cpptensor

cpptensor is a C++ tensor library focused on CPU performance. It pairs a generic implementation with build-time AVX2/AVX-512 specializations, dispatches between them at runtime based on the host CPU, and can offload linear algebra to OpenBLAS. Early CUDA support is available behind a build flag.

What you get#

• N-dimensional Tensor with views and contiguous materialization.
• A broad op set — arithmetic, unary math, reductions, comparisons, plus reshape / transpose / concat / stack.
• Linear algebra — matmul, dot, tensordot, svd, and eig.
• Runtime ISA dispatch — generic, AVX2, and AVX-512 kernels selected after validating CPU feature bits and OS state.
• Lazy kernel registry — ops initialize the kernel registry on first use, so a fresh process can call A + B without manual setup.
• Numerically careful reductions — sum, mean, and dot use widened, compensated accumulation.
• Checkpoints — Tensor::save / Tensor::load with a versioned binary format.
• Tests & benchmarks — a Catch2 suite and Google Benchmark targets.

Platform support#

At a glance#

#include "cpptensor/tensor/tensor.hpp"
#include "cpptensor/ops/arithmetic/add.hpp"
 
int main() {
    using namespace cpptensor;
 
    Tensor a = Tensor::full({2, 2}, 1.0f);
    Tensor b = Tensor::full({2, 2}, 2.0f);
    Tensor c = a + b;
    c.print();
}

Where to next#

• New here? Start with Installation and the Quickstart.
• Learn the data model in Tensors & views and the Operations catalog.
• Tuning performance? See Backends & ISA and the Build options.