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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * 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
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
50 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
51 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
52 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
53 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
54 float32x4_t vnGHIJ = vmlaq_f32(vmagic_bias, vzGHIJ, vminus_log2e);
55
56 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
57 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
58 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
59 const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
60 const int32x4_t veGHIJ = vshlq_n_s32(vreinterpretq_s32_f32(vnGHIJ), 17);
61
62 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
63 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
64 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
65 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
66 const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
67 const uint64x2_t vidxGHIJ = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnGHIJ), vindex_mask));
68
69 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
70 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
71 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
72 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
73 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
74 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
75 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
76 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
77 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
78 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
79 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
80 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
81 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
82 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
83 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
84 float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
85 const uint64_t vidxGH = vgetq_lane_u64(vidxGHIJ, 0);
86 const uint64_t vidxIJ = vgetq_lane_u64(vidxGHIJ, 1);
87 float32x2_t vlGH = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxGH]);
88 float32x2_t vlIJ = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxIJ]);
89
90 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
91 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
92 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
93 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
94 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
95 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
96 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
97 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
98 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
99 vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
100 vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
101 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
102 vlGH = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxGH >> 32)], vlGH, 1);
103 vlIJ = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxIJ >> 32)], vlIJ, 1);
104 const float32x4_t vlGHIJ = vcombine_f32(vlGH, vlIJ);
105
106 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
107 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
108 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
109 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
110 const float32x4_t vsGHIJ = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlGHIJ), veGHIJ));
111
112 vn0123 = vsubq_f32(vn0123, vmagic_bias);
113 vn4567 = vsubq_f32(vn4567, vmagic_bias);
114 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
115 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
116 vnGHIJ = vsubq_f32(vnGHIJ, vmagic_bias);
117
118 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
119 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
120 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
121 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
122 float32x4_t vtGHIJ = vmlaq_f32(vzGHIJ, vnGHIJ, vln2_hi);
123
124 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
125 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
126 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
127 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
128 vtGHIJ = vmlaq_f32(vtGHIJ, vnGHIJ, vln2_lo);
129
130 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
131 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
132 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
133 float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
134 float32x4_t vpGHIJ = vmulq_f32(vtGHIJ, vc2);
135
136 vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
137 vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
138 vp89AB = vmlsq_f32(vt89AB, vp89AB, vt89AB);
139 vpCDEF = vmlsq_f32(vtCDEF, vpCDEF, vtCDEF);
140 vpGHIJ = vmlsq_f32(vtGHIJ, vpGHIJ, vtGHIJ);
141
142 const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
143 const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
144 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB);
145 const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF);
146 const float32x4_t vyGHIJ = vmlsq_f32(vsGHIJ, vsGHIJ, vpGHIJ);
147
148 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
149 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
150 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
151 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
152 const float32x4_t vdGHIJ = vaddq_f32(vyGHIJ, vone);
153
154 float32x4_t vr0123 = vrecpeq_f32(vd0123);
155 float32x4_t vr4567 = vrecpeq_f32(vd4567);
156 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
157 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
158 float32x4_t vrGHIJ = vrecpeq_f32(vdGHIJ);
159
160 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
161 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
162 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
163 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
164 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
165
166 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
167 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
168 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
169 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
170 vrGHIJ = vmulq_f32(vrGHIJ, vrecpsq_f32(vrGHIJ, vdGHIJ));
171
172 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
173 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
174 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
175 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
176 float32x4_t vfGHIJ = vmulq_f32(vyGHIJ, vrGHIJ);
177
178 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
179 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
180 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
181 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
182 vfGHIJ = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfGHIJ), vcagtq_f32(vxGHIJ, vdenorm_cutoff)));
183
184 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
185 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
186 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
187 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
188 const uint32x4_t vmGHIJ = vcltq_f32(vxGHIJ, vmovq_n_f32(0.0f));
189
190 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
191 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
192 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
193 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
194 vfGHIJ = vbslq_f32(vmGHIJ, vfGHIJ, vsubq_f32(vone, vfGHIJ));
195
196 vst1q_f32(y, vf0123); y += 4;
197 vst1q_f32(y, vf4567); y += 4;
198 vst1q_f32(y, vf89AB); y += 4;
199 vst1q_f32(y, vfCDEF); y += 4;
200 vst1q_f32(y, vfGHIJ); y += 4;
201 }
202 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
203 const float32x4_t vx = vld1q_f32(x); x += 4;
204
205 const float32x4_t vz = vabsq_f32(vx);
206
207 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
208 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
209
210 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
211 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
212 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
213 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
214 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
215 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
216 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
217 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
218
219 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
220 vn = vsubq_f32(vn, vmagic_bias);
221 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
222 vt = vmlaq_f32(vt, vn, vln2_lo);
223
224 float32x4_t vp = vmulq_f32(vt, vc2);
225 vp = vmlsq_f32(vt, vp, vt);
226
227 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
228 const float32x4_t vd = vaddq_f32(vy, vone);
229
230 float32x4_t vr = vrecpeq_f32(vd);
231 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
232 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
233
234 float32x4_t vf = vmulq_f32(vy, vr);
235 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
236 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
237 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
238
239 vst1q_f32(y, vf); y += 4;
240 }
241 if XNN_UNLIKELY(n != 0) {
242 const float32x4_t vx = vld1q_f32(x);
243
244 const float32x4_t vz = vabsq_f32(vx);
245
246 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
247 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
248
249 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
250 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
251 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
252 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
253 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
254 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
255 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
256 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
257
258 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
259 vn = vsubq_f32(vn, vmagic_bias);
260 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
261 vt = vmlaq_f32(vt, vn, vln2_lo);
262
263 float32x4_t vp = vmulq_f32(vt, vc2);
264 vp = vmlsq_f32(vt, vp, vt);
265
266 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
267 const float32x4_t vd = vaddq_f32(vy, vone);
268
269 float32x4_t vr = vrecpeq_f32(vd);
270 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
271 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
272
273 float32x4_t vf = vmulq_f32(vy, vr);
274 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
275 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
276 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
277
278 float32x2_t vf_lo = vget_low_f32(vf);
279 if (n & (2 * sizeof(float))) {
280 vst1_f32(y, vf_lo); y += 2;
281 vf_lo = vget_high_f32(vf);
282 }
283 if (n & (1 * sizeof(float))) {
284 vst1_lane_f32(y, vf_lo, 0);
285 }
286 }
287 }
288