Chris@10
|
1 (*
|
Chris@10
|
2 * Copyright (c) 1997-1999 Massachusetts Institute of Technology
|
Chris@10
|
3 * Copyright (c) 2003, 2007-11 Matteo Frigo
|
Chris@10
|
4 * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
|
Chris@10
|
5 *
|
Chris@10
|
6 * This program is free software; you can redistribute it and/or modify
|
Chris@10
|
7 * it under the terms of the GNU General Public License as published by
|
Chris@10
|
8 * the Free Software Foundation; either version 2 of the License, or
|
Chris@10
|
9 * (at your option) any later version.
|
Chris@10
|
10 *
|
Chris@10
|
11 * This program is distributed in the hope that it will be useful,
|
Chris@10
|
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
Chris@10
|
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
Chris@10
|
14 * GNU General Public License for more details.
|
Chris@10
|
15 *
|
Chris@10
|
16 * You should have received a copy of the GNU General Public License
|
Chris@10
|
17 * along with this program; if not, write to the Free Software
|
Chris@10
|
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
Chris@10
|
19 *
|
Chris@10
|
20 *)
|
Chris@10
|
21
|
Chris@10
|
22 open Util
|
Chris@10
|
23 open Genutil
|
Chris@10
|
24 open C
|
Chris@10
|
25
|
Chris@10
|
26
|
Chris@10
|
27 let usage = "Usage: " ^ Sys.argv.(0) ^ " -n <number>"
|
Chris@10
|
28
|
Chris@10
|
29 let urs = ref Stride_variable
|
Chris@10
|
30 let ucsr = ref Stride_variable
|
Chris@10
|
31 let ucsi = ref Stride_variable
|
Chris@10
|
32 let uivs = ref Stride_variable
|
Chris@10
|
33 let uovs = ref Stride_variable
|
Chris@10
|
34 let dftIII_flag = ref false
|
Chris@10
|
35
|
Chris@10
|
36 let speclist = [
|
Chris@10
|
37 "-with-rs",
|
Chris@10
|
38 Arg.String(fun x -> urs := arg_to_stride x),
|
Chris@10
|
39 " specialize for given real-array stride";
|
Chris@10
|
40
|
Chris@10
|
41 "-with-csr",
|
Chris@10
|
42 Arg.String(fun x -> ucsr := arg_to_stride x),
|
Chris@10
|
43 " specialize for given complex-array real stride";
|
Chris@10
|
44
|
Chris@10
|
45 "-with-csi",
|
Chris@10
|
46 Arg.String(fun x -> ucsi := arg_to_stride x),
|
Chris@10
|
47 " specialize for given complex-array imaginary stride";
|
Chris@10
|
48
|
Chris@10
|
49 "-with-ivs",
|
Chris@10
|
50 Arg.String(fun x -> uivs := arg_to_stride x),
|
Chris@10
|
51 " specialize for given input vector stride";
|
Chris@10
|
52
|
Chris@10
|
53 "-with-ovs",
|
Chris@10
|
54 Arg.String(fun x -> uovs := arg_to_stride x),
|
Chris@10
|
55 " specialize for given output vector stride";
|
Chris@10
|
56
|
Chris@10
|
57 "-dft-III",
|
Chris@10
|
58 Arg.Unit(fun () -> dftIII_flag := true),
|
Chris@10
|
59 " produce shifted dftIII-style codelets"
|
Chris@10
|
60 ]
|
Chris@10
|
61
|
Chris@10
|
62 let hcdftIII sign n input =
|
Chris@10
|
63 let input' i =
|
Chris@10
|
64 if (i mod 2 == 0) then
|
Chris@10
|
65 Complex.zero
|
Chris@10
|
66 else
|
Chris@10
|
67 let i' = (i - 1) / 2 in
|
Chris@10
|
68 if (2 * i' < n - 1) then (input i')
|
Chris@10
|
69 else if (2 * i' == n - 1) then
|
Chris@10
|
70 Complex.real (input i')
|
Chris@10
|
71 else
|
Chris@10
|
72 Complex.conj (input (n - 1 - i'))
|
Chris@10
|
73 in Fft.dft sign (2 * n) input'
|
Chris@10
|
74
|
Chris@10
|
75 let generate n =
|
Chris@10
|
76 let ar0 = "R0" and ar1 = "R1" and acr = "Cr" and aci = "Ci"
|
Chris@10
|
77 and rs = "rs" and csr = "csr" and csi = "csi"
|
Chris@10
|
78 and i = "i" and v = "v"
|
Chris@10
|
79 and transform = if !dftIII_flag then hcdftIII else Trig.hdft
|
Chris@10
|
80 in
|
Chris@10
|
81
|
Chris@10
|
82 let sign = !Genutil.sign
