binary_ops.hpp

// toolboxes/builtin/src/lang/operators/backends/binary_ops_loops.hpp
//
// Private declarations for the backend-split binary-op inner loops.
// Implemented by exactly one of backends/MStdBinaryOps_{portable,simd}.cpp
// depending on NK_WITH_SIMD. Callers (binary_ops.cpp) use these
// only for the 2D same-shape DOUBLE fast path — every other shape
// (broadcasting, 3D, type mismatch) goes through the general
// elementwiseDouble() helper in helpers.hpp unchanged.
 
#pragma once
 
#include <cstddef>
 
namespace nk::ops::detail {
 
void plusLoop   (const double *a, const double *b, double *out, std::size_t n);
void minusLoop  (const double *a, const double *b, double *out, std::size_t n);
void timesLoop  (const double *a, const double *b, double *out, std::size_t n);
void rdivideLoop(const double *a, const double *b, double *out, std::size_t n);
 
// Shared scalar bodies for the small-N gate ↔ portable fallback. The *Loop
// entry points above forward here (portable) or call it from their small-N gate
// (Highway), so the gate and the fallback are one definition and the loop lives
// in a lambda-free TU (binary_scalar.cpp) that auto-vectorizes regardless of the
// dispatcher's parallel_for capture. See bugs/ops/cheap-elementwise-simd-small-n.
void plusScalar   (const double *a, const double *b, double *out, std::size_t n);
void minusScalar  (const double *a, const double *b, double *out, std::size_t n);
void timesScalar  (const double *a, const double *b, double *out, std::size_t n);
void rdivideScalar(const double *a, const double *b, double *out, std::size_t n);
 
// Matrix multiply over column-major real doubles.
//  A: M×K at memory layout a[k*M + i]
//  B: K×N at memory layout b[j*K + k]
//  C: M×N written as c[j*M + i]; caller allocates, this function zeros
// and then accumulates. Loop order is (j, k, i) — SAXPY down columns of
// A with the (k,j) element of B as the scalar — cache-friendly and
// directly vectorisable with MulAdd.
void matmulDoubleLoop(const double *a, const double *b, double *c,
                      std::size_t M, std::size_t N, std::size_t K);
 
} // namespace nk::ops::detail