Module simd::gemm

#include <std/simd/gemm.h> source

single-precision matrix multiply on the CPU.

sgemm is a cache-blocked GEMM: A and B are packed into contiguous panels and multiplied by a 4 x 16 register-tile kernel on vec<float, 8>, so it uses whatever vector unit the target has (SSE2/AVX on x86-64, NEON on AArch64, WebAssembly SIMD128, or scalar code elsewhere). Large products are split into tiles of C that run on std/sync/parallel.h's thread pool; small ones run on the calling thread.

Matrices are row-major; 'ld' is the distance in elements between the starts of two rows (>= the row length).

Vendor BLAS: on hosted targets (not Windows yet), a program that links a CBLAS gets its kernels for sgemm's larger products: AOCL-BLAS on AMD CPUs, oneMKL on Intel, Accelerate on Apple. Except with Accelerate, std runs the library single-threaded and splits the product across its own thread pool (the libraries' OpenMP workers busy-wait between calls and would compete with the pool for cores); that setting applies to the program's own calls into the library too. safeguard's "blas" feature links the right one for the build machine; otherwise link one by hand (e.g. -lblis-mt, -lmkl_rt, -framework Accelerate). The choice is made at link time; with no BLAS linked the built-in kernel runs.

Functions

sgemm	C = alpha * op(A) * op(B) + beta * C, where op(A) is m x k and op(B) is k x n, and op(X) = X, or X transposed when its trans flag is SGEMM_TRANS (then X is stored k x m / n x k). Matrices are row-major with leading dimensions lda / ldb / ldc (the stride between rows). beta == 0 overwrites C without reading it, so C may start uninitialized. Returns 1, or 0 (nothing computed) when a slice is too short for its dimensions or a leading dimension is smaller than its row.
sgemm_serial	sgemm, but always on the calling thread (no pool, no wake-ups; a vendor library, if linked, runs single-threaded): for code that is already parallel, or that must not block on other threads.
sgemm_vendor	1 if sgemm is using a linked vendor BLAS, and which one: "AOCL-BLAS (BLIS)", "oneMKL", "Accelerate", "OpenBLAS", "CBLAS" (another library) or "built-in".
sgemm_backend

Macros

SGEMM_NO_TRANS
SGEMM_TRANS