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__neon_rr2_lut64_p2_nr2recps_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__neon_rr2_lut64_p2_nr2recps_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->neon_rr2_lut64_p2.magic_bias);
29 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neon_rr2_lut64_p2.minus_log2e);
30 const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x3F));
31 const float32x4_t vln2_hi = vld1q_dup_f32(¶ms->neon_rr2_lut64_p2.ln2_hi);
32 const float32x4_t vln2_lo = vld1q_dup_f32(¶ms->neon_rr2_lut64_p2.ln2_lo);
33 const float32x4_t vc2 = vld1q_dup_f32(¶ms->neon_rr2_lut64_p2.c2);
34 const float32x4_t vone = vmovq_n_f32(1.0f);
35 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neon_rr2_lut64_p2.denorm_cutoff);
36
37 for (; n >= 24 * sizeof(float); n -= 24 * 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 const float32x4_t vxGHIJ = vld1q_f32(x); x += 4;
43 const float32x4_t vxKLMN = vld1q_f32(x); x += 4;
44
45 const float32x4_t vz0123 = vabsq_f32(vx0123);
46 const float32x4_t vz4567 = vabsq_f32(vx4567);
47 const float32x4_t vz89AB = vabsq_f32(vx89AB);
48 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
49 const float32x4_t vzGHIJ = vabsq_f32(vxGHIJ);
50 const float32x4_t vzKLMN = vabsq_f32(vxKLMN);
51
52 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
53 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
54 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
55 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
56 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
57 float32x4_t vnKLMN = vmlaq_f32(vmagic_bias, vzKLMN, vminus_log2e);
58
59 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
60 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
61 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
62 const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
63 const int32x4_t veGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 17);
64 const int32x4_t veKLMN = vshlq_n_s32(vreinterpretq_s32_f32(vnKLMN), 17);
65
66 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
67 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
68 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
69 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
70 const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
71 const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask));
72 const uint64x2_t vidxKLMN = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnKLMN), vindex_mask));
73
74 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
75 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
76 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
77 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
78 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
79 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
80 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
81 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
82 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
83 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
84 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
85 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
86 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
87 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
88 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
89 float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
90 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
91 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
92 float32x2_t vlGH = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxGH]);
93 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]);
94 const uint64_t vidxKL = vgetq_lane_u64(vidxKLMN, 0);
95 const uint64_t vidxMN = vgetq_lane_u64(vidxKLMN, 1);
96 float32x2_t vlKL = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxKL]);
97 float32x2_t vlMN = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxMN]);
98
99 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
100 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
101 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
102 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
103 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
104 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
105 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
106 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
107 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
108 vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
109 vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
110 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
111 vlGH = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxGH >> 32)], vlGH, 1);
112 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1);
113 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ);
114 vlKL = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxKL >> 32)], vlKL, 1);
115 vlMN = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxMN >> 32)], vlMN, 1);
116 const float32x4_t vlKLMN = vcombine_f32(vlKL, vlMN);
117
118 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
119 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
120 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
121 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
122 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ));
123 const float32x4_t vsKLMN = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlKLMN), veKLMN));
124
125 vn0123 = vsubq_f32(vn0123, vmagic_bias);
126 vn4567 = vsubq_f32(vn4567, vmagic_bias);
127 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
128 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
129 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
130 vnKLMN = vsubq_f32(vnKLMN, vmagic_bias);
131
132 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
133 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
134 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
135 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
136 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);
137 float32x4_t vtKLMN = vmlaq_f32(vzKLMN, vnKLMN, vln2_hi);
138
139 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
140 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
141 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
142 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
143 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);
144 vtKLMN = vmlaq_f32(vtKLMN, vnKLMN, vln2_lo);
145
146 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
147 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
148 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
149 float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
150 float32x4_t vpGHIJ = vmulq_f32(vtGHIJ, vc2);
151 float32x4_t vpKLMN = vmulq_f32(vtKLMN, vc2);
152
153 vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
154 vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
