1/* @ajax/mesh-ref-matmul: programs/matmul's exact mode on the CPU (mesh-ref matmul.rs), the seed of GPU
2 * inference. Row-major: A is M x K, B is K x N, C is M x N.
3 *
4 * - C = A B: every element a sequential sum over k from +0.0 (`acc = acc + a * b`, a correctly
5 *   rounded multiply then a correctly rounded add, both FTZ, never fused).
6 * - y = A x: per row, lane l sums k = l, l + 32, .. sequentially from +0.0, then the warp butterfly. */
7#ifndef MESH_REF_MATMUL_H
8#define MESH_REF_MATMUL_H
9
10#include <stddef.h>
11#include <stdint.h>
12
13/* Element i of the matrix with seed s: (h >> 8) 2^-23 - 1, in [-1, 1) (exact on every backend). */
14float mesh_matmul_value(uint32_t seed, uint32_t i);
15/* out[0..n) = value(seed, 0..n). */
16void mesh_matmul_matrix(uint32_t seed, size_t n, float *out);
17/* c (m x n) = a (m x k) b (k x n) in the exact order, rows split over `threads` threads (the result
18 * is the same for any count). */
19void mesh_matmul(const float *a, const float *b, size_t m, size_t k, size_t n, size_t threads, float *c);
20/* y (m) = a (m x k) x (k) in matvec's order. */
21void mesh_matvec(const float *a, const float *x, size_t m, size_t k, float *y);
22/* FNV-1a 64 over the words of a matrix: matmul's digest. */
23uint64_t mesh_matmul_digest(const float *c, size_t n);
24
25#endif