1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker // Template: src/cs16-bfly4/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_bfly4_ukernel__scalar_x4(size_t samples,int16_t * data,const size_t stride,const int16_t * twiddle)18*4bdc9457SAndroid Build Coastguard Worker void xnn_cs16_bfly4_ukernel__scalar_x4(
19*4bdc9457SAndroid Build Coastguard Worker size_t samples,
20*4bdc9457SAndroid Build Coastguard Worker int16_t* data,
21*4bdc9457SAndroid Build Coastguard Worker const size_t stride,
22*4bdc9457SAndroid Build Coastguard Worker const int16_t* twiddle)
23*4bdc9457SAndroid Build Coastguard Worker {
24*4bdc9457SAndroid Build Coastguard Worker const int16_t* tw1 = twiddle;
25*4bdc9457SAndroid Build Coastguard Worker const int16_t* tw2 = twiddle;
26*4bdc9457SAndroid Build Coastguard Worker const int16_t* tw3 = twiddle;
27*4bdc9457SAndroid Build Coastguard Worker int16_t* data0 = data;
28*4bdc9457SAndroid Build Coastguard Worker int16_t* data1 = data + samples * 2;
29*4bdc9457SAndroid Build Coastguard Worker int16_t* data2 = data + samples * 4;
30*4bdc9457SAndroid Build Coastguard Worker int16_t* data3 = data + samples * 6;
31*4bdc9457SAndroid Build Coastguard Worker
32*4bdc9457SAndroid Build Coastguard Worker assert(samples != 0);
33*4bdc9457SAndroid Build Coastguard Worker assert(data != NULL);
34*4bdc9457SAndroid Build Coastguard Worker assert(stride != 0);
35*4bdc9457SAndroid Build Coastguard Worker assert(twiddle != NULL);
36*4bdc9457SAndroid Build Coastguard Worker
37*4bdc9457SAndroid Build Coastguard Worker for (; samples >= 4; samples -= 4) {
38*4bdc9457SAndroid Build Coastguard Worker int32_t vout0r0 = (int32_t) data0[0];
39*4bdc9457SAndroid Build Coastguard Worker int32_t vout0i0 = (int32_t) data0[1];
40*4bdc9457SAndroid Build Coastguard Worker int32_t vout0r1 = (int32_t) data0[2];
41*4bdc9457SAndroid Build Coastguard Worker int32_t vout0i1 = (int32_t) data0[3];
42*4bdc9457SAndroid Build Coastguard Worker int32_t vout0r2 = (int32_t) data0[4];
43*4bdc9457SAndroid Build Coastguard Worker int32_t vout0i2 = (int32_t) data0[5];
44*4bdc9457SAndroid Build Coastguard Worker int32_t vout0r3 = (int32_t) data0[6];
45*4bdc9457SAndroid Build Coastguard Worker int32_t vout0i3 = (int32_t) data0[7];
46*4bdc9457SAndroid Build Coastguard Worker int32_t vout1r0 = (int32_t) data1[0];
47*4bdc9457SAndroid Build Coastguard Worker int32_t vout1i0 = (int32_t) data1[1];
48*4bdc9457SAndroid Build Coastguard Worker int32_t vout1r1 = (int32_t) data1[2];
49*4bdc9457SAndroid Build Coastguard Worker int32_t vout1i1 = (int32_t) data1[3];
50*4bdc9457SAndroid Build Coastguard Worker int32_t vout1r2 = (int32_t) data1[4];
51*4bdc9457SAndroid Build Coastguard Worker int32_t vout1i2 = (int32_t) data1[5];
52*4bdc9457SAndroid Build Coastguard Worker int32_t vout1r3 = (int32_t) data1[6];
53*4bdc9457SAndroid Build Coastguard Worker int32_t vout1i3 = (int32_t) data1[7];
54*4bdc9457SAndroid Build Coastguard Worker int32_t vout2r0 = (int32_t) data2[0];
55*4bdc9457SAndroid Build Coastguard Worker int32_t vout2i0 = (int32_t) data2[1];
56*4bdc9457SAndroid Build Coastguard Worker int32_t vout2r1 = (int32_t) data2[2];
57*4bdc9457SAndroid Build Coastguard Worker int32_t vout2i1 = (int32_t) data2[3];
58*4bdc9457SAndroid Build Coastguard Worker int32_t vout2r2 = (int32_t) data2[4];
59*4bdc9457SAndroid Build Coastguard Worker int32_t vout2i2 = (int32_t) data2[5];
60*4bdc9457SAndroid Build Coastguard Worker int32_t vout2r3 = (int32_t) data2[6];
61*4bdc9457SAndroid Build Coastguard Worker int32_t vout2i3 = (int32_t) data2[7];
62*4bdc9457SAndroid Build Coastguard Worker int32_t vout3r0 = (int32_t) data3[0];
