xref: /aosp_15_r20/external/XNNPACK/src/cs16-fftr/gen/scalar-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/cs16-fftr/scalar.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2022 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <stddef.h>
12*4bdc9457SAndroid Build Coastguard Worker #include <stdint.h>
13*4bdc9457SAndroid Build Coastguard Worker 
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/fft.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_cs16_fftr_ukernel__scalar_x2(size_t samples,int16_t * data,const int16_t * twiddle)18*4bdc9457SAndroid Build Coastguard Worker void xnn_cs16_fftr_ukernel__scalar_x2(
19*4bdc9457SAndroid Build Coastguard Worker     size_t samples,
20*4bdc9457SAndroid Build Coastguard Worker     int16_t* data,
21*4bdc9457SAndroid Build Coastguard Worker     const int16_t* twiddle)
22*4bdc9457SAndroid Build Coastguard Worker {
23*4bdc9457SAndroid Build Coastguard Worker   assert(samples >= 2);
24*4bdc9457SAndroid Build Coastguard Worker   assert(samples % 2 == 0);
25*4bdc9457SAndroid Build Coastguard Worker   assert(data != NULL);
26*4bdc9457SAndroid Build Coastguard Worker   assert(data != NULL);
27*4bdc9457SAndroid Build Coastguard Worker   assert(twiddle != NULL);
28*4bdc9457SAndroid Build Coastguard Worker 
29*4bdc9457SAndroid Build Coastguard Worker   int16_t* dl = data;
30*4bdc9457SAndroid Build Coastguard Worker   int16_t* dr = data + samples * 2;
31*4bdc9457SAndroid Build Coastguard Worker   int32_t vdcr = (int32_t) dl[0];
32*4bdc9457SAndroid Build Coastguard Worker   int32_t vdci = (int32_t) dl[1];
33*4bdc9457SAndroid Build Coastguard Worker 
34*4bdc9457SAndroid Build Coastguard Worker   vdcr = math_asr_s32(vdcr * 16383 + 16384, 15);
35*4bdc9457SAndroid Build Coastguard Worker   vdci = math_asr_s32(vdci * 16383 + 16384, 15);
36*4bdc9457SAndroid Build Coastguard Worker 
37*4bdc9457SAndroid Build Coastguard Worker   dl[0] = vdcr + vdci;
38*4bdc9457SAndroid Build Coastguard Worker   dl[1] = 0;
39*4bdc9457SAndroid Build Coastguard Worker   dl += 2;
40*4bdc9457SAndroid Build Coastguard Worker   dr[0] = vdcr - vdci;
41*4bdc9457SAndroid Build Coastguard Worker   dr[1] = 0;
42*4bdc9457SAndroid Build Coastguard Worker 
43*4bdc9457SAndroid Build Coastguard Worker   samples >>= 1;
44*4bdc9457SAndroid Build Coastguard Worker 
45*4bdc9457SAndroid Build Coastguard Worker   for (; samples >= 2; samples -= 2) {
46*4bdc9457SAndroid Build Coastguard Worker     dr -= 2 * 2;
47*4bdc9457SAndroid Build Coastguard Worker     int32_t vilr0 = dl[0];
48*4bdc9457SAndroid Build Coastguard Worker     int32_t vili0 = dl[1];
49*4bdc9457SAndroid Build Coastguard Worker     int32_t vilr1 = dl[2];
50*4bdc9457SAndroid Build Coastguard Worker     int32_t vili1 = dl[3];
51*4bdc9457SAndroid Build Coastguard Worker     int32_t virr0 =  (int32_t) dr[2];
52*4bdc9457SAndroid Build Coastguard Worker     int32_t viri0 = -(int32_t) dr[3];
53*4bdc9457SAndroid Build Coastguard Worker     int32_t virr1 =  (int32_t) dr[0];
54*4bdc9457SAndroid Build Coastguard Worker     int32_t viri1 = -(int32_t) dr[1];
55*4bdc9457SAndroid Build Coastguard Worker     const int32_t vtwr0 = twiddle[0];
56*4bdc9457SAndroid Build Coastguard Worker     const int32_t vtwi0 = twiddle[1];
57*4bdc9457SAndroid Build Coastguard Worker     const int32_t vtwr1 = twiddle[2];
58*4bdc9457SAndroid Build Coastguard Worker     const int32_t vtwi1 = twiddle[3];
59*4bdc9457SAndroid Build Coastguard Worker     twiddle += 2 * 2;
60*4bdc9457SAndroid Build Coastguard Worker 
61*4bdc9457SAndroid Build Coastguard Worker     vilr0 = math_asr_s32(vilr0 * 16383 + 16384, 15);
62*4bdc9457SAndroid Build Coastguard Worker     virr0 = math_asr_s32(virr0 * 16383 + 16384, 15);
63*4bdc9457SAndroid Build Coastguard Worker     vilr1 = math_asr_s32(vilr1 * 16383 + 16384, 15);
64*4bdc9457SAndroid Build Coastguard Worker     virr1 = math_asr_s32(virr1 * 16383 + 16384, 15);
65*4bdc9457SAndroid Build Coastguard Worker     vili0 = math_asr_s32(vili0 * 16383 + 16384, 15);
66*4bdc9457SAndroid Build Coastguard Worker     viri0 = math_asr_s32(viri0 * 16383 + 16384, 15);
67*4bdc9457SAndroid Build Coastguard Worker     vili1 = math_asr_s32(vili1 * 16383 + 16384, 15);
68*4bdc9457SAndroid Build Coastguard Worker     viri1 = math_asr_s32(viri1 * 16383 + 16384, 15);
69*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc1r0 = vilr0 + virr0;
70*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc2r0 = vilr0 - virr0;
