xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-neonfma-rr1-p5-div-x12.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_x12(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_x12(
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 >= 12 * sizeof(float); n -= 12 * sizeof(float)) {
38     const float32x4_t vx0123 = vld1q_f32(x); x += 4;
39     const float32x4_t vx4567 = vld1q_f32(x); x += 4;
40     const float32x4_t vx89AB = vld1q_f32(x); x += 4;
41 
42     const float32x4_t vz0123 = vabsq_f32(vx0123);
43     const float32x4_t vz4567 = vabsq_f32(vx4567);
44     const float32x4_t vz89AB = vabsq_f32(vx89AB);
45 
46     float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
47     float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
48     float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
49 
50     const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
51     const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
52     const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
53 
54     vn0123 = vsubq_f32(vn0123, vmagic_bias);
55     vn4567 = vsubq_f32(vn4567, vmagic_bias);
56     vn89AB = vsubq_f32(vn89AB, vmagic_bias);
57 
58     float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
59     float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
60     float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
61 
62     float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
63     float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
64     float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB);
65 
66     vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
67     vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
68     vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
69 
70     vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
71     vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
72     vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
73 
74     vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
75     vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
76     vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB);
77 
78     vt0123 = vmulq_f32(vt0123, vs0123);
79     vt4567 = vmulq_f32(vt4567, vs4567);
80     vt89AB = vmulq_f32(vt89AB, vs89AB);
81 
82     const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
83     const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
84     const float32x4_t ve89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB);
85 
86     const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
87     const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
88     const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
89 
90     float32x4_t vf0123 = vdivq_f32(ve0123, vd0123);
91     float32x4_t vf4567 = vdivq_f32(ve4567, vd4567);
92     float32x4_t vf89AB = vdivq_f32(ve89AB, vd89AB);
93 
94     vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
95     vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
96     vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
97 
98     const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
99     const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
100     const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
101 
102     vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
103     vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
104     vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
105 
106     vst1q_f32(y, vf0123); y += 4;
107     vst1q_f32(y, vf4567); y += 4;
108     vst1q_f32(y, vf89AB); y += 4;
109   }
110   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
111     const float32x4_t vx = vld1q_f32(x); x += 4;
112 
113     const float32x4_t vz = vabsq_f32(vx);
114 
115     float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
116     const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
117     vn = vsubq_f32(vn, vmagic_bias);
118     float32x4_t vt = vfmaq_f32(vz, vn, vln2);
119 
120     float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
121     vp = vfmaq_f32(vc3, vp, vt);
122     vp = vfmaq_f32(vc2, vp, vt);
123     vp = vfmaq_f32(vc1, vp, vt);
124 
125     vt = vmulq_f32(vt, vs);
126     const float32x4_t ve = vfmaq_f32(vs, vp, vt);
127     const float32x4_t vd = vaddq_f32(ve, vone);
128 
129     float32x4_t vf = vdivq_f32(ve, vd);
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 = vfmaq_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 = vfmaq_f32(vz, vn, vln2);
145 
146     float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
147     vp = vfmaq_f32(vc3, vp, vt);
148     vp = vfmaq_f32(vc2, vp, vt);
149     vp = vfmaq_f32(vc1, vp, vt);
150 
151     vt = vmulq_f32(vt, vs);
152     const float32x4_t ve = vfmaq_f32(vs, vp, vt);
153     const float32x4_t vd = vaddq_f32(ve, vone);
154 
155     float32x4_t vf = vdivq_f32(ve, vd);
156     vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
157     const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
158     vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
159 
160     float32x2_t vf_lo = vget_low_f32(vf);
161     if (n & (2 * sizeof(float))) {
162       vst1_f32(y, vf_lo); y += 2;
163       vf_lo = vget_high_f32(vf);
164     }
165     if (n & (1 * sizeof(float))) {
166       vst1_lane_f32(y, vf_lo, 0);
167     }
168   }
169 }
170