63*4bdc9457SAndroid Build Coastguard Worker int32_t vout3i0 = (int32_t) data3[1];
64*4bdc9457SAndroid Build Coastguard Worker int32_t vout3r1 = (int32_t) data3[2];
65*4bdc9457SAndroid Build Coastguard Worker int32_t vout3i1 = (int32_t) data3[3];
66*4bdc9457SAndroid Build Coastguard Worker int32_t vout3r2 = (int32_t) data3[4];
67*4bdc9457SAndroid Build Coastguard Worker int32_t vout3i2 = (int32_t) data3[5];
68*4bdc9457SAndroid Build Coastguard Worker int32_t vout3r3 = (int32_t) data3[6];
69*4bdc9457SAndroid Build Coastguard Worker int32_t vout3i3 = (int32_t) data3[7];
70*4bdc9457SAndroid Build Coastguard Worker
71*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1r0 = (const int32_t) tw1[0];
72*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1i0 = (const int32_t) tw1[1];
73*4bdc9457SAndroid Build Coastguard Worker tw1 += stride * 2;
74*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1r1 = (const int32_t) tw1[0];
75*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1i1 = (const int32_t) tw1[1];
76*4bdc9457SAndroid Build Coastguard Worker tw1 += stride * 2;
77*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1r2 = (const int32_t) tw1[0];
78*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1i2 = (const int32_t) tw1[1];
79*4bdc9457SAndroid Build Coastguard Worker tw1 += stride * 2;
80*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1r3 = (const int32_t) tw1[0];
81*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1i3 = (const int32_t) tw1[1];
82*4bdc9457SAndroid Build Coastguard Worker tw1 += stride * 2;
83*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2r0 = (const int32_t) tw2[0];
84*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2i0 = (const int32_t) tw2[1];
85*4bdc9457SAndroid Build Coastguard Worker tw2 += stride * 4;
86*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2r1 = (const int32_t) tw2[0];
87*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2i1 = (const int32_t) tw2[1];
88*4bdc9457SAndroid Build Coastguard Worker tw2 += stride * 4;
89*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2r2 = (const int32_t) tw2[0];
90*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2i2 = (const int32_t) tw2[1];
91*4bdc9457SAndroid Build Coastguard Worker tw2 += stride * 4;
92*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2r3 = (const int32_t) tw2[0];
93*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2i3 = (const int32_t) tw2[1];
94*4bdc9457SAndroid Build Coastguard Worker tw2 += stride * 4;
95*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3r0 = (const int32_t) tw3[0];
96*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3i0 = (const int32_t) tw3[1];
97*4bdc9457SAndroid Build Coastguard Worker tw3 += stride * 6;
98*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3r1 = (const int32_t) tw3[0];
99*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3i1 = (const int32_t) tw3[1];
100*4bdc9457SAndroid Build Coastguard Worker tw3 += stride * 6;
101*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3r2 = (const int32_t) tw3[0];
102*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3i2 = (const int32_t) tw3[1];
103*4bdc9457SAndroid Build Coastguard Worker tw3 += stride * 6;
104*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3r3 = (const int32_t) tw3[0];
105*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3i3 = (const int32_t) tw3[1];
106*4bdc9457SAndroid Build Coastguard Worker tw3 += stride * 6;
107*4bdc9457SAndroid Build Coastguard Worker
108*4bdc9457SAndroid Build Coastguard Worker // Note 32767 / 4 = 8191. Should be 8192.