155 vp89AB = vmlsq_f32(vt89AB, vp89AB, vt89AB);
156 vpCDEF = vmlsq_f32(vtCDEF, vpCDEF, vtCDEF);
157 vpGHIJ = vmlsq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
158 vpKLMN = vmlsq_f32(vtKLMN, vpKLMN, vtKLMN);
159
160 const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
161 const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
162 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB);
163 const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF);
164 const float32x4_t vyGHIJ = vmlsq_f32(vsGHIJ, vsGHIJ, vpGHIJ);
165 const float32x4_t vyKLMN = vmlsq_f32(vsKLMN, vsKLMN, vpKLMN);
166
167 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
168 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
169 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
170 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
171 const float32x4_t vdGHIJ = vaddq_f32(vyGHIJ, vone);
172 const float32x4_t vdKLMN = vaddq_f32(vyKLMN, vone);
173
174 float32x4_t vr0123 = vrecpeq_f32(vd0123);
175 float32x4_t vr4567 = vrecpeq_f32(vd4567);
176 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
177 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
178 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
179 float32x4_t vrKLMN = vrecpeq_f32(vdKLMN);
180
181 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
182 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
183 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
184 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
185 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
186 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
187
188 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
189 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
190 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
191 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
192 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
193 vrKLMN = vmulq_f32(vrKLMN, vrecpsq_f32(vrKLMN, vdKLMN));
194
195 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
196 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
197 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
198 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
199 float32x4_t vfGHIJ = vmulq_f32(vyGHIJ, vrGHIJ);
200 float32x4_t vfKLMN = vmulq_f32(vyKLMN, vrKLMN);
201
202 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
203 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
204 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
205 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
206 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
207 vfKLMN = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfKLMN), vcagtq_f32(vxKLMN, vdenorm_cutoff)));
208
209 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
210 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
211 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
212 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
213 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
214 const uint32x4_t vmKLMN = vcltq_f32(vxKLMN, vmovq_n_f32(0.0f));
215
216 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
217 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
218 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
219 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
220 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
221 vfKLMN = vbslq_f32(vmKLMN, vfKLMN, vsubq_f32(vone, vfKLMN));
222
223 vst1q_f32(y, vf0123); y += 4;
224 vst1q_f32(y, vf4567); y += 4;
225 vst1q_f32(y, vf89AB); y += 4;
226 vst1q_f32(y, vfCDEF); y += 4;
227 vst1q_f32(y, vfGHIJ); y += 4;
228 vst1q_f32(y, vfKLMN); y += 4;
229 }
230 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
231 const float32x4_t vx = vld1q_f32(x); x += 4;
232
233 const float32x4_t vz = vabsq_f32(vx);
234
235 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
236 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
237
238 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
239 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
240 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
241 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
242 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
243 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
244 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
245 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
246
247 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
248 vn = vsubq_f32(vn, vmagic_bias);
249 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
250 vt = vmlaq_f32(vt, vn, vln2_lo);
251
252 float32x4_t vp = vmulq_f32(vt, vc2);
253 vp = vmlsq_f32(vt, vp, vt);
254
255 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
256 const float32x4_t vd = vaddq_f32(vy, vone);
257
258 float32x4_t vr = vrecpeq_f32(vd);
259 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
260 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
261
262 float32x4_t vf = vmulq_f32(vy, vr);
263 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
264 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
265 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
266
267 vst1q_f32(y, vf); y += 4;
268 }
269 if XNN_UNLIKELY(n != 0) {
270 const float32x4_t vx = vld1q_f32(x);
271
272 const float32x4_t vz = vabsq_f32(vx);
273
274 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
275 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
276
277 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
278 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
279 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
280 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
281 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
282 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
283 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
284 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
285
286 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
287 vn = vsubq_f32(vn, vmagic_bias);
288 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
289 vt = vmlaq_f32(vt, vn, vln2_lo);
290
291 float32x4_t vp = vmulq_f32(vt, vc2);
292 vp = vmlsq_f32(vt, vp, vt);
293
294 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
295 const float32x4_t vd = vaddq_f32(vy, vone);
296
297 float32x4_t vr = vrecpeq_f32(vd);
298 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
299 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
300
301 float32x4_t vf = vmulq_f32(vy, vr);
302 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
303 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
304 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
305
306 float32x2_t vf_lo = vget_low_f32(vf);
307 if (n & (2 * sizeof(float))) {
308 vst1_f32(y, vf_lo); y += 2;
309 vf_lo = vget_high_f32(vf);
310 }
311 if (n & (1 * sizeof(float))) {
312 vst1_lane_f32(y, vf_lo, 0);
313 }
314 }
315 }
316