1 // Auto-generated file. Do not edit!
2 // Template: src/cs16-fftr/scalar.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11 #include <stddef.h>
12 #include <stdint.h>
13
14 #include <xnnpack/math.h>
15 #include <xnnpack/fft.h>
16
17
xnn_cs16_fftr_ukernel__scalar_x3(size_t samples,int16_t * data,const int16_t * twiddle)18 void xnn_cs16_fftr_ukernel__scalar_x3(
19 size_t samples,
20 int16_t* data,
21 const int16_t* twiddle)
22 {
23 assert(samples >= 2);
24 assert(samples % 2 == 0);
25 assert(data != NULL);
26 assert(data != NULL);
27 assert(twiddle != NULL);
28
29 int16_t* dl = data;
30 int16_t* dr = data + samples * 2;
31 int32_t vdcr = (int32_t) dl[0];
32 int32_t vdci = (int32_t) dl[1];
33
34 vdcr = math_asr_s32(vdcr * 16383 + 16384, 15);
35 vdci = math_asr_s32(vdci * 16383 + 16384, 15);
36
37 dl[0] = vdcr + vdci;
38 dl[1] = 0;
39 dl += 2;
40 dr[0] = vdcr - vdci;
41 dr[1] = 0;
42
43 samples >>= 1;
44
45 for (; samples >= 3; samples -= 3) {
46 dr -= 3 * 2;
47 int32_t vilr0 = dl[0];
48 int32_t vili0 = dl[1];
49 int32_t vilr1 = dl[2];
50 int32_t vili1 = dl[3];
51 int32_t vilr2 = dl[4];
52 int32_t vili2 = dl[5];
53 int32_t virr0 = (int32_t) dr[4];
54 int32_t viri0 = -(int32_t) dr[5];
55 int32_t virr1 = (int32_t) dr[2];
56 int32_t viri1 = -(int32_t) dr[3];
57 int32_t virr2 = (int32_t) dr[0];
58 int32_t viri2 = -(int32_t) dr[1];
59 const int32_t vtwr0 = twiddle[0];
60 const int32_t vtwi0 = twiddle[1];
61 const int32_t vtwr1 = twiddle[2];
62 const int32_t vtwi1 = twiddle[3];
63 const int32_t vtwr2 = twiddle[4];
64 const int32_t vtwi2 = twiddle[5];
65 twiddle += 3 * 2;
66
67 vilr0 = math_asr_s32(vilr0 * 16383 + 16384, 15);
68 virr0 = math_asr_s32(virr0 * 16383 + 16384, 15);
69 vilr1 = math_asr_s32(vilr1 * 16383 + 16384, 15);
70 virr1 = math_asr_s32(virr1 * 16383 + 16384, 15);
71 vilr2 = math_asr_s32(vilr2 * 16383 + 16384, 15);
72 virr2 = math_asr_s32(virr2 * 16383 + 16384, 15);
73 vili0 = math_asr_s32(vili0 * 16383 + 16384, 15);
74 viri0 = math_asr_s32(viri0 * 16383 + 16384, 15);
75 vili1 = math_asr_s32(vili1 * 16383 + 16384, 15);
76 viri1 = math_asr_s32(viri1 * 16383 + 16384, 15);
77 vili2 = math_asr_s32(vili2 * 16383 + 16384, 15);
78 viri2 = math_asr_s32(viri2 * 16383 + 16384, 15);
79 const int32_t vacc1r0 = vilr0 + virr0;
80 const int32_t vacc2r0 = vilr0 - virr0;
81 const int32_t vacc1r1 = vilr1 + virr1;
82 const int32_t vacc2r1 = vilr1 - virr1;
83 const int32_t vacc1r2 = vilr2 + virr2;
84 const int32_t vacc2r2 = vilr2 - virr2;
85 const int32_t vacc1i0 = vili0 + viri0;
86 const int32_t vacc2i0 = vili0 - viri0;
87 const int32_t vacc1i1 = vili1 + viri1;
88 const int32_t vacc2i1 = vili1 - viri1;
89 const int32_t vacc1i2 = vili2 + viri2;
90 const int32_t vacc2i2 = vili2 - viri2;
91
92 const int32_t twr0 = math_asr_s32(vacc2r0 * vtwr0 - vacc2i0 * vtwi0 + 16384, 15);
93 const int32_t twr1 = math_asr_s32(vacc2r1 * vtwr1 - vacc2i1 * vtwi1 + 16384, 15);
94 const int32_t twr2 = math_asr_s32(vacc2r2 * vtwr2 - vacc2i2 * vtwi2 + 16384, 15);
95 const int32_t twi0 = math_asr_s32(vacc2r0 * vtwi0 + vacc2i0 * vtwr0 + 16384, 15);
96 const int32_t twi1 = math_asr_s32(vacc2r1 * vtwi1 + vacc2i1 * vtwr1 + 16384, 15);
97 const int32_t twi2 = math_asr_s32(vacc2r2 * vtwi2 + vacc2i2 * vtwr2 + 16384, 15);
98
99 dl[0] = math_asr_s32(vacc1r0 + twr0, 1);
100 dl[1] = math_asr_s32(vacc1i0 + twi0, 1);
101 dl[2] = math_asr_s32(vacc1r1 + twr1, 1);
102 dl[3] = math_asr_s32(vacc1i1 + twi1, 1);
103 dl[4] = math_asr_s32(vacc1r2 + twr2, 1);
104 dl[5] = math_asr_s32(vacc1i2 + twi2, 1);
105 dr[4] = math_asr_s32(vacc1r0 - twr0, 1);
106 dr[5] = math_asr_s32(twi0 - vacc1i0, 1);
107 dr[2] = math_asr_s32(vacc1r1 - twr1, 1);
108 dr[3] = math_asr_s32(twi1 - vacc1i1, 1);
109 dr[0] = math_asr_s32(vacc1r2 - twr2, 1);
110 dr[1] = math_asr_s32(twi2 - vacc1i2, 1);
111 dl += 3 * 2;
112 }
113
114 if XNN_UNLIKELY(samples != 0) {
115 do {
116 dr -= 2;
117 int32_t vilr = dl[0];
118 int32_t vili = dl[1];
119 int32_t virr = (int32_t) dr[0];
120 int32_t viri = -(int32_t) dr[1];
121 const int32_t vtwr = twiddle[0];
122 const int32_t vtwi = twiddle[1];
123 twiddle += 2;
124
125 vilr = math_asr_s32(vilr * 16383 + 16384, 15);
126 vili = math_asr_s32(vili * 16383 + 16384, 15);
127 virr = math_asr_s32(virr * 16383 + 16384, 15);
128 viri = math_asr_s32(viri * 16383 + 16384, 15);
129 const int32_t vacc1r = vilr + virr;
130 const int32_t vacc1i = vili + viri;
131 const int32_t vacc2r = vilr - virr;
132 const int32_t vacc2i = vili - viri;
133
134 const int32_t twr = math_asr_s32(vacc2r * vtwr - vacc2i * vtwi + 16384, 15);
135 const int32_t twi = math_asr_s32(vacc2r * vtwi + vacc2i * vtwr + 16384, 15);
136
137 dl[0] = math_asr_s32(vacc1r + twr, 1);
138 dl[1] = math_asr_s32(vacc1i + twi, 1);
139 dr[0] = math_asr_s32(vacc1r - twr, 1);
140 dr[1] = math_asr_s32(twi - vacc1i, 1);
141 dl += 2;
142 } while (--samples != 0);
143 }
144 }
145