(** All utilities which should be included in all source files are defined here *)

(** {1 Functions from libm} *)

open Constants
open Lacaml.D

let factmax = 150

external erf_float : float -> float = "erf_float_bytecode" "erf_float"
  [@@unboxed] [@@noalloc]

external erfc_float : float -> float = "erfc_float_bytecode" "erfc_float"
  [@@unboxed] [@@noalloc]

external gamma_float : float -> float = "gamma_float_bytecode" "gamma_float"
  [@@unboxed] [@@noalloc]






(* Incomplete gamma function : Int_0^x   exp(-t) t^(a-1) dt  
  p:  1 / Gamma(a) * Int_0^x   exp(-t) t^(a-1) dt  
  q:  1 / Gamma(a) * Int_x^inf exp(-t) t^(a-1) dt  

  reference - Haruhiko Okumura: C-gengo niyoru saishin algorithm jiten
  (New Algorithm handbook in C language) (Gijyutsu hyouron
  sha, Tokyo, 1991) p.227 [in Japanese] *)

let incomplete_gamma ~alpha x = 
  let a = alpha in
  let a_inv = 1./. a in
  let gf = gamma_float alpha in
  let loggamma_a = log gf in
  let rec p_gamma x =
    if x >= 1. +. a then 1. -. q_gamma x 
    else if x = 0. then 0.
    else
      let rec pg_loop prev res term k =
        if k > 1000. then failwith "p_gamma did not converge."
        else if prev = res then res
        else
          let term = term *. x /. (a +. k) in
          pg_loop res (res +. term) term (k +. 1.)
      in
      let r0 =  exp (a *. log x -. x -. loggamma_a) *. a_inv in
      pg_loop min_float r0 r0 1.

  and q_gamma x =
    if x < 1. +. a then 1. -. p_gamma x 
    else
      let rec qg_loop prev res la lb w k =
        if k > 1000. then failwith "q_gamma did not converge."
        else if prev = res then res
        else
          let k_inv = 1. /. k in
          let kma = (k -. 1. -. a) *. k_inv in
          let la, lb =
            lb, kma *. (lb -. la) +. (k +. x) *. lb *. k_inv
          in
          let w = w *. kma  in
          let prev, res = res, res +. w /. (la *. lb) in
          qg_loop prev res la lb w (k +. 1.)
      in
      let w = exp (a *. log x -. x -. loggamma_a) in
      let lb = (1. +. x -. a) in
      qg_loop min_float (w /. lb) 1. lb w 2.0
  in
  gf *. p_gamma x 





let fact_memo =
  let rec aux accu_l accu = function
    | 0 ->  aux [1.] 1. 1
    | i when (i = factmax) ->
      let x = (float_of_int factmax) *. accu in
      List.rev (x::accu_l)
    | i ->  let x = (float_of_int i) *. accu in
      aux (x::accu_l) x (i+1)
  in
  aux [] 0. 0
  |> Array.of_list



let fact = function
  | i when (i < 0) ->
    raise (Invalid_argument "Argument of factorial should be non-negative")
  | i when (i > 150) ->
    raise (Invalid_argument "Result of factorial is infinite")
  | i -> fact_memo.(i)


let rec pow a = function
  | 0 -> 1.
  | 1 -> a
  | 2 -> a *. a
  | 3 -> a *. a *. a
  | -1 -> 1. /. a
  | n when  n > 0  ->
    let b = pow a (n / 2) in
    b *. b *. (if n mod 2 = 0 then 1. else a)
  | n when  n < 0  -> pow (1./.a) (-n)
  | _ -> assert false




let chop f g =
  if (abs_float f) < Constants.epsilon then 0.
  else f *. (g ())



(** Generalized Boys function. 
    maxm : Maximum total angular momentum
*)
let boys_function ~maxm t =
  match maxm with
  | 0 ->
    begin
      if t = 0. then [| 1. |] else
        let sq_t = sqrt t in
        [| (sq_pi_over_two /. sq_t) *.  erf_float sq_t |]
    end
  | _ ->
    begin
      let result =
        Array.init (maxm+1) (fun m -> 1. /. float_of_int (2*m+1))
      in
      (*
      assert (abs_float t > 1.e-10);
      *)
      if t <> 0. then
        begin
          let fmax = 
            let t_inv = sqrt (1. /. t) in
            let n = float_of_int maxm in
            let dm = 0.5 +. n in
            let f  = (pow t_inv (maxm+maxm+1) ) in
            match classify_float f with
              | FP_zero
              | FP_subnormal
              | FP_normal ->
                (incomplete_gamma dm t) *. 0.5 *. f
              | _ -> invalid_arg "zero_m overflow"
          in
          let emt = exp (-. t) in
          result.(maxm) <- fmax;
          for n=maxm-1 downto 0 do
            result.(n) <- ( (t+.t) *. result.(n+1) +. emt) *. result.(n)
          done
        end;
      result
    end