109*4bdc9457SAndroid Build Coastguard Worker vout0r0 = math_asr_s32(vout0r0 * 8191 + 16384, 15);
110*4bdc9457SAndroid Build Coastguard Worker vout0r1 = math_asr_s32(vout0r1 * 8191 + 16384, 15);
111*4bdc9457SAndroid Build Coastguard Worker vout0r2 = math_asr_s32(vout0r2 * 8191 + 16384, 15);
112*4bdc9457SAndroid Build Coastguard Worker vout0r3 = math_asr_s32(vout0r3 * 8191 + 16384, 15);
113*4bdc9457SAndroid Build Coastguard Worker vout0i0 = math_asr_s32(vout0i0 * 8191 + 16384, 15);
114*4bdc9457SAndroid Build Coastguard Worker vout0i1 = math_asr_s32(vout0i1 * 8191 + 16384, 15);
115*4bdc9457SAndroid Build Coastguard Worker vout0i2 = math_asr_s32(vout0i2 * 8191 + 16384, 15);
116*4bdc9457SAndroid Build Coastguard Worker vout0i3 = math_asr_s32(vout0i3 * 8191 + 16384, 15);
117*4bdc9457SAndroid Build Coastguard Worker vout1r0 = math_asr_s32(vout1r0 * 8191 + 16384, 15);
118*4bdc9457SAndroid Build Coastguard Worker vout1r1 = math_asr_s32(vout1r1 * 8191 + 16384, 15);
119*4bdc9457SAndroid Build Coastguard Worker vout1r2 = math_asr_s32(vout1r2 * 8191 + 16384, 15);
120*4bdc9457SAndroid Build Coastguard Worker vout1r3 = math_asr_s32(vout1r3 * 8191 + 16384, 15);
121*4bdc9457SAndroid Build Coastguard Worker vout1i0 = math_asr_s32(vout1i0 * 8191 + 16384, 15);
122*4bdc9457SAndroid Build Coastguard Worker vout1i1 = math_asr_s32(vout1i1 * 8191 + 16384, 15);
123*4bdc9457SAndroid Build Coastguard Worker vout1i2 = math_asr_s32(vout1i2 * 8191 + 16384, 15);
124*4bdc9457SAndroid Build Coastguard Worker vout1i3 = math_asr_s32(vout1i3 * 8191 + 16384, 15);
125*4bdc9457SAndroid Build Coastguard Worker vout2r0 = math_asr_s32(vout2r0 * 8191 + 16384, 15);
126*4bdc9457SAndroid Build Coastguard Worker vout2r1 = math_asr_s32(vout2r1 * 8191 + 16384, 15);
127*4bdc9457SAndroid Build Coastguard Worker vout2r2 = math_asr_s32(vout2r2 * 8191 + 16384, 15);
128*4bdc9457SAndroid Build Coastguard Worker vout2r3 = math_asr_s32(vout2r3 * 8191 + 16384, 15);
129*4bdc9457SAndroid Build Coastguard Worker vout2i0 = math_asr_s32(vout2i0 * 8191 + 16384, 15);
130*4bdc9457SAndroid Build Coastguard Worker vout2i1 = math_asr_s32(vout2i1 * 8191 + 16384, 15);
131*4bdc9457SAndroid Build Coastguard Worker vout2i2 = math_asr_s32(vout2i2 * 8191 + 16384, 15);
132*4bdc9457SAndroid Build Coastguard Worker vout2i3 = math_asr_s32(vout2i3 * 8191 + 16384, 15);
133*4bdc9457SAndroid Build Coastguard Worker vout3r0 = math_asr_s32(vout3r0 * 8191 + 16384, 15);
134*4bdc9457SAndroid Build Coastguard Worker vout3r1 = math_asr_s32(vout3r1 * 8191 + 16384, 15);
135*4bdc9457SAndroid Build Coastguard Worker vout3r2 = math_asr_s32(vout3r2 * 8191 + 16384, 15);
136*4bdc9457SAndroid Build Coastguard Worker vout3r3 = math_asr_s32(vout3r3 * 8191 + 16384, 15);
137*4bdc9457SAndroid Build Coastguard Worker vout3i0 = math_asr_s32(vout3i0 * 8191 + 16384, 15);
138*4bdc9457SAndroid Build Coastguard Worker vout3i1 = math_asr_s32(vout3i1 * 8191 + 16384, 15);
139*4bdc9457SAndroid Build Coastguard Worker vout3i2 = math_asr_s32(vout3i2 * 8191 + 16384, 15);
140*4bdc9457SAndroid Build Coastguard Worker vout3i3 = math_asr_s32(vout3i3 * 8191 + 16384, 15);
