tensor-compiler/taco

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

1366 stars · 200 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TACO (The Tensor Algebra Compiler) is a C++ library that computes sparse and dense tensor algebra and linear algebra expressions, generating kernels with performance competitive with hand-optimized code. It can be used as a C++ or Python library, a command-line code generator, and supports OpenMP and CUDA execution.

Use cases

• compute sparse matrix-vector multiplication (SpMV) efficiently
• compute MTTKRP for sparse tensor factorization
• compute SDDMM for machine learning factor analysis
• generate optimized C kernels for tensor expressions
• load and compute on sparse tensors in CSR, COO, or CSF formats
• run tensor algebra on GPUs with CUDA
• prototype sparse linear algebra in Python with NumPy/SciPy

When to choose

• you need high-performance sparse tensor or linear algebra kernels without hand-optimizing them
• you work with mixed sparse/dense tensor computations in C++ or Python
• you want to generate custom optimized kernels for specific tensor expressions
• you need CSR, COO, CSF, or custom sparse storage format support

When to avoid

• you only need dense linear algebra, where BLAS libraries like Eigen or MKL are simpler
• you need deep learning autodiff or neural network training rather than tensor algebra kernels
• you need a pure Python solution without a C++ build toolchain

Facets

library · maturity maintenance

compiler math machine-learning data-science machine-learning performance cpp python cli sparse-tensors tensor-algebra linear-algebra code-generator cuda openmp spmv mttkrp sddmm numerical-computing code-generation algorithms scientific-computing linux macos gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tensor-compiler/taco")