21
21
/* This file was automatically generated --- DO NOT EDIT */
22
/* Generated on Sat Jul 5 21:43:15 EDT 2003 */
22
/* Generated on Tue Mar 7 10:08:03 EST 2006 */
24
24
#include "codelet-dft.h"
26
/* Generated by: /homee/stevenj/cvs/fftw3.0.1/genfft/gen_twiddle_c -simd -compact -variables 4 -n 4 -name t1fv_4 -include t1f.h */
28
/* Generated by: ../../../genfft/gen_twiddle_c -fma -reorder-insns -schedule-for-pipeline -simd -compact -variables 4 -pipeline-latency 8 -n 4 -name t1fv_4 -include t1f.h */
31
* This function contains 11 FP additions, 8 FP multiplications,
32
* (or, 9 additions, 6 multiplications, 2 fused multiply/add),
33
* 13 stack variables, and 8 memory accesses
37
* $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
38
* $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
39
* $Id: gen_twiddle_c.ml,v 1.14 2006-02-12 23:34:12 athena Exp $
44
static const R *t1fv_4(R *ri, R *ii, const R *W, stride ios, INT m, INT dist)
49
for (i = m; i > 0; i = i - VL, x = x + (VL * dist), W = W + (TWVL * 6), MAKE_VOLATILE_STRIDE(ios)) {
50
V T1, T7, T2, T5, T8, T3, T6;
51
T1 = LD(&(x[0]), dist, &(x[0]));
52
T7 = LD(&(x[WS(ios, 3)]), dist, &(x[WS(ios, 1)]));
53
T2 = LD(&(x[WS(ios, 2)]), dist, &(x[0]));
54
T5 = LD(&(x[WS(ios, 1)]), dist, &(x[WS(ios, 1)]));
55
T8 = BYTWJ(&(W[TWVL * 4]), T7);
56
T3 = BYTWJ(&(W[TWVL * 2]), T2);
57
T6 = BYTWJ(&(W[0]), T5);
64
ST(&(x[0]), VADD(Ta, Tb), dist, &(x[0]));
65
ST(&(x[WS(ios, 2)]), VSUB(Ta, Tb), dist, &(x[0]));
66
ST(&(x[WS(ios, 3)]), VFMAI(T9, T4), dist, &(x[WS(ios, 1)]));
67
ST(&(x[WS(ios, 1)]), VFNMSI(T9, T4), dist, &(x[WS(ios, 1)]));
73
static const tw_instr twinstr[] = {
80
static const ct_desc desc = { 4, "t1fv_4", twinstr, &GENUS, {9, 6, 2, 0}, 0, 0, 0 };
82
void X(codelet_t1fv_4) (planner *p) {
83
X(kdft_dit_register) (p, t1fv_4, &desc);
87
/* Generated by: ../../../genfft/gen_twiddle_c -simd -compact -variables 4 -pipeline-latency 8 -n 4 -name t1fv_4 -include t1f.h */
29
90
* This function contains 11 FP additions, 6 FP multiplications,
35
* $Id: algsimp.ml,v 1.7 2003/03/15 20:29:42 stevenj Exp $
36
* $Id: fft.ml,v 1.2 2003/03/15 20:29:42 stevenj Exp $
37
* $Id: gen_twiddle_c.ml,v 1.7 2003/04/16 19:51:27 athena Exp $
96
* $Id: algsimp.ml,v 1.9 2006-02-12 23:34:12 athena Exp $
97
* $Id: fft.ml,v 1.4 2006-01-05 03:04:27 stevenj Exp $
98
* $Id: gen_twiddle_c.ml,v 1.14 2006-02-12 23:34:12 athena Exp $
42
static const R *t1fv_4(R *ri, R *ii, const R *W, stride ios, int m, int dist)
103
static const R *t1fv_4(R *ri, R *ii, const R *W, stride ios, INT m, INT dist)
48
for (i = m; i > 0; i = i - VL, x = x + (VL * dist), W = W + (TWVL * 6)) {
108
for (i = m; i > 0; i = i - VL, x = x + (VL * dist), W = W + (TWVL * 6), MAKE_VOLATILE_STRIDE(ios)) {
49
109
V T1, T8, T3, T6, T7, T2, T5;
50
110
T1 = LD(&(x[0]), dist, &(x[0]));
51
111
T7 = LD(&(x[WS(ios, 3)]), dist, &(x[WS(ios, 1)]));
80
static const ct_desc desc = { 4, "t1fv_4", twinstr, {11, 6, 0, 0}, &GENUS, 0, 0, 0 };
139
static const ct_desc desc = { 4, "t1fv_4", twinstr, &GENUS, {11, 6, 0, 0}, 0, 0, 0 };
82
141
void X(codelet_t1fv_4) (planner *p) {
83
142
X(kdft_dit_register) (p, t1fv_4, &desc);
144
#endif /* HAVE_FMA */