2
* Copyright (c) 2003, 2007-8 Matteo Frigo
3
* Copyright (c) 2003, 2007-8 Massachusetts Institute of Technology
2
* Copyright (c) 2003, 2007-11 Matteo Frigo
3
* Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
5
5
* This program is free software; you can redistribute it and/or modify
6
6
* it under the terms of the GNU General Public License as published by
21
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sun Jul 12 06:46:20 EDT 2009 */
22
/* Generated on Wed Jul 27 06:19:01 EDT 2011 */
24
24
#include "codelet-rdft.h"
28
/* Generated by: ../../../genfft/gen_r2cb -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cbIII_4 -dft-III -include r2cbIII.h */
28
/* Generated by: ../../../genfft/gen_r2cb.native -fma -reorder-insns -schedule-for-pipeline -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cbIII_4 -dft-III -include r2cbIII.h */
31
31
* This function contains 6 FP additions, 4 FP multiplications,
39
39
DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
40
40
DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
42
for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
43
E T1, T2, T4, T5, T3, T6;
48
R0[0] = KP2_000000000 * (T1 + T2);
50
R0[WS(rs, 1)] = KP2_000000000 * (T5 - T4);
52
R1[WS(rs, 1)] = -(KP1_414213562 * (T3 + T6));
53
R1[0] = KP1_414213562 * (T3 - T6);
43
for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
44
E T1, T2, T4, T5, T3, T6;
49
R0[0] = KP2_000000000 * (T1 + T2);
51
R0[WS(rs, 1)] = KP2_000000000 * (T5 - T4);
53
R1[WS(rs, 1)] = -(KP1_414213562 * (T3 + T6));
54
R1[0] = KP1_414213562 * (T3 - T6);
63
65
#else /* HAVE_FMA */
65
/* Generated by: ../../../genfft/gen_r2cb -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cbIII_4 -dft-III -include r2cbIII.h */
67
/* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cbIII_4 -dft-III -include r2cbIII.h */
68
70
* This function contains 6 FP additions, 4 FP multiplications,
76
78
DK(KP1_414213562, +1.414213562373095048801688724209698078569671875);
77
79
DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
79
for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
80
E T1, T2, T3, T4, T5, T6;
87
R0[0] = KP2_000000000 * (T1 + T2);
88
R0[WS(rs, 1)] = KP2_000000000 * (T5 - T4);
89
R1[0] = KP1_414213562 * (T3 - T6);
90
R1[WS(rs, 1)] = -(KP1_414213562 * (T3 + T6));
82
for (i = v; i > 0; i = i - 1, R0 = R0 + ovs, R1 = R1 + ovs, Cr = Cr + ivs, Ci = Ci + ivs, MAKE_VOLATILE_STRIDE(rs), MAKE_VOLATILE_STRIDE(csr), MAKE_VOLATILE_STRIDE(csi)) {
83
E T1, T2, T3, T4, T5, T6;
90
R0[0] = KP2_000000000 * (T1 + T2);
91
R0[WS(rs, 1)] = KP2_000000000 * (T5 - T4);
92
R1[0] = KP1_414213562 * (T3 - T6);
93
R1[WS(rs, 1)] = -(KP1_414213562 * (T3 + T6));