang
g++-14  -std=c++20 -O3 -fopenmp      main.cpp -o main   # GCC
clang++ -std=c++20 -O3 -fopenmp      main.cpp -o main   # LLVM clang + libomp

#What Changed Since v0.0.2

The full v0.0.2 reference is at v0.0.2 → C++ Library, and the three-version comparison is on the versions page.

#Plots

New in v0.0.3: the Python package ships a matplotlib layer. net.train(..., plot=True) draws the loss curve when training finishes, and seven further diagrams cover the architecture, the weights, the fit and the loss surface. Every one renders in light and dark.

#Reference

The C++ Library

The header-only engine in core/. The float signatures, the SIMD pragmas, gradient accumulation and batching, the derivative lookup, and what is still open.

The Python Library

The gradiex package: install, the API, what the binding guarantees the C++ API does not, and the plot gallery — loss curves, architecture diagrams, decision boundaries, weight maps and 3‑D loss landscapes.

Performance

Benchmarks against v0.0.2 on a shared harness: 1.2–3.2× per sample, 1.7–5.5× with mini-batching, and what the -fopenmp-simd flag is worth at each layer width.

If you are looking for something specific:

• What switching to float costs
• Whether the SIMD pragmas do anything
• Gradient accumulation and batching
• How derivatives are resolved
• Why two translation units will not link
• Known bugs and limitations
• The plot gallery

#At a Glance

#Project State

v0.0.3 keeps v0.0.2's flat memory layout and changes what runs over it. Three things drive the release: halving the width of every scalar so twice as much fits in cache and a vector register holds eight lanes instead of four; annotating the hot loops so a compiler will vectorise them; and restoring mini-batching, which the v0.0.2 rewrite dropped and v0.0.1 had.

Two v0.0.2 defects are fixed along the way. computeLoss now calls the layer's configured loss instead of always summing squared error, and the scratch pads are sized width rather than inputSize * width. A third is addressed in design: fetchAssociatedDerivative derives the derivative from the forward function, so a caller can no longer pair sigmoid with reluDerivative. That lookup works for five of the eight activations — see Derivative Lookup.

What remains is reach rather than correctness. The engine builds and trains; python/ and tests/ were never touched, so the release has no automated verification at all — the gradient check, the zero-allocation assertion and the binding tests are all sitting in double-typed files that no longer compile. The new batching path has never been asserted against anything. Porting the suite is the highest-value work left; see Test Suite.