71*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc1r1 = vilr1 + virr1;
72*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc2r1 = vilr1 - virr1;
73*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc1i0 = vili0 + viri0;
74*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc2i0 = vili0 - viri0;
75*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc1i1 = vili1 + viri1;
76*4bdc9457SAndroid Build Coastguard Worker     const int32_t vacc2i1 = vili1 - viri1;
77*4bdc9457SAndroid Build Coastguard Worker 
78*4bdc9457SAndroid Build Coastguard Worker     const int32_t twr0 = math_asr_s32(vacc2r0 * vtwr0 - vacc2i0 * vtwi0 + 16384, 15);
79*4bdc9457SAndroid Build Coastguard Worker     const int32_t twr1 = math_asr_s32(vacc2r1 * vtwr1 - vacc2i1 * vtwi1 + 16384, 15);
80*4bdc9457SAndroid Build Coastguard Worker     const int32_t twi0 = math_asr_s32(vacc2r0 * vtwi0 + vacc2i0 * vtwr0 + 16384, 15);
81*4bdc9457SAndroid Build Coastguard Worker     const int32_t twi1 = math_asr_s32(vacc2r1 * vtwi1 + vacc2i1 * vtwr1 + 16384, 15);
82*4bdc9457SAndroid Build Coastguard Worker 
83*4bdc9457SAndroid Build Coastguard Worker     dl[0] = math_asr_s32(vacc1r0 + twr0, 1);
84*4bdc9457SAndroid Build Coastguard Worker     dl[1] = math_asr_s32(vacc1i0 + twi0, 1);
85*4bdc9457SAndroid Build Coastguard Worker     dl[2] = math_asr_s32(vacc1r1 + twr1, 1);
86*4bdc9457SAndroid Build Coastguard Worker     dl[3] = math_asr_s32(vacc1i1 + twi1, 1);
87*4bdc9457SAndroid Build Coastguard Worker     dr[2] = math_asr_s32(vacc1r0 - twr0, 1);
88*4bdc9457SAndroid Build Coastguard Worker     dr[3] = math_asr_s32(twi0 - vacc1i0, 1);
89*4bdc9457SAndroid Build Coastguard Worker     dr[0] = math_asr_s32(vacc1r1 - twr1, 1);
90*4bdc9457SAndroid Build Coastguard Worker     dr[1] = math_asr_s32(twi1 - vacc1i1, 1);
91*4bdc9457SAndroid Build Coastguard Worker     dl += 2 * 2;
92*4bdc9457SAndroid Build Coastguard Worker   }
93*4bdc9457SAndroid Build Coastguard Worker 
94*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(samples != 0) {
95*4bdc9457SAndroid Build Coastguard Worker     do {
96*4bdc9457SAndroid Build Coastguard Worker       dr -= 2;
97*4bdc9457SAndroid Build Coastguard Worker       int32_t vilr = dl[0];
98*4bdc9457SAndroid Build Coastguard Worker       int32_t vili = dl[1];
99*4bdc9457SAndroid Build Coastguard Worker       int32_t virr =  (int32_t) dr[0];
100*4bdc9457SAndroid Build Coastguard Worker       int32_t viri = -(int32_t) dr[1];
101*4bdc9457SAndroid Build Coastguard Worker       const int32_t vtwr = twiddle[0];
102*4bdc9457SAndroid Build Coastguard Worker       const int32_t vtwi = twiddle[1];
103*4bdc9457SAndroid Build Coastguard Worker       twiddle += 2;
104*4bdc9457SAndroid Build Coastguard Worker 
105*4bdc9457SAndroid Build Coastguard Worker       vilr =  math_asr_s32(vilr * 16383 + 16384, 15);
106*4bdc9457SAndroid Build Coastguard Worker       vili =  math_asr_s32(vili * 16383 + 16384, 15);
107*4bdc9457SAndroid Build Coastguard Worker       virr = math_asr_s32(virr * 16383 + 16384, 15);
108*4bdc9457SAndroid Build Coastguard Worker       viri = math_asr_s32(viri * 16383 + 16384, 15);
109*4bdc9457SAndroid Build Coastguard Worker       const int32_t vacc1r = vilr + virr;
110*4bdc9457SAndroid Build Coastguard Worker       const int32_t vacc1i = vili + viri;
111*4bdc9457SAndroid Build Coastguard Worker       const int32_t vacc2r = vilr - virr;
112*4bdc9457SAndroid Build Coastguard Worker       const int32_t vacc2i = vili - viri;
113*4bdc9457SAndroid Build Coastguard Worker 
114*4bdc9457SAndroid Build Coastguard Worker       const int32_t twr = math_asr_s32(vacc2r * vtwr - vacc2i * vtwi + 16384, 15);
115*4bdc9457SAndroid Build Coastguard Worker       const int32_t twi = math_asr_s32(vacc2r * vtwi + vacc2i * vtwr + 16384, 15);
116*4bdc9457SAndroid Build Coastguard Worker 
117*4bdc9457SAndroid Build Coastguard Worker       dl[0] = math_asr_s32(vacc1r + twr, 1);
118*4bdc9457SAndroid Build Coastguard Worker       dl[1] = math_asr_s32(vacc1i + twi, 1);
119*4bdc9457SAndroid Build Coastguard Worker       dr[0] = math_asr_s32(vacc1r - twr, 1);
120*4bdc9457SAndroid Build Coastguard Worker       dr[1] = math_asr_s32(twi - vacc1i, 1);
121*4bdc9457SAndroid Build Coastguard Worker       dl += 2;
122*4bdc9457SAndroid Build Coastguard Worker     } while (--samples != 0);
123*4bdc9457SAndroid Build Coastguard Worker   }
124*4bdc9457SAndroid Build Coastguard Worker }
125