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_nr1recps1fma_x16(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_nr1recps1fma_x16(
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(¶ms->neonfma_rr1_p5.magic_bias);
27 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neonfma_rr1_p5.minus_log2e);
28 const float32x4_t vln2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.ln2);
29 const float32x4_t vc5 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c5);
30 const float32x4_t vc4 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c4);
31 const float32x4_t vc3 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c3);
32 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c2);
33 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neonfma_rr1_p5.c1);
34 const float32x4_t vone = vmovq_n_f32(1.0f);
35 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neonfma_rr1_p5.denorm_cutoff);
36
37 for (; n >= 16 * sizeof(float); n -= 16 * 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 const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
42
43 const float32x4_t vz0123 = vabsq_f32(vx0123);
44 const float32x4_t vz4567 = vabsq_f32(vx4567);
45 const float32x4_t vz89AB = vabsq_f32(vx89AB);
46 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
47
48 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
49 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
50 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
51 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
52
53 const float32x4_t vs0123 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 23));
54 const float32x4_t vs4567 = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 23));
55 const float32x4_t vs89AB = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 23));
56 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 23));
57
58 vn0123 = vsubq_f32(vn0123, vmagic_bias);
59 vn4567 = vsubq_f32(vn4567, vmagic_bias);
60 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
61 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
62
63 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
64 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
65 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
66 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
67
68 float32x4_t vp0123 = vfmaq_f32(vc4, vc5, vt0123);
69 float32x4_t vp4567 = vfmaq_f32(vc4, vc5, vt4567);
70 float32x4_t vp89AB = vfmaq_f32(vc4, vc5, vt89AB);
71 float32x4_t vpCDEF = vfmaq_f32(vc4, vc5, vtCDEF);
72
73 vp0123 = vfmaq_f32(vc3, vp0123, vt0123);
74 vp4567 = vfmaq_f32(vc3, vp4567, vt4567);
75 vp89AB = vfmaq_f32(vc3, vp89AB, vt89AB);
76 vpCDEF = vfmaq_f32(vc3, vpCDEF, vtCDEF);
77
78 vp0123 = vfmaq_f32(vc2, vp0123, vt0123);
79 vp4567 = vfmaq_f32(vc2, vp4567, vt4567);
80 vp89AB = vfmaq_f32(vc2, vp89AB, vt89AB);
81 vpCDEF = vfmaq_f32(vc2, vpCDEF, vtCDEF);
82
83 vp0123 = vfmaq_f32(vc1, vp0123, vt0123);
84 vp4567 = vfmaq_f32(vc1, vp4567, vt4567);
85 vp89AB = vfmaq_f32(vc1, vp89AB, vt89AB);
86 vpCDEF = vfmaq_f32(vc1, vpCDEF, vtCDEF);
87
88 vt0123 = vmulq_f32(vt0123, vs0123);
89 vt4567 = vmulq_f32(vt4567, vs4567);
90 vt89AB = vmulq_f32(vt89AB, vs89AB);
91 vtCDEF = vmulq_f32(vtCDEF, vsCDEF);
92
93 const float32x4_t ve0123 = vfmaq_f32(vs0123, vp0123, vt0123);
94 const float32x4_t ve4567 = vfmaq_f32(vs4567, vp4567, vt4567);
95 const float32x4_t ve89AB = vfmaq_f32(vs89AB, vp89AB, vt89AB);
96 const float32x4_t veCDEF = vfmaq_f32(vsCDEF, vpCDEF, vtCDEF);
97
98 const float32x4_t vd0123 = vaddq_f32(ve0123, vone);
99 const float32x4_t vd4567 = vaddq_f32(ve4567, vone);
100 const float32x4_t vd89AB = vaddq_f32(ve89AB, vone);
101 const float32x4_t vdCDEF = vaddq_f32(veCDEF, vone);
102
103 float32x4_t vr0123 = vrecpeq_f32(vd0123);
104 float32x4_t vr4567 = vrecpeq_f32(vd4567);
105 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
106 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
107
108 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
109 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
110 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
111 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
112
113 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
114 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
115 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB));
116 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF));
117
118 float32x4_t vf0123 = vmulq_f32(ve0123, vr0123);
119 float32x4_t vf4567 = vmulq_f32(ve4567, vr4567);
120 float32x4_t vf89AB = vmulq_f32(ve89AB, vr89AB);
121 float32x4_t vfCDEF = vmulq_f32(veCDEF, vrCDEF);
122
123 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
124 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
125 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
126 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
127
128 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
129 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
130 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
131 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
132
133 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
134 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
135 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
136 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
137
138 vst1q_f32(y, vf0123); y += 4;
139 vst1q_f32(y, vf4567); y += 4;
140 vst1q_f32(y, vf89AB); y += 4;
141 vst1q_f32(y, vfCDEF); y += 4;
142 }
143 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
144 const float32x4_t vx = vld1q_f32(x); x += 4;
145
146 const float32x4_t vz = vabsq_f32(vx);
147
148 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
149 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
150 vn = vsubq_f32(vn, vmagic_bias);
151 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
152
153 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
154 vp = vfmaq_f32(vc3, vp, vt);
155 vp = vfmaq_f32(vc2, vp, vt);
156 vp = vfmaq_f32(vc1, vp, vt);
157
158 vt = vmulq_f32(vt, vs);
159 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
160 const float32x4_t vd = vaddq_f32(ve, vone);
161
162 float32x4_t vr = vrecpeq_f32(vd);
163 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
164 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
165
166 float32x4_t vf = vmulq_f32(ve, vr);
167 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
168 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
169 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
170
171 vst1q_f32(y, vf); y += 4;
172 }
173 if XNN_UNLIKELY(n != 0) {
174 const float32x4_t vx = vld1q_f32(x);
175
176 const float32x4_t vz = vabsq_f32(vx);
177
178 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
179 const float32x4_t vs = vreinterpretq_f32_s32(vshlq_n_s32(vreinterpretq_s32_f32(vn), 23));
180 vn = vsubq_f32(vn, vmagic_bias);
181 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
182
183 float32x4_t vp = vfmaq_f32(vc4, vc5, vt);
184 vp = vfmaq_f32(vc3, vp, vt);
185 vp = vfmaq_f32(vc2, vp, vt);
186 vp = vfmaq_f32(vc1, vp, vt);
187
188 vt = vmulq_f32(vt, vs);
189 const float32x4_t ve = vfmaq_f32(vs, vp, vt);
190 const float32x4_t vd = vaddq_f32(ve, vone);
191
192 float32x4_t vr = vrecpeq_f32(vd);
193 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
194 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
195
196 float32x4_t vf = vmulq_f32(ve, vr);
197 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
198 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
199 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
200
201 float32x2_t vf_lo = vget_low_f32(vf);
202 if (n & (2 * sizeof(float))) {
203 vst1_f32(y, vf_lo); y += 2;
204 vf_lo = vget_high_f32(vf);
205 }
206 if (n & (1 * sizeof(float))) {
207 vst1_lane_f32(y, vf_lo, 0);
208 }
209 }
210 }
211