(** {2 List functions} *)

let list_some l =
  List.filter (function None -> false | _ -> true) l
  |> List.map (function Some x -> x | _ -> assert false)


(** {2 Stream functions} *)

let stream_range ?(start=0) n = 
  Stream.from (fun i ->
    let result = i+start in
    if result <= n then
      Some result
    else None
  )


let list_range ?(start=0) n = 
  Array.init n (fun i -> start+i) |> Array.to_list




(** {2 Linear algebra} *)


let array_sum a = 
   Array.fold_left ( +. ) 0. a

let array_product a = 
   Array.fold_left ( *. ) 0. a


let diagonalize_symm m_H = 
  let m_V = lacpy m_H in
  let result = 
    syevd ~vectors:true m_V
  in
  m_V, result

let xt_o_x ~o ~x =
  gemm o x
  |> gemm ~transa:`T x


let canonical_ortho ?thresh:(thresh=1.e-6) ~overlap c =
  let d, u, _ = gesvd (lacpy overlap) in
  let d_sqrt = Vec.sqrt d in
  let n = (* Number of non-negligible singular vectors *)
    Vec.fold (fun accu x -> if x > thresh then accu + 1 else accu) 0 d
  in
  let d_inv_sq = (* D^{-1/2} *)
    Vec.map (fun x ->
      if x >= thresh then 1. /. x
      else 0. ) ~y:d d_sqrt
  in
  if n < Vec.dim d_sqrt then 
    Printf.printf "Removed linear dependencies below %f\n" (1. /. d.{n})
  ;
  Mat.scal_cols u d_inv_sq ;
  gemm c u



(** {2 Printers} *)

let pp_float_array_size ppf a = 
  Format.fprintf ppf "@[<2>@[ %d:@[<2>" (Array.length a);
  Array.iter (fun f -> Format.fprintf ppf "@[%10f@]@ " f) a;
  Format.fprintf ppf "]@]@]"

let pp_float_array ppf a = 
  Format.fprintf ppf "@[<2>[@ ";
  Array.iter (fun f -> Format.fprintf ppf "@[%10f@]@ " f) a;
  Format.fprintf ppf "]@]"

let pp_float_2darray ppf a = 
  Format.fprintf ppf "@[<2>[@ ";
  Array.iter (fun f -> Format.fprintf ppf "@[%a@]@ " pp_float_array f) a;
  Format.fprintf ppf "]@]"

let pp_float_2darray_size ppf a = 
  Format.fprintf ppf "@[<2>@[ %d:@[" (Array.length a);
  Array.iter (fun f -> Format.fprintf ppf "@[%a@]@ " pp_float_array_size f) a;
  Format.fprintf ppf "]@]@]"

let pp_matrix ppf m = 
  let open Lacaml.Io in
  let rows = Mat.dim1 m
  and cols = Mat.dim2 m
  in
  let rec aux first last =
    if (first <= last) then begin
        Format.fprintf ppf "@[\n\n    %a@]@ " (Lacaml.Io.pp_lfmat
          ~row_labels:
            (Array.init rows (fun i -> Printf.sprintf "%d  " (i + 1)))
          ~col_labels:
            (Array.init (min 5 (cols-first+1)) (fun i -> Printf.sprintf "-- %d --" (i + first) ))
          ~print_right:false
          ~print_foot:false
          () ) (lacpy ~ac:first ~n:(min 5 (cols-first+1)) m);
        aux (first+5) last
    end
  in
  aux 1 cols




let string_of_matrix m = 
  Format.asprintf "%a" pp_matrix m

let debug_matrix name a =
  Format.printf "@[%s =\n@[%a@]@]@ " name pp_matrix a


let matrix_of_file filename =
  let ic = Scanf.Scanning.open_in filename in
  let rec read_line accu =
    let result =
      try
        Some (Scanf.bscanf ic " %d %d %f" (fun i j v ->
          (i,j,v) :: accu))
      with End_of_file -> None
    in
    match result with
    | Some accu -> read_line accu
    | None      -> List.rev accu
  in
  let data = read_line [] in
  Scanf.Scanning.close_in ic;
  let isize, jsize = 
    List.fold_left (fun (accu_i,accu_j) (i,j,v) ->
      (max i accu_i, max j accu_j)) (0,0) data
  in
  let result =
    Lacaml.D.Mat.of_array
      (Array.make_matrix isize jsize 0.)
  in
  List.iter (fun (i,j,v) -> result.{i,j} <- v) data;
  result


let sym_matrix_of_file filename =
  let result = 
    matrix_of_file filename
  in
  for j=1 to Mat.dim1 result do
    for i=1 to j do
      result.{j,i} <- result.{i,j}
    done;
  done;
  result