1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/neon-lut64-p2.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
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19
xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr1recps1fma_x24(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n % sizeof(float) == 0);
27
28 const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->neonfma_rr1_lut64_p2.magic_bias);
29 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neonfma_rr1_lut64_p2.minus_log2e);
30 const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31 const float32x4_t vln2 = vld1q_dup_f32(¶ms->neonfma_rr1_lut64_p2.ln2);
32 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neonfma_rr1_lut64_p2.c2);
33 const float32x4_t vone = vmovq_n_f32(1.0f);
34 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neonfma_rr1_lut64_p2.denorm_cutoff);
35
36 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
37 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
38 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
39 const float32x4_t vx89AB = vld1q_f32(x); x += 4;
40 const float32x4_t vxCDEF = vld1q_f32(x); x += 4;
41 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
42 const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
43
44 const float32x4_t vz0123 = vabsq_f32(vx0123);
45 const float32x4_t vz4567 = vabsq_f32(vx4567);
46 const float32x4_t vz89AB = vabsq_f32(vx89AB);
47 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
48 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
49 const float32x4_t vzKLMN = vabsq_f32(vxKLMN);
50
51 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
52 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
53 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
54 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
55 float32x4_t vnGHIJ = vfmaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
56 float32x4_t vnKLMN = vfmaq_f32(vmagic_bias, vzKLMN, vminus_log2e);
57
58 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
59 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
60 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
61 const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
62 const int32x4_t veGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 17);
63 const int32x4_t veKLMN = vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 17);
64
65 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
66 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
67 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
68 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
69 const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
70 const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask));
71 const uint64x2_t vidxKLMN = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnKLMN), vindex_mask));
72
73 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
74 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
75 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
76 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
77 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
78 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
79 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
80 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
81 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
82 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
83 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
84 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
85 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
86 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
87 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
88 float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
89 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
90 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
91 float32x2_t vlGH = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxGH]);
92 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]);
93 const uint64_t vidxKL = vgetq_lane_u64(vidxKLMN, 0);
94 const uint64_t vidxMN = vgetq_lane_u64(vidxKLMN, 1);
95 float32x2_t vlKL = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxKL]);
96 float32x2_t vlMN = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxMN]);
97
98 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
99 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
100 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
101 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
102 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
103 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
104 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
105 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
106 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
107 vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
108 vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
109 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
110 vlGH = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxGH >> 32)], vlGH, 1);
111 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1);
112 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ);
113 vlKL = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxKL >> 32)], vlKL, 1);
114 vlMN = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxMN >> 32)], vlMN, 1);
115 const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN);
116
117 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
118 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
119 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
120 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
121 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ));
122 const float32x4_t vsKLMN = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlKLMN), veKLMN));
123
124 vn0123 = vsubq_f32(vn0123, vmagic_bias);
125 vn4567 = vsubq_f32(vn4567, vmagic_bias);
126 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
127 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
128 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
129 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
130
131 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
132 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
133 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
134 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
135 float32x4_t vtGHIJ = vfmaq_f32(vzGHIJ, vnGHIJ, vln2);
136 float32x4_t vtKLMN = vfmaq_f32(vzKLMN, vnKLMN, vln2);
137
138 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
139 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
140 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
141 float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
142 float32x4_t vpGHIJ = vmulq_f32(vtGHIJ, vc2);
143 float32x4_t vpKLMN = vmulq_f32(vtKLMN, vc2);
144
145 vp0123 = vfmsq_f32(vt0123, vp0123, vt0123);
