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 06:45:41 EDT 2009 */
22
/* Generated on Wed Jul 27 06:18:29 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 r2cb_4 -include r2cb.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 r2cb_4 -include r2cb.h */
31
31
* This function contains 6 FP additions, 4 FP multiplications,
37
37
static void r2cb_4(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
39
39
DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
41
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)) {
42
E T4, T6, T1, T2, T3, T5;
49
R1[0] = FNMS(KP2_000000000, T6, T5);
50
R1[WS(rs, 1)] = FMA(KP2_000000000, T6, T5);
51
R0[0] = FMA(KP2_000000000, T4, T3);
52
R0[WS(rs, 1)] = FNMS(KP2_000000000, T4, T3);
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 T4, T6, T1, T2, T3, T5;
50
R1[0] = FNMS(KP2_000000000, T6, T5);
51
R1[WS(rs, 1)] = FMA(KP2_000000000, T6, T5);
52
R0[0] = FMA(KP2_000000000, T4, T3);
53
R0[WS(rs, 1)] = FNMS(KP2_000000000, T4, T3);
62
64
#else /* HAVE_FMA */
64
/* Generated by: ../../../genfft/gen_r2cb -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cb_4 -include r2cb.h */
66
/* Generated by: ../../../genfft/gen_r2cb.native -compact -variables 4 -pipeline-latency 4 -sign 1 -n 4 -name r2cb_4 -include r2cb.h */
67
69
* This function contains 6 FP additions, 2 FP multiplications,
73
75
static void r2cb_4(R *R0, R *R1, R *Cr, R *Ci, stride rs, stride csr, stride csi, INT v, INT ivs, INT ovs)
75
77
DK(KP2_000000000, +2.000000000000000000000000000000000000000000000);
77
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)) {
82
T5 = KP2_000000000 * T4;
84
T8 = KP2_000000000 * T7;
80
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)) {
85
T5 = KP2_000000000 * T4;
87
T8 = KP2_000000000 * T7;
93
R0[WS(rs, 1)] = T3 - T5;
94
R1[WS(rs, 1)] = T6 + T8;
90
R0[WS(rs, 1)] = T3 - T5;
91
R1[WS(rs, 1)] = T6 + T8;