141*4bdc9457SAndroid Build Coastguard Worker
142*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0r0 = math_asr_s32(vout1r0 * vtw1r0 - vout1i0 * vtw1i0 + 16384, 15);
143*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0r1 = math_asr_s32(vout1r1 * vtw1r1 - vout1i1 * vtw1i1 + 16384, 15);
144*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0r2 = math_asr_s32(vout1r2 * vtw1r2 - vout1i2 * vtw1i2 + 16384, 15);
145*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0r3 = math_asr_s32(vout1r3 * vtw1r3 - vout1i3 * vtw1i3 + 16384, 15);
146*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0i0 = math_asr_s32(vout1r0 * vtw1i0 + vout1i0 * vtw1r0 + 16384, 15);
147*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0i1 = math_asr_s32(vout1r1 * vtw1i1 + vout1i1 * vtw1r1 + 16384, 15);
148*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0i2 = math_asr_s32(vout1r2 * vtw1i2 + vout1i2 * vtw1r2 + 16384, 15);
149*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0i3 = math_asr_s32(vout1r3 * vtw1i3 + vout1i3 * vtw1r3 + 16384, 15);
150*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1r0 = math_asr_s32(vout2r0 * vtw2r0 - vout2i0 * vtw2i0 + 16384, 15);
151*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1r1 = math_asr_s32(vout2r1 * vtw2r1 - vout2i1 * vtw2i1 + 16384, 15);
152*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1r2 = math_asr_s32(vout2r2 * vtw2r2 - vout2i2 * vtw2i2 + 16384, 15);
153*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1r3 = math_asr_s32(vout2r3 * vtw2r3 - vout2i3 * vtw2i3 + 16384, 15);
154*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1i0 = math_asr_s32(vout2r0 * vtw2i0 + vout2i0 * vtw2r0 + 16384, 15);
155*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1i1 = math_asr_s32(vout2r1 * vtw2i1 + vout2i1 * vtw2r1 + 16384, 15);
156*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1i2 = math_asr_s32(vout2r2 * vtw2i2 + vout2i2 * vtw2r2 + 16384, 15);
157*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1i3 = math_asr_s32(vout2r3 * vtw2i3 + vout2i3 * vtw2r3 + 16384, 15);
158*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2r0 = math_asr_s32(vout3r0 * vtw3r0 - vout3i0 * vtw3i0 + 16384, 15);
159*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2r1 = math_asr_s32(vout3r1 * vtw3r1 - vout3i1 * vtw3i1 + 16384, 15);
160*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2r2 = math_asr_s32(vout3r2 * vtw3r2 - vout3i2 * vtw3i2 + 16384, 15);
161*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2r3 = math_asr_s32(vout3r3 * vtw3r3 - vout3i3 * vtw3i3 + 16384, 15);
162*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2i0 = math_asr_s32(vout3r0 * vtw3i0 + vout3i0 * vtw3r0 + 16384, 15);
163*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2i1 = math_asr_s32(vout3r1 * vtw3i1 + vout3i1 * vtw3r1 + 16384, 15);
164*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2i2 = math_asr_s32(vout3r2 * vtw3i2 + vout3i2 * vtw3r2 + 16384, 15);
165*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2i3 = math_asr_s32(vout3r3 * vtw3i3 + vout3i3 * vtw3r3 + 16384, 15);
