Module ml::tensor_blas

BLAS-accelerated Tensor matmul.

tensor.sc's tensor_matmul is a portable, dependency-free triple loop -- deliberately so (see tensor.h's own header comment: no framework dependency baked into the portable core). Real BLAS libraries (Accelerate on macOS, OpenBLAS/MKL elsewhere) are cache-blocked, hand-vectorized, and multithreaded in ways a compiler-auto-vectorized triple loop doesn't match — see the Benchmarks page. This file closes that gap by calling the same underlying library PyTorch does (cblas_sgemm, float32 like Tensor itself), instead of re-implementing a hand-tuned GEMM from scratch.

The products go through std/simd/gemm.h's sgemm: any standard CBLAS the program links (Accelerate on macOS, AOCL-BLAS, oneMKL or OpenBLAS elsewhere; '-framework Accelerate' / '-lblis' / '-lopenblas' ...) takes the large ones, and without one sgemm's own vectorized, threaded kernel runs -- so linking a BLAS is optional, not required. The one Apple-only piece, vDSP_vadd in the backward pass, has a portable fallback elsewhere (see tensor_blas.sc). There is no runtime "unavailable" case the way GPU backends have one.