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_x16(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_x16(
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 >= 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 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
49 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
50 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
51 float32x4_t vnCDEF = vmlaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
52
53 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
54 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
55 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
56 const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
57
58 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
59 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
60 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
61 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
62 const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
63
64 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
65 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
66 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
67 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
68 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
69 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
70 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
71 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
72 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
73 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
74 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
75 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
76 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
77 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
78 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
79 float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
80
81 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
82 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
83 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
84 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
85 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
86 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
87 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
88 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
89 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
90 vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
91 vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
92 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
93
94 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
95 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
96 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
97 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
98
99 vn0123 = vsubq_f32(vn0123, vmagic_bias);
100 vn4567 = vsubq_f32(vn4567, vmagic_bias);
101 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
102 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
103
104 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
105 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
106 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
107 float32x4_t vtCDEF = vmlaq_f32(vzCDEF, vnCDEF, vln2_hi);
108
109 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
110 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
111 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
112 vtCDEF = vmlaq_f32(vtCDEF, vnCDEF, vln2_lo);
113
114 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
115 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
116 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
117 float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
118
119 vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
120 vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
121 vp89AB = vmlsq_f32(vt89AB, vp89AB, vt89AB);
122 vpCDEF = vmlsq_f32(vtCDEF, vpCDEF, vtCDEF);
123
124 const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
125 const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
126 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB);
127 const float32x4_t vyCDEF = vmlsq_f32(vsCDEF, vsCDEF, vpCDEF);
128
129 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
130 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
131 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
132 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
133
134 float32x4_t vr0123 = vrecpeq_f32(vd0123);
135 float32x4_t vr4567 = vrecpeq_f32(vd4567);
136 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
137 float32x4_t vrCDEF = vrecpeq_f32(vdCDEF);
138
139 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
140 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
141 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
142 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
143
144 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
145 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
146 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
147 vrCDEF = vmulq_f32(vrCDEF, vrecpsq_f32(vrCDEF, vdCDEF));
148
149 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
150 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
151 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
152 float32x4_t vfCDEF = vmulq_f32(vyCDEF, vrCDEF);
153
154 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
155 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
156 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
157 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
158
159 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
160 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
161 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
162 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
163
164 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
165 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
166 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
167 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
168
169 vst1q_f32(y, vf0123); y += 4;
170 vst1q_f32(y, vf4567); y += 4;
171 vst1q_f32(y, vf89AB); y += 4;
172 vst1q_f32(y, vfCDEF); y += 4;
173 }
174 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
175 const float32x4_t vx = vld1q_f32(x); x += 4;
176
177 const float32x4_t vz = vabsq_f32(vx);
178
179 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
180 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
181
182 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
183 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
184 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
185 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
186 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
187 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
188 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
189 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
190
191 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
192 vn = vsubq_f32(vn, vmagic_bias);
193 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
194 vt = vmlaq_f32(vt, vn, vln2_lo);
195
196 float32x4_t vp = vmulq_f32(vt, vc2);
197 vp = vmlsq_f32(vt, vp, vt);
198
199 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
200 const float32x4_t vd = vaddq_f32(vy, vone);
201
202 float32x4_t vr = vrecpeq_f32(vd);
203 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
204 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
205
206 float32x4_t vf = vmulq_f32(vy, vr);
207 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
208 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
209 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
210
211 vst1q_f32(y, vf); y += 4;
212 }
213 if XNN_UNLIKELY(n != 0) {
214 const float32x4_t vx = vld1q_f32(x);
215
216 const float32x4_t vz = vabsq_f32(vx);
217
218 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
219 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
220
221 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
222 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
223 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
224 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
225 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
226 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
227 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
228 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
229
230 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
231 vn = vsubq_f32(vn, vmagic_bias);
232 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
233 vt = vmlaq_f32(vt, vn, vln2_lo);
234
235 float32x4_t vp = vmulq_f32(vt, vc2);
236 vp = vmlsq_f32(vt, vp, vt);
237
238 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
239 const float32x4_t vd = vaddq_f32(vy, vone);
240
241 float32x4_t vr = vrecpeq_f32(vd);
242 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
243 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
244
245 float32x4_t vf = vmulq_f32(vy, vr);
246 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
247 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
248 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
249
250 float32x2_t vf_lo = vget_low_f32(vf);
251 if (n & (2 * sizeof(float))) {
252 vst1_f32(y, vf_lo); y += 2;
253 vf_lo = vget_high_f32(vf);
254 }
255 if (n & (1 * sizeof(float))) {
256 vst1_lane_f32(y, vf_lo, 0);
257 }
258 }
259 }
260