166*4bdc9457SAndroid Build Coastguard Worker
167*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5r0 = vout0r0 - vtmp1r0;
168*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5r1 = vout0r1 - vtmp1r1;
169*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5r2 = vout0r2 - vtmp1r2;
170*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5r3 = vout0r3 - vtmp1r3;
171*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5i0 = vout0i0 - vtmp1i0;
172*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5i1 = vout0i1 - vtmp1i1;
173*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5i2 = vout0i2 - vtmp1i2;
174*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5i3 = vout0i3 - vtmp1i3;
175*4bdc9457SAndroid Build Coastguard Worker vout0r0 += vtmp1r0;
176*4bdc9457SAndroid Build Coastguard Worker vout0r1 += vtmp1r1;
177*4bdc9457SAndroid Build Coastguard Worker vout0r2 += vtmp1r2;
178*4bdc9457SAndroid Build Coastguard Worker vout0r3 += vtmp1r3;
179*4bdc9457SAndroid Build Coastguard Worker vout0i0 += vtmp1i0;
180*4bdc9457SAndroid Build Coastguard Worker vout0i1 += vtmp1i1;
181*4bdc9457SAndroid Build Coastguard Worker vout0i2 += vtmp1i2;
182*4bdc9457SAndroid Build Coastguard Worker vout0i3 += vtmp1i3;
183*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3r0 = vtmp0r0 + vtmp2r0;
184*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3r1 = vtmp0r1 + vtmp2r1;
185*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3r2 = vtmp0r2 + vtmp2r2;
186*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3r3 = vtmp0r3 + vtmp2r3;
187*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3i0 = vtmp0i0 + vtmp2i0;
188*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3i1 = vtmp0i1 + vtmp2i1;
189*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3i2 = vtmp0i2 + vtmp2i2;
190*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3i3 = vtmp0i3 + vtmp2i3;
191*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4r0 = vtmp0r0 - vtmp2r0;
192*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4r1 = vtmp0r1 - vtmp2r1;
193*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4r2 = vtmp0r2 - vtmp2r2;
194*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4r3 = vtmp0r3 - vtmp2r3;
195*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4i0 = vtmp0i0 - vtmp2i0;
196*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4i1 = vtmp0i1 - vtmp2i1;
197*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4i2 = vtmp0i2 - vtmp2i2;
198*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4i3 = vtmp0i3 - vtmp2i3;
199*4bdc9457SAndroid Build Coastguard Worker vout2r0 = vout0r0 - vtmp3r0;
200*4bdc9457SAndroid Build Coastguard Worker vout2r1 = vout0r1 - vtmp3r1;
201*4bdc9457SAndroid Build Coastguard Worker vout2r2 = vout0r2 - vtmp3r2;
202*4bdc9457SAndroid Build Coastguard Worker vout2r3 = vout0r3 - vtmp3r3;
203*4bdc9457SAndroid Build Coastguard Worker vout2i0 = vout0i0 - vtmp3i0;
204*4bdc9457SAndroid Build Coastguard Worker vout2i1 = vout0i1 - vtmp3i1;
205*4bdc9457SAndroid Build Coastguard Worker vout2i2 = vout0i2 - vtmp3i2;
206*4bdc9457SAndroid Build Coastguard Worker vout2i3 = vout0i3 - vtmp3i3;
