xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-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__neonfma_rr1_p5_div_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__neonfma_rr1_p5_div_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->neonfma_rr1_p5.magic_bias);
27   const float32x4_t vminus_log2e = vld1q_dup_f32(&params->neonfma_rr1_p5.minus_log2e);
28   const float32x4_t vln2 = vld1q_dup_f32(&params->neonfma_rr1_p5.ln2);
29   const float32x4_t vc5 = vld1q_dup_f32(&params->neonfma_rr1_p5.c5);
30   const float32x4_t vc4 = vld1q_dup_f32(&params->neonfma_rr1_p5.c4);
31   const float32x4_t vc3 = vld1q_dup_f32(&params->neonfma_rr1_p5.c3);
32   const float32x4_t vc2 = vld1q_dup_f32(&params->neonfma_rr1_p5.c2);
33   const float32x4_t vc1 = vld1q_dup_f32(&params->neonfma_rr1_p5.c1);
34   const float32x4_t vone = vmovq_n_f32(1.0f);
35   const float32x4_t vdenorm_cutoff = vld1q_dup_f32(&params->neonfma_rr1_p5.denorm_cutoff);
36 
37   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
38     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40 
41     const float32x4_t vz0123 = vabsq_f32(vx0123);
42     const float32x4_t vz4567 = vabsq_f32(vx4567);
43 
44     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
45     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
46 
47     const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
48     const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
49 
50     vn0123 = vsubq_f32(vn0123, vmagic_bias);
51     vn4567 = vsubq_f32(vn4567, vmagic_bias);
52 
53     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
54     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
55 
56     float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
57     float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
58 
59     vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
60     vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
61 
62     vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
63     vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
64 
65     vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
66     vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
67 
68     vt0123 = vmulq_f32(vt0123, vs0123);
69     vt4567 = vmulq_f32(vt4567, vs4567);
70 
71     const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
72     const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
73 
74     const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
75     const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
76 
77     float32x4_t vf0123 = vdivq_f32(ve0123, vd0123);
78     float32x4_t vf4567 = vdivq_f32(ve4567, vd4567);
79 
80     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
81     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
82 
83     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
84     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
85 
86     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
87     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
88 
89     vst1q_f32(y, vf0123); y += 4;
90     vst1q_f32(y, vf4567); y += 4;
91   }
92   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
93     const float32x4_t vx = vld1q_f32(x); x += 4;
94 
95     const float32x4_t vz = vabsq_f32(vx);
96 
97     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
98     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
99     vn = vsubq_f32(vn, vmagic_bias);
100     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
101 
102     float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
103     vp = vfmaq_f32(vc3, vp, vt);
104     vp = vfmaq_f32(vc2, vp, vt);
105     vp = vfmaq_f32(vc1, vp, vt);
106 
107     vt = vmulq_f32(vt, vs);
108     const float32x4_t ve = vfmaq_f32(vs, vp, vt);
109     const float32x4_t vd = vaddq_f32(ve, vone);
110 
111     float32x4_t vf = vdivq_f32(ve, vd);
112     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
113     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
114     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
115 
116     vst1q_f32(y, vf); y += 4;
117   }
118   if XNN_UNLIKELY(n != 0) {
119     const float32x4_t vx = vld1q_f32(x);
120 
121     const float32x4_t vz = vabsq_f32(vx);
122 
123     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
124     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
125     vn = vsubq_f32(vn, vmagic_bias);
126     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
127 
128     float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
129     vp = vfmaq_f32(vc3, vp, vt);
130     vp = vfmaq_f32(vc2, vp, vt);
131     vp = vfmaq_f32(vc1, vp, vt);
132 
133     vt = vmulq_f32(vt, vs);
134     const float32x4_t ve = vfmaq_f32(vs, vp, vt);
135     const float32x4_t vd = vaddq_f32(ve, vone);
136 
137     float32x4_t vf = vdivq_f32(ve, vd);
138     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
139     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
140     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
141 
142     float32x2_t vf_lo = vget_low_f32(vf);
143     if (n & (2 * sizeof(float))) {
144       vst1_f32(y, vf_lo); y += 2;
145       vf_lo = vget_high_f32(vf);
146     }
147     if (n & (1 * sizeof(float))) {
148       vst1_lane_f32(y, vf_lo, 0);
149     }
150   }
151 }
152