|
Chris@10
|
83 and name = !Magic.codelet_name in
|
Chris@10
|
84
|
Chris@10
|
85 let vrs = either_stride (!urs) (C.SVar rs)
|
Chris@10
|
86 and vcsr = either_stride (!ucsr) (C.SVar csr)
|
Chris@10
|
87 and vcsi = either_stride (!ucsi) (C.SVar csi)
|
Chris@10
|
88 in
|
Chris@10
|
89
|
Chris@10
|
90 let sovs = stride_to_string "ovs" !uovs in
|
Chris@10
|
91 let sivs = stride_to_string "ivs" !uivs in
|
Chris@10
|
92
|
Chris@10
|
93 let locations = unique_array_c n in
|
Chris@10
|
94 let input =
|
Chris@10
|
95 locative_array_c n
|
Chris@10
|
96 (C.array_subscript acr vcsr)
|
Chris@10
|
97 (C.array_subscript aci vcsi)
|
Chris@10
|
98 locations sivs in
|
Chris@10
|
99 let output = transform sign n (load_array_hc n input) in
|
Chris@10
|
100 let oloce =
|
Chris@10
|
101 locative_array_c n
|
Chris@10
|
102 (C.array_subscript ar0 vrs)
|
Chris@10
|
103 (C.array_subscript "BUG" vrs)
|
Chris@10
|
104 locations sovs
|
Chris@10
|
105 and oloco =
|
Chris@10
|
106 locative_array_c n
|
Chris@10
|
107 (C.array_subscript ar1 vrs)
|
Chris@10
|
108 (C.array_subscript "BUG" vrs)
|
Chris@10
|
109 locations sovs in
|
Chris@10
|
110 let oloc i = if i mod 2 == 0 then oloce (i/2) else oloco ((i-1)/2) in
|
Chris@10
|
111 let odag = store_array_r n oloc output in
|
Chris@10
|
112 let annot = standard_optimizer odag in
|
Chris@10
|
113
|
Chris@10
|
114 let body = Block (
|
Chris@10
|
115 [Decl ("INT", i)],
|
Chris@10
|
116 [For (Expr_assign (CVar i, CVar v),
|
Chris@10
|
117 Binop (" > ", CVar i, Integer 0),
|
Chris@10
|
118 list_to_comma
|
Chris@10
|
119 [Expr_assign (CVar i, CPlus [CVar i; CUminus (Integer 1)]);
|
Chris@10
|
120 Expr_assign (CVar ar0, CPlus [CVar ar0; CVar sovs]);
|
Chris@10
|
121 Expr_assign (CVar ar1, CPlus [CVar ar1; CVar sovs]);
|
Chris@10
|
122 Expr_assign (CVar acr, CPlus [CVar acr; CVar sivs]);
|
Chris@10
|
123 Expr_assign (CVar aci, CPlus [CVar aci; CVar sivs]);
|
Chris@10
|
124 make_volatile_stride (4*n) (CVar rs);
|
Chris@10
|
125 make_volatile_stride (4*n) (CVar csr);
|
Chris@10
|
126 make_volatile_stride (4*n) (CVar csi)
|
Chris@10
|
127 ],
|
Chris@10
|
128 Asch annot)
|
Chris@10
|
129 ])
|
Chris@10
|
130 in
|
Chris@10
|
131
|
Chris@10
|
132 let tree =
|
Chris@10
|
133 Fcn ((if !Magic.standalone then "void" else "static void"), name,
|
Chris@10
|
134 ([Decl (C.realtypep, ar0);
|
Chris@10
|
135 Decl (C.realtypep, ar1);
|
Chris@10
|
136 Decl (C.realtypep, acr);
|
Chris@10
|
137 Decl (C.realtypep, aci);
|
Chris@10
|
138 Decl (C.stridetype, rs);
|
Chris@10
|
139 Decl (C.stridetype, csr);
|
Chris@10
|
140 Decl (C.stridetype, csi);
|
Chris@10
|
141 Decl ("INT", v);
|
Chris@10
|
142 Decl ("INT", "ivs");
|
Chris@10
|
143 Decl ("INT", "ovs")]),
|
Chris@10
|
144 finalize_fcn body)
|
Chris@10
|
145
|
Chris@10
|
146 in let desc =
|
Chris@10
|
147 Printf.sprintf
|
Chris@10
|
148 "static const kr2c_desc desc = { %d, \"%s\", %s, &GENUS };\n\n"
|
Chris@10
|
149 n name (flops_of tree)
|
Chris@10
|
150
|
Chris@10
|
151 and init =
|
Chris@10
|
152 (declare_register_fcn name) ^
|
Chris@10
|
153 "{" ^
|
Chris@10
|
154 " X(kr2c_register)(p, " ^ name ^ ", &desc);\n" ^
|
Chris@10
|
155 "}\n"
|
Chris@10
|
156
|
Chris@10
|
157 in
|
Chris@10
|
158 (unparse tree) ^ "\n" ^ (if !Magic.standalone then "" else desc ^ init)
|
Chris@10
|
159
|
Chris@10
|
160
|
Chris@10
|
161 let main () =
|
Chris@10
|
162 begin
|
Chris@10
|
163 parse speclist usage;
|
Chris@10
|
164 print_string (generate (check_size ()));
|
Chris@10
|
165 end
|
Chris@10
|
166
|
Chris@10
|
167 let _ = main()
|