#pragma once #include // C export layer — lets Python ctypes / Rust FFI call the C++20 kernels // without name-mangling. Every function here maps to a theorem in ArrayLang/. #ifdef __cplusplus extern "C" { #endif // ── Paper I: NAND ────────────────────────────────────────────────────────────── // nand(a,b) = !(a && b) [maps to andGate_eq, notGate_eq, orGate_eq] int sovarr_nand(int a, int b); int sovarr_not(int a); int sovarr_and(int a, int b); int sovarr_or(int a, int b); // ── Paper II: Softmax + Face Centroid ───────────────────────────────────────── // softmax(in, out, n): out[i] = exp(in[i]) / Σ exp(in[j]) // [maps to softmax_is_pmap, softmax_shift_invariant] void sovarr_softmax(const float* in, float* out, size_t n); // face_centroid(support, support_len, out, n): // out[i] = 1/|F| if i ∈ support, else 0 // [maps to faceCentroid, faceCentroid_support] void sovarr_face_centroid(const int* support, size_t support_len, float* out, size_t n); // ── Paper III: Attention ─────────────────────────────────────────────────────── // nand_attention(q,k,v,out,n): scores_i = Σ_j q_i*k_j, w=softmax(scores), out_i = Σ_j w_i*v_j // [maps to attention_is_pmap] void sovarr_nand_attention(const float* q, const float* k, const float* v, float* out, size_t n); // ── Broadcast ───────────────────────────────────────────────────────────────── // broadcast_1d(v, w, out, n): out[i] = v[i] + w[i] // (1D case of broadcast_is_pullback) void sovarr_broadcast_1d(const float* v, const float* w, float* out, size_t n); // ── Consistency probe ───────────────────────────────────────────────────────── // Returns build-time git SHA + kernel version string. const char* sovarr_version(void); #ifdef __cplusplus } #endif