// bend-blas, the JS twins: plain loops, correct and slow, for the
// interpreter and the tests. An Array<F32> is a JS array of numbers.
// A failure is io_fail(code): the runtime's message for a code is not
// this library's, so the code is the contract on this lane.
function bend_fresh(n) {
  let slots = 1;
  while (slots < n) slots *= 2;
  return new Array(slots).fill(0);
}
function bend_copy(src, n) {
  const c = bend_fresh(n);
  for (let i = 0; i < n; i++) c[i] = src[i];
  return c;
}
function bend_short(a, n) {
  return n > 0x7fffffff || a.length < n;
}
function bend_gemm(ta, tb, m, n, k, alpha, a, b, beta, c) {
  const out = bend_copy(c, m * n);
  for (let i = 0; i < m; i++) for (let j = 0; j < n; j++) {
    let acc = 0;
    for (let l = 0; l < k; l++) acc += (ta ? a[l * m + i] : a[i * k + l]) * (tb ? b[j * k + l] : b[l * n + j]);
    out[i * n + j] = alpha * acc + beta * out[i * n + j];
  }
  return out;
}
// the device buffer table of the JS twins: host arrays
if (typeof globalThis.bend_devs === "undefined") globalThis.bend_devs = [];
function cublas_upload(n, a) {
  if (bend_short(a, n)) return io_fail(2);
  globalThis.bend_devs.push(bend_copy(a, n));
  return io_done(globalThis.bend_devs.length - 1);
}
io_eff(CID(Cublas.upload), cublas_upload);