cannam@167
|
1 (*
|
cannam@167
|
2 * Copyright (c) 1997-1999 Massachusetts Institute of Technology
|
cannam@167
|
3 * Copyright (c) 2003, 2007-14 Matteo Frigo
|
cannam@167
|
4 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
cannam@167
|
5 *
|
cannam@167
|
6 * This program is free software; you can redistribute it and/or modify
|
cannam@167
|
7 * it under the terms of the GNU General Public License as published by
|
cannam@167
|
8 * the Free Software Foundation; either version 2 of the License, or
|
cannam@167
|
9 * (at your option) any later version.
|
cannam@167
|
10 *
|
cannam@167
|
11 * This program is distributed in the hope that it will be useful,
|
cannam@167
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
cannam@167
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
cannam@167
|
14 * GNU General Public License for more details.
|
cannam@167
|
15 *
|
cannam@167
|
16 * You should have received a copy of the GNU General Public License
|
cannam@167
|
17 * along with this program; if not, write to the Free Software
|
cannam@167
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
cannam@167
|
19 *
|
cannam@167
|
20 *)
|
cannam@167
|
21
|
cannam@167
|
22 open Util
|
cannam@167
|
23 open Genutil
|
cannam@167
|
24 open C
|
cannam@167
|
25
|
cannam@167
|
26 type ditdif = DIT | DIF
|
cannam@167
|
27 let ditdif = ref DIT
|
cannam@167
|
28
|
cannam@167
|
29 let usage = "Usage: " ^ Sys.argv.(0) ^ " -n <number> [ -dit | -dif ]"
|
cannam@167
|
30
|
cannam@167
|
31 let reload_twiddle = ref false
|
cannam@167
|
32
|
cannam@167
|
33 let urs = ref Stride_variable
|
cannam@167
|
34 let uvs = ref Stride_variable
|
cannam@167
|
35 let ums = ref Stride_variable
|
cannam@167
|
36
|
cannam@167
|
37 let speclist = [
|
cannam@167
|
38 "-dit",
|
cannam@167
|
39 Arg.Unit(fun () -> ditdif := DIT),
|
cannam@167
|
40 " generate a DIT codelet";
|
cannam@167
|
41
|
cannam@167
|
42 "-dif",
|
cannam@167
|
43 Arg.Unit(fun () -> ditdif := DIF),
|
cannam@167
|
44 " generate a DIF codelet";
|
cannam@167
|
45
|
cannam@167
|
46 "-reload-twiddle",
|
cannam@167
|
47 Arg.Unit(fun () -> reload_twiddle := true),
|
cannam@167
|
48 " do not collect common twiddle factors";
|
cannam@167
|
49
|
cannam@167
|
50 "-with-rs",
|
cannam@167
|
51 Arg.String(fun x -> urs := arg_to_stride x),
|
cannam@167
|
52 " specialize for given input stride";
|
cannam@167
|
53
|
cannam@167
|
54 "-with-vs",
|
cannam@167
|
55 Arg.String(fun x -> uvs := arg_to_stride x),
|
cannam@167
|
56 " specialize for given vector stride";
|
cannam@167
|
57
|
cannam@167
|
58 "-with-ms",
|
cannam@167
|
59 Arg.String(fun x -> ums := arg_to_stride x),
|
cannam@167
|
60 " specialize for given ms"
|
cannam@167
|
61 ]
|
cannam@167
|
62
|
cannam@167
|
63 let generate n =
|
cannam@167
|
64 let rioarray = "x"
|
cannam@167
|
65 and rs = "rs" and vs = "vs"
|
cannam@167
|
66 and twarray = "W"
|
cannam@167
|
67 and m = "m" and mb = "mb" and me = "me" and ms = "ms" in
|
cannam@167
|
68
|
cannam@167
|
69 let sign = !Genutil.sign
|
cannam@167
|
70 and name = !Magic.codelet_name
|
cannam@167
|
71 and byvl x = choose_simd x (ctimes (CVar "VL", x))
|
cannam@167
|
72 and bytwvl x = choose_simd x (ctimes (CVar "TWVL", x))
|
cannam@167
|
73 and bytwvl_vl x = choose_simd x (ctimes (CVar "(TWVL/VL)", x)) in
|
cannam@167
|
74 let ename = expand_name name in
|
cannam@167
|
75
|
cannam@167
|
76 let (bytwiddle, num_twiddles, twdesc) = Twiddle.twiddle_policy 0 true in
|
cannam@167
|
77 let nt = num_twiddles n in
|
cannam@167
|
78
|
cannam@167
|
79 let svs = either_stride (!uvs) (C.SVar vs)
|
cannam@167
|
80 and srs = either_stride (!urs) (C.SVar rs) in
|
cannam@167
|
81 let sms = stride_to_string "ms" !ums in
|
cannam@167
|
82
|
cannam@167
|
83 let byw =
|
cannam@167
|
84 if !reload_twiddle then
|
cannam@167
|
85 array n (fun v -> bytwiddle n sign (twiddle_array nt twarray))
|
cannam@167
|
86 else
|
cannam@167
|
87 let a = bytwiddle n sign (twiddle_array nt twarray)
|
cannam@167
|
88 in fun v -> a
|
cannam@167
|
89 in
|
cannam@167
|
90
|
cannam@167
