xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neon-rr2-p5-nr2recps-x8.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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
39     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
40     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
41 
42     const float32x4_t vz0123 = vabsq_f32(vx0123);
43     const float32x4_t vz4567 = vabsq_f32(vx4567);
44 
45     float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
46     float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
47 
48     const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
49     const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
50 
51     vn0123 = vsubq_f32(vn0123, vmagic_bias);
52     vn4567 = vsubq_f32(vn4567, vmagic_bias);
53 
54     float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
55     float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
56 
57     vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
58     vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
59 
60     float32x4_t vp0123 = vmlaq_f32(vc4, vc5, vt0123);
61     float32x4_t vp4567 = vmlaq_f32(vc4, vc5, vt4567);
62 
63     vp0123 = vmlaq_f32(vc3, vp0123, vt0123);
64     vp4567 = vmlaq_f32(vc3, vp4567, vt4567);
65 
66     vp0123 = vmlaq_f32(vc2, vp0123, vt0123);
67     vp4567 = vmlaq_f32(vc2, vp4567, vt4567);
68 
69     vp0123 = vmlaq_f32(vc1, vp0123, vt0123);
70     vp4567 = vmlaq_f32(vc1, vp4567, vt4567);
71 
72     vt0123 = vmulq_f32(vt0123, vs0123);
73     vt4567 = vmulq_f32(vt4567, vs4567);
74 
75     const float32x4_t ve0123 = vmlaq_f32(vs0123, vp0123, vt0123);
76     const float32x4_t ve4567 = vmlaq_f32(vs4567, vp4567, vt4567);
77 
78     const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
79     const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
80 
81     float32x4_t vr0123 = vrecpeq_f32(vd0123);
82     float32x4_t vr4567 = vrecpeq_f32(vd4567);
83 
84     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
85     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
86 
87     vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
88     vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
89 
90     float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
91     float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
92 
93     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
94     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
95 
96     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
97     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
98 
99     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
100     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
101 
102     vst1q_f32(y, vf0123); y += 4;
103     vst1q_f32(y, vf4567); y += 4;
104   }
105   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
106     const float32x4_t vx = vld1q_f32(x); x += 4;
107 
108     const float32x4_t vz = vabsq_f32(vx);
109 
110     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
111     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
112     vn = vsubq_f32(vn, vmagic_bias);
113     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
114     vt = vmlaq_f32(vt, vn, vln2_lo);
115 
116     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
117     vp = vmlaq_f32(vc3, vp, vt);
118     vp = vmlaq_f32(vc2, vp, vt);
119     vp = vmlaq_f32(vc1, vp, vt);
120 
121     vt = vmulq_f32(vt, vs);
122     const float32x4_t ve = vmlaq_f32(vs, vp, vt);
123     const float32x4_t vd = vaddq_f32(ve, vone);
124 
125     float32x4_t vr = vrecpeq_f32(vd);
126     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
127     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
128 
129     float32x4_t vf = vmulq_f32(ve, vr);
130     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
131     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
132     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
133 
134     vst1q_f32(y, vf); y += 4;
135   }
136   if XNN_UNLIKELY(n != 0) {
137     const float32x4_t vx = vld1q_f32(x);
138 
139     const float32x4_t vz = vabsq_f32(vx);
140 
141     float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
142     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
143     vn = vsubq_f32(vn, vmagic_bias);
144     float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
145     vt = vmlaq_f32(vt, vn, vln2_lo);
146 
147     float32x4_t vp = vmlaq_f32(vc4, vc5, vt);
148     vp = vmlaq_f32(vc3, vp, vt);
149     vp = vmlaq_f32(vc2, vp, vt);
150     vp = vmlaq_f32(vc1, vp, vt);
151 
152     vt = vmulq_f32(vt, vs);
153     const float32x4_t ve = vmlaq_f32(vs, vp, vt);
154     const float32x4_t vd = vaddq_f32(ve, vone);
155 
156     float32x4_t vr = vrecpeq_f32(vd);
157     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
158     vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
159 
160     float32x4_t vf = vmulq_f32(ve, vr);
161     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
162     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
163     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
164 
165     float32x2_t vf_lo = vget_low_f32(vf);
166     if (n & (2 * sizeof(float))) {
167       vst1_f32(y, vf_lo); y += 2;
168       vf_lo = vget_high_f32(vf);
169     }
170     if (n & (1 * sizeof(float))) {
171       vst1_lane_f32(y, vf_lo, 0);
172     }
173   }
174 }
175