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_x12(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_x12(
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 >= 12 * sizeof(float); n -= 12 * 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
42 const float32x4_t vz0123 = vabsq_f32(vx0123);
43 const float32x4_t vz4567 = vabsq_f32(vx4567);
44 const float32x4_t vz89AB = vabsq_f32(vx89AB);
45
46 float32x4_t vn0123 = vmlaq_f32(vmagic_bias, vz0123, vminus_log2e);
47 float32x4_t vn4567 = vmlaq_f32(vmagic_bias, vz4567, vminus_log2e);
48 float32x4_t vn89AB = vmlaq_f32(vmagic_bias, vz89AB, vminus_log2e);
49
50 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
51 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
52 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
53
54 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
55 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
56 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
57 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
58
59 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
60 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
61 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
62 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
63 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
64 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
65 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
66 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
67 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
68 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
69 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
70 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
71
72 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
73 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
74 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
75 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
76 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
77 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
78 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
79 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
80 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
81
82 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
83 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
84 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
85
86 vn0123 = vsubq_f32(vn0123, vmagic_bias);
87 vn4567 = vsubq_f32(vn4567, vmagic_bias);
88 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
89
90 float32x4_t vt0123 = vmlaq_f32(vz0123, vn0123, vln2_hi);
91 float32x4_t vt4567 = vmlaq_f32(vz4567, vn4567, vln2_hi);
92 float32x4_t vt89AB = vmlaq_f32(vz89AB, vn89AB, vln2_hi);
93
94 vt0123 = vmlaq_f32(vt0123, vn0123, vln2_lo);
95 vt4567 = vmlaq_f32(vt4567, vn4567, vln2_lo);
96 vt89AB = vmlaq_f32(vt89AB, vn89AB, vln2_lo);
97
98 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
99 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
100 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
101
102 vp0123 = vmlsq_f32(vt0123, vp0123, vt0123);
103 vp4567 = vmlsq_f32(vt4567, vp4567, vt4567);
104 vp89AB = vmlsq_f32(vt89AB, vp89AB, vt89AB);
105
106 const float32x4_t vy0123 = vmlsq_f32(vs0123, vs0123, vp0123);
107 const float32x4_t vy4567 = vmlsq_f32(vs4567, vs4567, vp4567);
108 const float32x4_t vy89AB = vmlsq_f32(vs89AB, vs89AB, vp89AB);
109
110 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
111 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
112 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
113
114 float32x4_t vr0123 = vrecpeq_f32(vd0123);
115 float32x4_t vr4567 = vrecpeq_f32(vd4567);
116 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
117
118 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
119 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
120 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
121
122 vr0123 = vmulq_f32(vr0123, vrecpsq_f32(vr0123, vd0123));
123 vr4567 = vmulq_f32(vr4567, vrecpsq_f32(vr4567, vd4567));
124 vr89AB = vmulq_f32(vr89AB, vrecpsq_f32(vr89AB, vd89AB));
125
126 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
127 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
128 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
129
130 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
131 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
132 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
133
134 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
135 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
136 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
137
138 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
139 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
140 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
141
142 vst1q_f32(y, vf0123); y += 4;
143 vst1q_f32(y, vf4567); y += 4;
144 vst1q_f32(y, vf89AB); y += 4;
145 }
146 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
147 const float32x4_t vx = vld1q_f32(x); x += 4;
148
149 const float32x4_t vz = vabsq_f32(vx);
150
151 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
152 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
153
154 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
155 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
156 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
157 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
158 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
159 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
160 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
161 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
162
163 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
164 vn = vsubq_f32(vn, vmagic_bias);
165 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
166 vt = vmlaq_f32(vt, vn, vln2_lo);
167
168 float32x4_t vp = vmulq_f32(vt, vc2);
169 vp = vmlsq_f32(vt, vp, vt);
170
171 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
172 const float32x4_t vd = vaddq_f32(vy, vone);
173
174 float32x4_t vr = vrecpeq_f32(vd);
175 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
176 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
177
178 float32x4_t vf = vmulq_f32(vy, vr);
179 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
180 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
181 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
182
183 vst1q_f32(y, vf); y += 4;
184 }
185 if XNN_UNLIKELY(n != 0) {
186 const float32x4_t vx = vld1q_f32(x);
187
188 const float32x4_t vz = vabsq_f32(vx);
189
190 float32x4_t vn = vmlaq_f32(vmagic_bias, vz, vminus_log2e);
191 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
192
193 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
194 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
195 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
196 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
197 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
198 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
199 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
200 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
201
202 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
203 vn = vsubq_f32(vn, vmagic_bias);
204 float32x4_t vt = vmlaq_f32(vz, vn, vln2_hi);
205 vt = vmlaq_f32(vt, vn, vln2_lo);
206
207 float32x4_t vp = vmulq_f32(vt, vc2);
208 vp = vmlsq_f32(vt, vp, vt);
209
210 const float32x4_t vy = vmlsq_f32(vs, vs, vp);
211 const float32x4_t vd = vaddq_f32(vy, vone);
212
213 float32x4_t vr = vrecpeq_f32(vd);
214 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
215 vr = vmulq_f32(vr, vrecpsq_f32(vr, vd));
216
217 float32x4_t vf = vmulq_f32(vy, vr);
218 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
219 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
220 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
221
222 float32x2_t vf_lo = vget_low_f32(vf);
223 if (n & (2 * sizeof(float))) {
224 vst1_f32(y, vf_lo); y += 2;
225 vf_lo = vget_high_f32(vf);
226 }
227 if (n & (1 * sizeof(float))) {
228 vst1_lane_f32(y, vf_lo, 0);
229 }
230 }
231 }
232