207*4bdc9457SAndroid Build Coastguard Worker vout0r0 += vtmp3r0;
208*4bdc9457SAndroid Build Coastguard Worker vout0r1 += vtmp3r1;
209*4bdc9457SAndroid Build Coastguard Worker vout0r2 += vtmp3r2;
210*4bdc9457SAndroid Build Coastguard Worker vout0r3 += vtmp3r3;
211*4bdc9457SAndroid Build Coastguard Worker vout0i0 += vtmp3i0;
212*4bdc9457SAndroid Build Coastguard Worker vout0i1 += vtmp3i1;
213*4bdc9457SAndroid Build Coastguard Worker vout0i2 += vtmp3i2;
214*4bdc9457SAndroid Build Coastguard Worker vout0i3 += vtmp3i3;
215*4bdc9457SAndroid Build Coastguard Worker vout1r0 = vtmp5r0 + vtmp4i0;
216*4bdc9457SAndroid Build Coastguard Worker vout1r1 = vtmp5r1 + vtmp4i1;
217*4bdc9457SAndroid Build Coastguard Worker vout1r2 = vtmp5r2 + vtmp4i2;
218*4bdc9457SAndroid Build Coastguard Worker vout1r3 = vtmp5r3 + vtmp4i3;
219*4bdc9457SAndroid Build Coastguard Worker vout1i0 = vtmp5i0 - vtmp4r0;
220*4bdc9457SAndroid Build Coastguard Worker vout1i1 = vtmp5i1 - vtmp4r1;
221*4bdc9457SAndroid Build Coastguard Worker vout1i2 = vtmp5i2 - vtmp4r2;
222*4bdc9457SAndroid Build Coastguard Worker vout1i3 = vtmp5i3 - vtmp4r3;
223*4bdc9457SAndroid Build Coastguard Worker vout3r0 = vtmp5r0 - vtmp4i0;
224*4bdc9457SAndroid Build Coastguard Worker vout3r1 = vtmp5r1 - vtmp4i1;
225*4bdc9457SAndroid Build Coastguard Worker vout3r2 = vtmp5r2 - vtmp4i2;
226*4bdc9457SAndroid Build Coastguard Worker vout3r3 = vtmp5r3 - vtmp4i3;
227*4bdc9457SAndroid Build Coastguard Worker vout3i0 = vtmp5i0 + vtmp4r0;
228*4bdc9457SAndroid Build Coastguard Worker vout3i1 = vtmp5i1 + vtmp4r1;
229*4bdc9457SAndroid Build Coastguard Worker vout3i2 = vtmp5i2 + vtmp4r2;
230*4bdc9457SAndroid Build Coastguard Worker vout3i3 = vtmp5i3 + vtmp4r3;
231*4bdc9457SAndroid Build Coastguard Worker
232*4bdc9457SAndroid Build Coastguard Worker data0[0] = (int16_t) vout0r0;
233*4bdc9457SAndroid Build Coastguard Worker data0[1] = (int16_t) vout0i0;
234*4bdc9457SAndroid Build Coastguard Worker data0[2] = (int16_t) vout0r1;
235*4bdc9457SAndroid Build Coastguard Worker data0[3] = (int16_t) vout0i1;
236*4bdc9457SAndroid Build Coastguard Worker data0[4] = (int16_t) vout0r2;
237*4bdc9457SAndroid Build Coastguard Worker data0[5] = (int16_t) vout0i2;
238*4bdc9457SAndroid Build Coastguard Worker data0[6] = (int16_t) vout0r3;
239*4bdc9457SAndroid Build Coastguard Worker data0[7] = (int16_t) vout0i3;
240*4bdc9457SAndroid Build Coastguard Worker data0 += 4 * 2;
241*4bdc9457SAndroid Build Coastguard Worker data1[0] = (int16_t) vout1r0;
242*4bdc9457SAndroid Build Coastguard Worker data1[1] = (int16_t) vout1i0;
243*4bdc9457SAndroid Build Coastguard Worker data1[2] = (int16_t) vout1r1;
244*4bdc9457SAndroid Build Coastguard Worker data1[3] = (int16_t) vout1i1;
245*4bdc9457SAndroid Build Coastguard Worker data1[4] = (int16_t) vout1r2;
246*4bdc9457SAndroid Build Coastguard Worker data1[5] = (int16_t) vout1i2;
247*4bdc9457SAndroid Build Coastguard Worker data1[6] = (int16_t) vout1r3;
248*4bdc9457SAndroid Build Coastguard Worker data1[7] = (int16_t) vout1i3;
249*4bdc9457SAndroid Build Coastguard Worker data1 += 4 * 2;