|
91 let locations = unique_v_array_c n n in
|
cannam@167
|
92
|
cannam@167
|
93 let ioi =
|
cannam@167
|
94 locative_v_array_c n n
|
cannam@167
|
95 (C.varray_subscript rioarray svs srs)
|
cannam@167
|
96 (C.varray_subscript "BUG" svs srs)
|
cannam@167
|
97 locations sms
|
cannam@167
|
98 and ioo =
|
cannam@167
|
99 locative_v_array_c n n
|
cannam@167
|
100 (C.varray_subscript rioarray svs srs)
|
cannam@167
|
101 (C.varray_subscript "BUG" svs srs)
|
cannam@167
|
102 locations sms
|
cannam@167
|
103 in
|
cannam@167
|
104
|
cannam@167
|
105 let lioi = load_v_array_c n n ioi in
|
cannam@167
|
106 let fft = Trig.dft_via_rdft in
|
cannam@167
|
107 let output =
|
cannam@167
|
108 match !ditdif with
|
cannam@167
|
109 | DIT -> array n (fun v -> fft sign n (byw v (lioi v)))
|
cannam@167
|
110 | DIF -> array n (fun v -> byw v (fft sign n (lioi v)))
|
cannam@167
|
111 in
|
cannam@167
|
112
|
cannam@167
|
113 let odag = store_v_array_c n n ioo (transpose output) in
|
cannam@167
|
114 let annot = standard_optimizer odag in
|
cannam@167
|
115
|
cannam@167
|
116 let vm = CVar m and vmb = CVar mb and vme = CVar me in
|
cannam@167
|
117
|
cannam@167
|
118 let body = Block (
|
cannam@167
|
119 [Decl ("INT", m);
|
cannam@167
|
120 Decl (C.realtypep, rioarray)],
|
cannam@167
|
121 [Stmt_assign (CVar rioarray,
|
cannam@167
|
122 CVar (if (sign < 0) then "ri" else "ii"));
|
cannam@167
|
123 For (list_to_comma
|
cannam@167
|
124 [Expr_assign (vm, vmb);
|
cannam@167
|
125 Expr_assign (CVar twarray,
|
cannam@167
|
126 CPlus [CVar twarray;
|
cannam@167
|
127 ctimes (vmb,
|
cannam@167
|
128 bytwvl_vl (Integer nt))])],
|
cannam@167
|
129 Binop (" < ", vm, vme),
|
cannam@167
|
130 list_to_comma
|
cannam@167
|
131 [Expr_assign (vm, CPlus [vm; byvl (Integer 1)]);
|
cannam@167
|
132 Expr_assign (CVar rioarray, CPlus [CVar rioarray;
|
cannam@167
|
133 byvl (CVar sms)]);
|
cannam@167
|
134 Expr_assign (CVar twarray, CPlus [CVar twarray;
|
cannam@167
|
135 bytwvl (Integer nt)]);
|
cannam@167
|
136 make_volatile_stride (2*n) (CVar rs);
|
cannam@167
|
137 make_volatile_stride (2*n) (CVar vs)
|
cannam@167
|
138 ],
|
cannam@167
|
139 Asch annot)]) in
|
cannam@167
|
140
|
cannam@167
|
141 let tree =
|
cannam@167
|
142 Fcn (("static void"), ename,
|
cannam@167
|
143 [Decl (C.realtypep, "ri");
|
cannam@167
|
144 Decl (C.realtypep, "ii");
|
cannam@167
|
145 Decl (C.constrealtypep, twarray);
|
cannam@167
|
146 Decl (C.stridetype, rs);
|
cannam@167
|
147 Decl (C.stridetype, vs);
|
cannam@167
|
148 Decl ("INT", mb);
|
cannam@167
|
149 Decl ("INT", me);
|
cannam@167
|
150 Decl ("INT", ms)],
|
cannam@167
|
151 finalize_fcn body)
|
cannam@167
|
152 in
|
cannam@167
|
153 let twinstr =
|
cannam@167
|
154 Printf.sprintf "static const tw_instr twinstr[] = %s;\n\n"
|
cannam@167
|
155 (twinstr_to_string "VL" (twdesc n))
|
cannam@167
|
156
|
cannam@167
|
157 and desc =
|
cannam@167
|
158 Printf.sprintf
|
cannam@167
|
159 "static const ct_desc desc = {%d, %s, twinstr, &GENUS, %s, %s, %s, %s};\n\n"
|
cannam@167
|
160 n (stringify name) (flops_of tree)
|
cannam@167
|
161 (stride_to_solverparm !urs)
|
cannam@167
|
162 (stride_to_solverparm !uvs)
|
cannam@167
|
163 (stride_to_solverparm !ums)
|
cannam@167
|
164
|
cannam@167
|
165 and register =
|
cannam@167
|
166 match !ditdif with
|
cannam@167
|
167 | DIT -> "X(kdft_ditsq_register)"
|
cannam@167
|
168 | DIF -> "X(kdft_difsq_register)"
|
cannam@167
|
169 in
|
cannam@167
|
170 let init =
|
cannam@167
|
171 "\n" ^
|
cannam@167
|
172 twinstr ^
|
cannam@167
|
173 desc ^
|
cannam@167
|
174 (declare_register_fcn name) ^
|
cannam@167
|
175 (Printf.sprintf "{\n%s(p, %s, &desc);\n}" register ename)
|
cannam@167
|
176 in
|
cannam@167
|
177
|
cannam@167
|
178 (unparse tree) ^ "\n" ^ init
|
cannam@167
|
179
|
cannam@167
|
180
|
cannam@167
|
181 let main () =
|
cannam@167
|
182 begin
|
cannam@167
|
183 parse (speclist @ Twiddle.speclist) usage;
|
cannam@167
|
184 print_string (generate (check_size ()));
|
cannam@167
|
185 end
|
cannam@167
|
186
|
cannam@167
|
187 let _ = main()
|