numc

README

A fast, lightweight N-dimensional tensor library in C, built for high-performance scientific computing and AI workloads. Hand-written SIMD kernels, packed GEMM, arena memory, and OpenMP parallelization.

Warning, this is not production ready yet

#include <numc/numc.h>

int main(void) {
  NumcCtx *ctx = numc_ctx_create();

  size_t shape[] = {2, 3};
  NumcArray *a = numc_array_fill(ctx, shape, 2, NUMC_DTYPE_FLOAT32, &(float){1.0f});
  NumcArray *b = numc_array_fill(ctx, shape, 2, NUMC_DTYPE_FLOAT32, &(float){2.0f});
  NumcArray *c = numc_array_create(ctx, shape, 2, NUMC_DTYPE_FLOAT32);

  numc_add(a, b, c);       // c = a + b
  numc_array_print(c);     // [[3, 3, 3], [3, 3, 3]]

  numc_ctx_free(ctx);      // frees a, b, c in O(1)
}

Why numc

Fast — hand-written AVX2, AVX-512, NEON, SVE/SVE2, and RVV kernels. Packed GEMM with Goto's algorithm.
Complete — 10 dtypes (int8–int64, uint8–uint64, float32, float64), elementwise ops, reductions, matmul, broadcasting, slicing, random. Every operation supports every dtype.
Simple — arena allocator owns all memory. Create a context, use it, free it. No manual malloc/free per array.
Portable — pure C23 with scalar fallback. Builds on any platform. SIMD is a bonus, not a requirement.

Performance

Run ./run.sh bench to benchmark on your machine. Results are written to bench/numc/results.csv. See bench/README.md for details.

Build

git clone https://github.com/rizukirr/numc.git
cd numc
./run.sh release        # optimized build
./run.sh test           # tests with AddressSanitizer
./run.sh bench          # run benchmarks

Cross-compilation (requires cross-compilers + QEMU):

./run.sh neon test      # AArch64 NEON
./run.sh sve test       # AArch64 SVE
./run.sh rvv test       # RISC-V RVV
./run.sh avx512 test    # x86 AVX-512

CMake options: NUMC_ENABLE_ASAN (AddressSanitizer), NUMC_WERROR (-Werror), NUMC_OPTIMIZE_NATIVE (-march=native, default ON).

Examples

Examples and demo programs (including MNIST) live in a separate repository to keep this repo lightweight: numc-example

Documentation

Project Wiki — API reference, architecture details.

Hardware testing wanted

SIMD kernels pass correctness tests via QEMU but only AVX2 has native benchmarks. If you have access to AVX-512, NEON (Apple Silicon / Graviton), SVE, or RVV hardware, run ./run.sh bench and open a PR.

Contributing

See CONTRIBUTING.md.

License

Licensed under the MIT License.