xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x20.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/neon-p5.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__neon_rr2_p5_nr2recps_x20(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24   assert(n % sizeof(float) == 0);
25 
26   const float32x4_t vmagic_bias = vld1q_dup_f32(&params->neon_rr2_p5.magic_bias);
27   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neon_rr2_p5.minus_log2e);
28   const float32x4_t vln2_hi = vld1q_dup_f32(&params->neon_rr2_p5.ln2_hi);
29   const float32x4_t vln2_lo = vld1q_dup_f32(&params->neon_rr2_p5.ln2_lo);
30   const float32x4_t vc5 = vld1q_dup_f32(&params->neon_rr2_p5.c5);
31   const float32x4_t vc4 = vld1q_dup_f32(&params->neon_rr2_p5.c4);
32   const float32x4_t vc3 = vld1q_dup_f32(&params->neon_rr2_p5.c3);
33   const float32x4_t vc2 = vld1q_dup_f32(&params->neon_rr2_p5.c2);
34   const float32x4_t vc1 = vld1q_dup_f32(&params->neon_rr2_p5.c1);
35   const float32x4_t vone = vmovq_n_f32(1.0f);
36   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neon_rr2_p5.denorm_cutoff);
37 
38   for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
39     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
40     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
41     const float32x4_t vx89AB = vld1q_f32(x); x += 4;
42     const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
43     const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
44 
45     const float32x4_t vz0123 = vabsq_f32(vx0123);
46     const float32x4_t vz4567 = vabsq_f32(vx4567);
47     const float32x4_t vz89AB = vabsq_f32(vx89AB);
48     const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
49     const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
50 
51     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
52     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
53     float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
54     float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
55     float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
56 
57     const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
58     const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
59     const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
60     const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
61     const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 23));
62 
63     vn0123 = vsubq_f32(vn0123, vmagic_bias);
64     vn4567 = vsubq_f32(vn4567, vmagic_bias);
65     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
66     vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
67     vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
68 
69     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
70     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
71     float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
72     float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
73     float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);
74 
75     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
76     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
77     vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
78     vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
79     vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);
80 
81     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
82     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
83     float32x4_t vp89AB = vmlaq_f32(vc4, vc5, vt89AB);
84     float32x4_t vpCDEF = vmlaq_f32(vc4, vc5, vtCDEF);
85     float32x4_t vpGHIJ = vmlaq_f32(vc4, vc5, vtGHIJ);
86 
87     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
88     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
89     vp89AB = vmlaq_f32(vc3, vp89AB, vt89AB);
90     vpCDEF = vmlaq_f32(vc3, vpCDEF, vtCDEF);
91     vpGHIJ = vmlaq_f32(vc3, vpGHIJ, vtGHIJ);
92 
93     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
94     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
95     vp89AB = vmlaq_f32(vc2, vp89AB, vt89AB);
96     vpCDEF = vmlaq_f32(vc2, vpCDEF, vtCDEF);
97     vpGHIJ = vmlaq_f32(vc2, vpGHIJ, vtGHIJ);
98 
99     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
100     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
101     vp89AB = vmlaq_f32(vc1, vp89AB, vt89AB);
102     vpCDEF = vmlaq_f32(vc1, vpCDEF, vtCDEF);
103     vpGHIJ = vmlaq_f32(vc1, vpGHIJ, vtGHIJ);
104 
105     vt0123 = vmulq_f32(vt0123, vs0123);
106     vt4567 = vmulq_f32(vt4567, vs4567);
107     vt89AB = vmulq_f32(vt89AB, vs89AB);
108     vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
109     vtGHIJ = vmulq_f32(vtGHIJ, vsGHIJ);
110 
111     const float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);
112     const float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);
113     const float32x4_t ve89AB = vmlaq_f32(vs89AB, vp89AB, vt89AB);
114     const float32x4_t veCDEF = vmlaq_f32(vsCDEF, vpCDEF, vtCDEF);
115     const float32x4_t veGHIJ = vmlaq_f32(vsGHIJ, vpGHIJ, vtGHIJ);
116 
117     const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
118     const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
119     const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
120     const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
121     const float32x4_t vdGHIJ = vaddq_f32(veGHIJ, vone);
122 
123     float32x4_t vr0123 = vrecpeq_f32(vd0123);
124     float32x4_t vr4567 = vrecpeq_f32(vd4567);
125     float32x4_t vr89AB = vrecpeq_f32(vd89AB);
126     float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
127     float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
128 
129     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
130     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
131     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
132     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
133     vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
134 
135     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
136     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
137     vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
138     vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
139     vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
140 
141     float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
142     float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
143     float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
144     float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
145     float32x4_t vfGHIJ = vmulq_f32(veGHIJ, vrGHIJ);
146 
147     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
148     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
149     vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
150     vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
151     vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
152 
153     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
154     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
155     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
156     const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
157     const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
158 
159     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
160     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
161     vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
162     vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
163     vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
164 
165     vst1q_f32(y, vf0123); y += 4;
166     vst1q_f32(y, vf4567); y += 4;
167     vst1q_f32(y, vf89AB); y += 4;
168     vst1q_f32(y, vfCDEF); y += 4;
169     vst1q_f32(y, vfGHIJ); y += 4;
170   }
171   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
172     const float32x4_t vx = vld1q_f32(x); x += 4;
173 
174     const float32x4_t vz = vabsq_f32(vx);
175 
176     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
177     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
178     vn = vsubq_f32(vn, vmagic_bias);
179     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
180     vt = vmlaq_f32(vt, vn, vln2_lo);
181 
182     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
183     vp = vmlaq_f32(vc3, vp, vt);
184     vp = vmlaq_f32(vc2, vp, vt);
185     vp = vmlaq_f32(vc1, vp, vt);
186 
187     vt = vmulq_f32(vt, vs);
188     const float32x4_t ve = vmlaq_f32(vs, vp, vt);
189     const float32x4_t vd = vaddq_f32(ve, vone);
190 
191     float32x4_t vr = vrecpeq_f32(vd);
192     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
193     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
194 
195     float32x4_t vf = vmulq_f32(ve, vr);
196     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
197     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
198     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
199 
200     vst1q_f32(y, vf); y += 4;
201   }
202   if XNN_UNLIKELY(n != 0) {
203     const float32x4_t vx = vld1q_f32(x);
204 
205     const float32x4_t vz = vabsq_f32(vx);
206 
207     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
208     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
209     vn = vsubq_f32(vn, vmagic_bias);
210     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
211     vt = vmlaq_f32(vt, vn, vln2_lo);
212 
213     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
214     vp = vmlaq_f32(vc3, vp, vt);
215     vp = vmlaq_f32(vc2, vp, vt);
216     vp = vmlaq_f32(vc1, vp, vt);
217 
218     vt = vmulq_f32(vt, vs);
219     const float32x4_t ve = vmlaq_f32(vs, vp, vt);
220     const float32x4_t vd = vaddq_f32(ve, vone);
221 
222     float32x4_t vr = vrecpeq_f32(vd);
223     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
224     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
225 
226     float32x4_t vf = vmulq_f32(ve, vr);
227     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
228     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
229     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
230 
231     float32x2_t vf_lo = vget_low_f32(vf);
232     if (n & (2 * sizeof(float))) {
233       vst1_f32(y, vf_lo); y += 2;
234       vf_lo = vget_high_f32(vf);
235     }
236     if (n & (1 * sizeof(float))) {
237       vst1_lane_f32(y, vf_lo, 0);
238     }
239   }
240 }
241