250*4bdc9457SAndroid Build Coastguard Worker data2[0] = (int16_t) vout2r0;
251*4bdc9457SAndroid Build Coastguard Worker data2[1] = (int16_t) vout2i0;
252*4bdc9457SAndroid Build Coastguard Worker data2[2] = (int16_t) vout2r1;
253*4bdc9457SAndroid Build Coastguard Worker data2[3] = (int16_t) vout2i1;
254*4bdc9457SAndroid Build Coastguard Worker data2[4] = (int16_t) vout2r2;
255*4bdc9457SAndroid Build Coastguard Worker data2[5] = (int16_t) vout2i2;
256*4bdc9457SAndroid Build Coastguard Worker data2[6] = (int16_t) vout2r3;
257*4bdc9457SAndroid Build Coastguard Worker data2[7] = (int16_t) vout2i3;
258*4bdc9457SAndroid Build Coastguard Worker data2 += 4 * 2;
259*4bdc9457SAndroid Build Coastguard Worker data3[0] = (int16_t) vout3r0;
260*4bdc9457SAndroid Build Coastguard Worker data3[1] = (int16_t) vout3i0;
261*4bdc9457SAndroid Build Coastguard Worker data3[2] = (int16_t) vout3r1;
262*4bdc9457SAndroid Build Coastguard Worker data3[3] = (int16_t) vout3i1;
263*4bdc9457SAndroid Build Coastguard Worker data3[4] = (int16_t) vout3r2;
264*4bdc9457SAndroid Build Coastguard Worker data3[5] = (int16_t) vout3i2;
265*4bdc9457SAndroid Build Coastguard Worker data3[6] = (int16_t) vout3r3;
266*4bdc9457SAndroid Build Coastguard Worker data3[7] = (int16_t) vout3i3;
267*4bdc9457SAndroid Build Coastguard Worker data3 += 4 * 2;
268*4bdc9457SAndroid Build Coastguard Worker }
269*4bdc9457SAndroid Build Coastguard Worker
270*4bdc9457SAndroid Build Coastguard Worker if XNN_UNLIKELY(samples != 0) {
271*4bdc9457SAndroid Build Coastguard Worker do {
272*4bdc9457SAndroid Build Coastguard Worker int32_t vout0r = (int32_t) data0[0];
273*4bdc9457SAndroid Build Coastguard Worker int32_t vout0i = (int32_t) data0[1];
274*4bdc9457SAndroid Build Coastguard Worker int32_t vout1r = (int32_t) data1[0];
275*4bdc9457SAndroid Build Coastguard Worker int32_t vout1i = (int32_t) data1[1];
276*4bdc9457SAndroid Build Coastguard Worker int32_t vout2r = (int32_t) data2[0];
277*4bdc9457SAndroid Build Coastguard Worker int32_t vout2i = (int32_t) data2[1];
278*4bdc9457SAndroid Build Coastguard Worker int32_t vout3r = (int32_t) data3[0];
279*4bdc9457SAndroid Build Coastguard Worker int32_t vout3i = (int32_t) data3[1];
280*4bdc9457SAndroid Build Coastguard Worker
281*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1r = (const int32_t) tw1[0];
282*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw1i = (const int32_t) tw1[1];
283*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2r = (const int32_t) tw2[0];
284*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw2i = (const int32_t) tw2[1];
285*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3r = (const int32_t) tw3[0];
286*4bdc9457SAndroid Build Coastguard Worker const int32_t vtw3i = (const int32_t) tw3[1];
287*4bdc9457SAndroid Build Coastguard Worker tw1 += stride * 2;
288*4bdc9457SAndroid Build Coastguard Worker tw2 += stride * 4;
289*4bdc9457SAndroid Build Coastguard Worker tw3 += stride * 6;
290*4bdc9457SAndroid Build Coastguard Worker
291*4bdc9457SAndroid Build Coastguard Worker // Note 32767 / 4 = 8191. Should be 8192.