146 vp4567 = vfmsq_f32(vt4567, vp4567, vt4567);
147 vp89AB = vfmsq_f32(vt89AB, vp89AB, vt89AB);
148 vpCDEF = vfmsq_f32(vtCDEF, vpCDEF, vtCDEF);
149 vpGHIJ = vfmsq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
150 vpKLMN = vfmsq_f32(vtKLMN, vpKLMN, vtKLMN);
151
152 const float32x4_t vy0123 = vfmsq_f32(vs0123, vs0123, vp0123);
153 const float32x4_t vy4567 = vfmsq_f32(vs4567, vs4567, vp4567);
154 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB);
155 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF);
156 const float32x4_t vyGHIJ = vfmsq_f32(vsGHIJ, vsGHIJ, vpGHIJ);
157 const float32x4_t vyKLMN = vfmsq_f32(vsKLMN, vsKLMN, vpKLMN);
158
159 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
160 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
161 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
162 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
163 const float32x4_t vdGHIJ = vaddq_f32(vyGHIJ, vone);
164 const float32x4_t vdKLMN = vaddq_f32(vyKLMN, vone);
165
166 float32x4_t vr0123 = vrecpeq_f32(vd0123);
167 float32x4_t vr4567 = vrecpeq_f32(vd4567);
168 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
169 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
170 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
171 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN);
172
173 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
174 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
175 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
176 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
177 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
178 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
179
180 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
181 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
182 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB));
183 vrCDEF = vfmaq_f32(vrCDEF, vrCDEF, vfmsq_f32(vone, vrCDEF, vdCDEF));
184 vrGHIJ = vfmaq_f32(vrGHIJ, vrGHIJ, vfmsq_f32(vone, vrGHIJ, vdGHIJ));
185 vrKLMN = vfmaq_f32(vrKLMN, vrKLMN, vfmsq_f32(vone, vrKLMN, vdKLMN));
186
187 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
188 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
189 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
190 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
191 float32x4_t vfGHIJ = vmulq_f32(vyGHIJ, vrGHIJ);
192 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN);
193
194 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
195 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
196 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
197 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
198 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
199 vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm_cutoff)));
200
201 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
202 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
203 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
204 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
205 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
206 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
207
208 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
209 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
210 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
211 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
212 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
213 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN));
214
215 vst1q_f32(y, vf0123); y += 4;
216 vst1q_f32(y, vf4567); y += 4;
217 vst1q_f32(y, vf89AB); y += 4;
218 vst1q_f32(y, vfCDEF); y += 4;
219 vst1q_f32(y, vfGHIJ); y += 4;
220 vst1q_f32(y, vfKLMN); y += 4;
221 }
222 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
223 const float32x4_t vx = vld1q_f32(x); x += 4;
224
225 const float32x4_t vz = vabsq_f32(vx);
226
227 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
228 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
229
230 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
231 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
232 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
233 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
234 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
235 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
236 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
237 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
238
239 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
240 vn = vsubq_f32(vn, vmagic_bias);
241 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
242
243 float32x4_t vp = vmulq_f32(vt, vc2);
244 vp = vfmsq_f32(vt, vp, vt);
245
246 const float32x4_t vy = vfmsq_f32(vs, vs, vp);
247 const float32x4_t vd = vaddq_f32(vy, vone);
248
249 float32x4_t vr = vrecpeq_f32(vd);
250 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
251 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
252
253 float32x4_t vf = vmulq_f32(vy, vr);
254 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
255 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
256 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
257
258 vst1q_f32(y, vf); y += 4;
259 }
260 if XNN_UNLIKELY(n != 0) {
261 const float32x4_t vx = vld1q_f32(x);
262
263 const float32x4_t vz = vabsq_f32(vx);
264
265 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
266 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
267
268 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
269 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
270 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
271 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
272 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
273 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
274 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
275 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
276
277 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
278 vn = vsubq_f32(vn, vmagic_bias);
279 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
280
281 float32x4_t vp = vmulq_f32(vt, vc2);
282 vp = vfmsq_f32(vt, vp, vt);
283
284 const float32x4_t vy = vfmsq_f32(vs, vs, vp);
285 const float32x4_t vd = vaddq_f32(vy, vone);
286
287 float32x4_t vr = vrecpeq_f32(vd);
288 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
289 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
290
291 float32x4_t vf = vmulq_f32(vy, vr);
292 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
293 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
294 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
295
296 float32x2_t vf_lo = vget_low_f32(vf);
297 if (n & (2 * sizeof(float))) {
298 vst1_f32(y, vf_lo); y += 2;
299 vf_lo = vget_high_f32(vf);
300 }
301 if (n & (1 * sizeof(float))) {
302 vst1_lane_f32(y, vf_lo, 0);
303 }
304 }
305 }
306