// BatchMatMul should have exactly the same batch dimensions and
  // BatchMatMulV2 should have broadcastable batch dimensions.
  //
  // The last two dimensions are non-batch dimensions that don't need to
  // participate in batch dimension compatibility check.
  if (std::is_same<OpT, BatchMatMulOp>()) {
    for (const auto& dim_pairs : llvm::zip(x_batches, y_batches)) {
      int64_t x_dim = std::get<0>(dim_pairs);

		if newmHandoff.waiting {
			newmHandoff.waiting = false
			notewakeup(&newmHandoff.wake)
		}
		unlock(&newmHandoff.lock)
		// The M has not started yet, but the template thread does not
		// participate in STW, so it will always process queued Ms and
		// it is safe to releasem.
		releasem(getg().m)
		return
	}
	newm1(mp)
	releasem(getg().m)
}

func newm1(mp *m) {
	if iscgo {