292*4bdc9457SAndroid Build Coastguard Worker vout0r = math_asr_s32(vout0r * 8191 + 16384, 15);
293*4bdc9457SAndroid Build Coastguard Worker vout0i = math_asr_s32(vout0i * 8191 + 16384, 15);
294*4bdc9457SAndroid Build Coastguard Worker vout1r = math_asr_s32(vout1r * 8191 + 16384, 15);
295*4bdc9457SAndroid Build Coastguard Worker vout1i = math_asr_s32(vout1i * 8191 + 16384, 15);
296*4bdc9457SAndroid Build Coastguard Worker vout2r = math_asr_s32(vout2r * 8191 + 16384, 15);
297*4bdc9457SAndroid Build Coastguard Worker vout2i = math_asr_s32(vout2i * 8191 + 16384, 15);
298*4bdc9457SAndroid Build Coastguard Worker vout3r = math_asr_s32(vout3r * 8191 + 16384, 15);
299*4bdc9457SAndroid Build Coastguard Worker vout3i = math_asr_s32(vout3i * 8191 + 16384, 15);
300*4bdc9457SAndroid Build Coastguard Worker
301*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0r = math_asr_s32(vout1r * vtw1r - vout1i * vtw1i + 16384, 15);
302*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp0i = math_asr_s32(vout1r * vtw1i + vout1i * vtw1r + 16384, 15);
303*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1r = math_asr_s32(vout2r * vtw2r - vout2i * vtw2i + 16384, 15);
304*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp1i = math_asr_s32(vout2r * vtw2i + vout2i * vtw2r + 16384, 15);
305*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2r = math_asr_s32(vout3r * vtw3r - vout3i * vtw3i + 16384, 15);
306*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp2i = math_asr_s32(vout3r * vtw3i + vout3i * vtw3r + 16384, 15);
307*4bdc9457SAndroid Build Coastguard Worker
308*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5r = vout0r - vtmp1r;
309*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp5i = vout0i - vtmp1i;
310*4bdc9457SAndroid Build Coastguard Worker vout0r += vtmp1r;
311*4bdc9457SAndroid Build Coastguard Worker vout0i += vtmp1i;
312*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3r = vtmp0r + vtmp2r;
313*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp3i = vtmp0i + vtmp2i;
314*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4r = vtmp0r - vtmp2r;
315*4bdc9457SAndroid Build Coastguard Worker const int32_t vtmp4i = vtmp0i - vtmp2i;
316*4bdc9457SAndroid Build Coastguard Worker vout2r = vout0r - vtmp3r;
317*4bdc9457SAndroid Build Coastguard Worker vout2i = vout0i - vtmp3i;
318*4bdc9457SAndroid Build Coastguard Worker
319*4bdc9457SAndroid Build Coastguard Worker vout0r += vtmp3r;
320*4bdc9457SAndroid Build Coastguard Worker vout0i += vtmp3i;
321*4bdc9457SAndroid Build Coastguard Worker
322*4bdc9457SAndroid Build Coastguard Worker vout1r = vtmp5r + vtmp4i;
323*4bdc9457SAndroid Build Coastguard Worker vout1i = vtmp5i - vtmp4r;
324*4bdc9457SAndroid Build Coastguard Worker vout3r = vtmp5r - vtmp4i;
325*4bdc9457SAndroid Build Coastguard Worker vout3i = vtmp5i + vtmp4r;
326*4bdc9457SAndroid Build Coastguard Worker
327*4bdc9457SAndroid Build Coastguard Worker data0[0] = (int16_t) vout0r;
328*4bdc9457SAndroid Build Coastguard Worker data0[1] = (int16_t) vout0i;
329*4bdc9457SAndroid Build Coastguard Worker data1[0] = (int16_t) vout1r;
330*4bdc9457SAndroid Build Coastguard Worker data1[1] = (int16_t) vout1i;
331*4bdc9457SAndroid Build Coastguard Worker data2[0] = (int16_t) vout2r;
332*4bdc9457SAndroid Build Coastguard Worker data2[1] = (int16_t) vout2i;
333*4bdc9457SAndroid Build Coastguard Worker data3[0] = (int16_t) vout3r;
334*4bdc9457SAndroid Build Coastguard Worker data3[1] = (int16_t) vout3i;
335*4bdc9457SAndroid Build Coastguard Worker data0 += 2;
336*4bdc9457SAndroid Build Coastguard Worker data1 += 2;
337*4bdc9457SAndroid Build Coastguard Worker data2 += 2;
338*4bdc9457SAndroid Build Coastguard Worker data3 += 2;
339*4bdc9457SAndroid Build Coastguard Worker } while(--samples != 0);
340*4bdc9457SAndroid Build Coastguard Worker }
341*4bdc9457SAndroid Build Coastguard Worker }
342