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_div_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__neonfma_rr1_lut64_p2_div_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->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 >= 16 * sizeof(float); n -= 16 * 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
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 const float32x4_t vzCDEF = vabsq_f32(vxCDEF);
46
47 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
48 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
49 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
50 float32x4_t vnCDEF = vfmaq_f32(vmagic_bias, vzCDEF, vminus_log2e);
51
52 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 17);
53 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 17);
54 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 17);
55 const int32x4_t veCDEF = vshlq_n_s32(vreinterpretq_s32_f32(vnCDEF), 17);
56
57 // Use bits 0:6 bits of n, as integer, as an index for table lookup of l := 2**(n % 64).
58 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
59 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
60 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
61 const uint64x2_t vidxCDEF = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vnCDEF), vindex_mask));
62
63 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
64 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
65 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx01]);
66 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx23]);
67 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
68 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
69 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx45]);
70 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx67]);
71 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
72 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
73 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx89]);
74 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxAB]);
75 const uint64_t vidxCD = vgetq_lane_u64(vidxCDEF, 0);
76 const uint64_t vidxEF = vgetq_lane_u64(vidxCDEF, 1);
77 float32x2_t vlCD = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxCD]);
78 float32x2_t vlEF = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidxEF]);
79
80 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx01 >> 32)], vl01, 1);
81 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx23 >> 32)], vl23, 1);
82 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
83 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx45 >> 32)], vl45, 1);
84 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx67 >> 32)], vl67, 1);
85 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
86 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx89 >> 32)], vl89, 1);
87 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxAB >> 32)], vlAB, 1);
88 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
89 vlCD = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxCD >> 32)], vlCD, 1);
90 vlEF = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidxEF >> 32)], vlEF, 1);
91 const float32x4_t vlCDEF = vcombine_f32(vlCD, vlEF);
92
93 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
94 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
95 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
96 const float32x4_t vsCDEF = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vlCDEF), veCDEF));
97
98 vn0123 = vsubq_f32(vn0123, vmagic_bias);
99 vn4567 = vsubq_f32(vn4567, vmagic_bias);
100 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
101 vnCDEF = vsubq_f32(vnCDEF, vmagic_bias);
102
103 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
104 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
105 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
106 float32x4_t vtCDEF = vfmaq_f32(vzCDEF, vnCDEF, vln2);
107
108 float32x4_t vp0123 = vmulq_f32(vt0123, vc2);
109 float32x4_t vp4567 = vmulq_f32(vt4567, vc2);
110 float32x4_t vp89AB = vmulq_f32(vt89AB, vc2);
111 float32x4_t vpCDEF = vmulq_f32(vtCDEF, vc2);
112
113 vp0123 = vfmsq_f32(vt0123, vp0123, vt0123);
114 vp4567 = vfmsq_f32(vt4567, vp4567, vt4567);
115 vp89AB = vfmsq_f32(vt89AB, vp89AB, vt89AB);
116 vpCDEF = vfmsq_f32(vtCDEF, vpCDEF, vtCDEF);
117
118 const float32x4_t vy0123 = vfmsq_f32(vs0123, vs0123, vp0123);
119 const float32x4_t vy4567 = vfmsq_f32(vs4567, vs4567, vp4567);
120 const float32x4_t vy89AB = vfmsq_f32(vs89AB, vs89AB, vp89AB);
121 const float32x4_t vyCDEF = vfmsq_f32(vsCDEF, vsCDEF, vpCDEF);
122
123 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
124 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
125 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
126 const float32x4_t vdCDEF = vaddq_f32(vyCDEF, vone);
127
128 float32x4_t vf0123 = vdivq_f32(vy0123, vd0123);
129 float32x4_t vf4567 = vdivq_f32(vy4567, vd4567);
130 float32x4_t vf89AB = vdivq_f32(vy89AB, vd89AB);
131 float32x4_t vfCDEF = vdivq_f32(vyCDEF, vdCDEF);
132
133 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
134 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
135 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
136 vfCDEF = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vfCDEF), vcagtq_f32(vxCDEF, vdenorm_cutoff)));
137
138 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
139 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
140 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
141 const uint32x4_t vmCDEF = vcltq_f32(vxCDEF, vmovq_n_f32(0.0f));
142
143 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
144 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
145 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
146 vfCDEF = vbslq_f32(vmCDEF, vfCDEF, vsubq_f32(vone, vfCDEF));
147
148 vst1q_f32(y, vf0123); y += 4;
149 vst1q_f32(y, vf4567); y += 4;
150 vst1q_f32(y, vf89AB); y += 4;
151 vst1q_f32(y, vfCDEF); y += 4;
152 }
153 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
154 const float32x4_t vx = vld1q_f32(x); x += 4;
155
156 const float32x4_t vz = vabsq_f32(vx);
157
158 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
159 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
160
161 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
162 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
163 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
164 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
165 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
166 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
167 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
168 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
169
170 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
171 vn = vsubq_f32(vn, vmagic_bias);
172 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
173
174 float32x4_t vp = vmulq_f32(vt, vc2);
175 vp = vfmsq_f32(vt, vp, vt);
176
177 const float32x4_t vy = vfmsq_f32(vs, vs, vp);
178 const float32x4_t vd = vaddq_f32(vy, vone);
179
180 float32x4_t vf = vdivq_f32(vy, vd);
181 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
182 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
183 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
184
185 vst1q_f32(y, vf); y += 4;
186 }
187 if XNN_UNLIKELY(n != 0) {
188 const float32x4_t vx = vld1q_f32(x);
189
190 const float32x4_t vz = vabsq_f32(vx);
191
192 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
193 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 17);
194
195 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
196 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
197 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
198 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_lo]);
199 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) vidx_hi]);
200 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
201 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_64[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
202 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
203
204 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
205 vn = vsubq_f32(vn, vmagic_bias);
206 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
207
208 float32x4_t vp = vmulq_f32(vt, vc2);
209 vp = vfmsq_f32(vt, vp, vt);
210
211 const float32x4_t vy = vfmsq_f32(vs, vs, vp);
212 const float32x4_t vd = vaddq_f32(vy, vone);
213
214 float32x4_t vf = vdivq_f32(vy, vd);
215 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
216 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
217 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
218
219 float32x2_t vf_lo = vget_low_f32(vf);
220 if (n & (2 * sizeof(float))) {
221 vst1_f32(y, vf_lo); y += 2;
222 vf_lo = vget_high_f32(vf);
223 }
224 if (n & (1 * sizeof(float))) {
225 vst1_lane_f32(y, vf_lo, 0);
226 }
227 }
228 }
229