What it does

Tensor Comprehensions is a C++ library with Python bindings that JIT-compiles high-performance ML kernels from a concise mathematical notation. You write something resembling Einstein summation—O(n, c1, c3, h, w) +=! I0(...) * I1(...)—and TC handles the rest: Halide for scheduling, ISL for polyhedral optimization, NVRTC or LLVM for code generation. It plugs into PyTorch and Caffe2, so the resulting kernel drops into familiar tensor code.

The interesting bit

The autotuner uses a genetic search that evolves mapping options across generations. The README shows a tensordot kernel going from 7,345μs (naive) to 220μs in two generations on a P100. Crucially, you autotune once per operation definition, then reuse those options across different tensor sizes—amortizing the ~28-second search cost.

Key highlights

• Mathematical syntax close to index notation; no explicit loops or thread blocks
• Genetic autotuner searches the schedule space automatically
• Targets CUDA via NVRTC or CPU via LLVM; framework-agnostic at core
• Claims 80%+ of peak shared memory bandwidth after autotuning on some kernels
• Apache 2.0, with conda binaries and Docker builds available

Caveats

• Register-level optimizations are “still in the work”; peak compute is not guaranteed
• The project appears dormant—last significant activity years ago, and it targets pre-Ampere GPU architectures
• Caffe2 integration is a historical curiosity at this point

Verdict

Worth studying if you build ML compilers or need to understand Halide-style scheduling. For production PyTorch work, look at torch.compile, Triton, or CUDA graphs instead—this is more research artifact than living toolchain.