1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/neon-lut2048-p1.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_2048[2048];
19
xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_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__neonfma_rr1_lut2048_p1_nr2fma_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->neonfma_rr1_lut2048_p1.magic_bias);
29 const float32x4_t vminus_log2e = vld1q_dup_f32(¶ms->neonfma_rr1_lut2048_p1.minus_log2e);
30 const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
31 const float32x4_t vln2 = vld1q_dup_f32(¶ms->neonfma_rr1_lut2048_p1.ln2);
32 const float32x4_t vc1 = vld1q_dup_f32(¶ms->neonfma_rr1_lut2048_p1.c1);
33 const float32x4_t vone = vmovq_n_f32(1.0f);
34 const float32x4_t vdenorm_cutoff = vld1q_dup_f32(¶ms->neonfma_rr1_lut2048_p1.denorm_cutoff);
35
36 for (; n >= 12 * sizeof(float); n -= 12 * 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
41 const float32x4_t vz0123 = vabsq_f32(vx0123);
42 const float32x4_t vz4567 = vabsq_f32(vx4567);
43 const float32x4_t vz89AB = vabsq_f32(vx89AB);
44
45 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
46 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
47 float32x4_t vn89AB = vfmaq_f32(vmagic_bias, vz89AB, vminus_log2e);
48
49 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 12);
50 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 12);
51 const int32x4_t ve89AB = vshlq_n_s32(vreinterpretq_s32_f32(vn89AB), 12);
52
53 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
54 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
55 const uint64x2_t vidx89AB = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn89AB), vindex_mask));
56
57 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
58 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
59 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx01]);
60 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx23]);
61 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
62 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
63 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx45]);
64 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx67]);
65 const uint64_t vidx89 = vgetq_lane_u64(vidx89AB, 0);
66 const uint64_t vidxAB = vgetq_lane_u64(vidx89AB, 1);
67 float32x2_t vl89 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx89]);
68 float32x2_t vlAB = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidxAB]);
69
70 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx01 >> 32)], vl01, 1);
71 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx23 >> 32)], vl23, 1);
72 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
73 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx45 >> 32)], vl45, 1);
74 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx67 >> 32)], vl67, 1);
75 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
76 vl89 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx89 >> 32)], vl89, 1);
77 vlAB = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidxAB >> 32)], vlAB, 1);
78 const float32x4_t vl89AB = vcombine_f32(vl89, vlAB);
79
80 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
81 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
82 const float32x4_t vs89AB = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl89AB), ve89AB));
83
84 vn0123 = vsubq_f32(vn0123, vmagic_bias);
85 vn4567 = vsubq_f32(vn4567, vmagic_bias);
86 vn89AB = vsubq_f32(vn89AB, vmagic_bias);
87
88 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
89 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
90 float32x4_t vt89AB = vfmaq_f32(vz89AB, vn89AB, vln2);
91
92 const float32x4_t vp0123 = vmulq_f32(vt0123, vc1);
93 const float32x4_t vp4567 = vmulq_f32(vt4567, vc1);
94 const float32x4_t vp89AB = vmulq_f32(vt89AB, vc1);
95
96 const float32x4_t vy0123 = vfmaq_f32(vs0123, vs0123, vp0123);
97 const float32x4_t vy4567 = vfmaq_f32(vs4567, vs4567, vp4567);
98 const float32x4_t vy89AB = vfmaq_f32(vs89AB, vs89AB, vp89AB);
99
100 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
101 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
102 const float32x4_t vd89AB = vaddq_f32(vy89AB, vone);
103
104 float32x4_t vr0123 = vrecpeq_f32(vd0123);
105 float32x4_t vr4567 = vrecpeq_f32(vd4567);
106 float32x4_t vr89AB = vrecpeq_f32(vd89AB);
107
108 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
109 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
110 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB));
111
112 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
113 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
114 vr89AB = vfmaq_f32(vr89AB, vr89AB, vfmsq_f32(vone, vr89AB, vd89AB));
115
116 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
117 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
118 float32x4_t vf89AB = vmulq_f32(vy89AB, vr89AB);
119
120 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
121 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
122 vf89AB = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf89AB), vcagtq_f32(vx89AB, vdenorm_cutoff)));
123
124 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
125 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
126 const uint32x4_t vm89AB = vcltq_f32(vx89AB, vmovq_n_f32(0.0f));
127
128 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
129 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
130 vf89AB = vbslq_f32(vm89AB, vf89AB, vsubq_f32(vone, vf89AB));
131
132 vst1q_f32(y, vf0123); y += 4;
133 vst1q_f32(y, vf4567); y += 4;
134 vst1q_f32(y, vf89AB); y += 4;
135 }
136 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
137 const float32x4_t vx = vld1q_f32(x); x += 4;
138
139 const float32x4_t vz = vabsq_f32(vx);
140
141 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
142 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
143
144 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
145 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
146 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
147 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
148 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
149 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
150 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
151 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
152
153 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
154 vn = vsubq_f32(vn, vmagic_bias);
155 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
156
157 const float32x4_t vp = vmulq_f32(vt, vc1);
158
159 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
160 const float32x4_t vd = vaddq_f32(vy, vone);
161
162 float32x4_t vr = vrecpeq_f32(vd);
163 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
164 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
165
166 float32x4_t vf = vmulq_f32(vy, vr);
167 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
168 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
169 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
170
171 vst1q_f32(y, vf); y += 4;
172 }
173 if XNN_UNLIKELY(n != 0) {
174 const float32x4_t vx = vld1q_f32(x);
175
176 const float32x4_t vz = vabsq_f32(vx);
177
178 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
179 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
180
181 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
182 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
183 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
184 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
185 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
186 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
187 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
188 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
189
190 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
191 vn = vsubq_f32(vn, vmagic_bias);
192 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
193
194 const float32x4_t vp = vmulq_f32(vt, vc1);
195
196 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
197 const float32x4_t vd = vaddq_f32(vy, vone);
198
199 float32x4_t vr = vrecpeq_f32(vd);
200 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
201 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
202
203 float32x4_t vf = vmulq_f32(vy, vr);
204 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
205 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
206 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
207
208 float32x2_t vf_lo = vget_low_f32(vf);
209 if (n & (2 * sizeof(float))) {
210 vst1_f32(y, vf_lo); y += 2;
211 vf_lo = vget_high_f32(vf);
212 }
213 if (n & (1 * sizeof(float))) {
214 vst1_lane_f32(y, vf_lo, 0);
215 }
216 }
217 }
218