1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/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_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__neonfma_rr1_lut2048_p1_nr2fma_x8(
21 size_t n,
22 const float* x,
23 float* y,
24 const void* params) XNN_DISABLE_TSAN
25 {
26 assert(n % sizeof(float) == 0);
27
28 const float32x4_t vmagic_bias = vmovq_n_f32(0x1.800000p12f);
29 const float32x4_t vminus_log2e = vmovq_n_f32(-0x1.715476p0f);
30 const int32x4_t vindex_mask = vmovq_n_s32(INT32_C(0x7FF));
31 const float32x4_t vln2 = vmovq_n_f32(0x1.62E43p-1f);
32 const float32x4_t vc1 = vmovq_n_f32(-0x1.FFFFFEp-1f);
33 const float32x4_t vone = vmovq_n_f32(1.0f);
34 const float32x4_t vdenorm_cutoff = vmovq_n_f32(-0x1.5D589Ep+6f);
35
36 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
37 const float32x4_t vx0123 = vld1q_f32(x); x += 4;
38 const float32x4_t vx4567 = vld1q_f32(x); x += 4;
39
40 const float32x4_t vz0123 = vabsq_f32(vx0123);
41 const float32x4_t vz4567 = vabsq_f32(vx4567);
42
43 float32x4_t vn0123 = vfmaq_f32(vmagic_bias, vz0123, vminus_log2e);
44 float32x4_t vn4567 = vfmaq_f32(vmagic_bias, vz4567, vminus_log2e);
45
46 const int32x4_t ve0123 = vshlq_n_s32(vreinterpretq_s32_f32(vn0123), 12);
47 const int32x4_t ve4567 = vshlq_n_s32(vreinterpretq_s32_f32(vn4567), 12);
48
49 const uint64x2_t vidx0123 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn0123), vindex_mask));
50 const uint64x2_t vidx4567 = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn4567), vindex_mask));
51
52 const uint64_t vidx01 = vgetq_lane_u64(vidx0123, 0);
53 const uint64_t vidx23 = vgetq_lane_u64(vidx0123, 1);
54 float32x2_t vl01 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx01]);
55 float32x2_t vl23 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx23]);
56 const uint64_t vidx45 = vgetq_lane_u64(vidx4567, 0);
57 const uint64_t vidx67 = vgetq_lane_u64(vidx4567, 1);
58 float32x2_t vl45 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx45]);
59 float32x2_t vl67 = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx67]);
60
61 vl01 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx01 >> 32)], vl01, 1);
62 vl23 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx23 >> 32)], vl23, 1);
63 const float32x4_t vl0123 = vcombine_f32(vl01, vl23);
64 vl45 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx45 >> 32)], vl45, 1);
65 vl67 = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx67 >> 32)], vl67, 1);
66 const float32x4_t vl4567 = vcombine_f32(vl45, vl67);
67
68 const float32x4_t vs0123 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl0123), ve0123));
69 const float32x4_t vs4567 = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl4567), ve4567));
70
71 vn0123 = vsubq_f32(vn0123, vmagic_bias);
72 vn4567 = vsubq_f32(vn4567, vmagic_bias);
73
74 float32x4_t vt0123 = vfmaq_f32(vz0123, vn0123, vln2);
75 float32x4_t vt4567 = vfmaq_f32(vz4567, vn4567, vln2);
76
77 const float32x4_t vp0123 = vmulq_f32(vt0123, vc1);
78 const float32x4_t vp4567 = vmulq_f32(vt4567, vc1);
79
80 const float32x4_t vy0123 = vfmaq_f32(vs0123, vs0123, vp0123);
81 const float32x4_t vy4567 = vfmaq_f32(vs4567, vs4567, vp4567);
82
83 const float32x4_t vd0123 = vaddq_f32(vy0123, vone);
84 const float32x4_t vd4567 = vaddq_f32(vy4567, vone);
85
86 float32x4_t vr0123 = vrecpeq_f32(vd0123);
87 float32x4_t vr4567 = vrecpeq_f32(vd4567);
88
89 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
90 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
91
92 vr0123 = vfmaq_f32(vr0123, vr0123, vfmsq_f32(vone, vr0123, vd0123));
93 vr4567 = vfmaq_f32(vr4567, vr4567, vfmsq_f32(vone, vr4567, vd4567));
94
95 float32x4_t vf0123 = vmulq_f32(vy0123, vr0123);
96 float32x4_t vf4567 = vmulq_f32(vy4567, vr4567);
97
98 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
99 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
100
101 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
102 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
103
104 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
105 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
106
107 vst1q_f32(y, vf0123); y += 4;
108 vst1q_f32(y, vf4567); y += 4;
109 }
110 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
111 const float32x4_t vx = vld1q_f32(x); x += 4;
112
113 const float32x4_t vz = vabsq_f32(vx);
114
115 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
116 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
117
118 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
119 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
120 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
121 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
122 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
123 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
124 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
125 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
126
127 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
128 vn = vsubq_f32(vn, vmagic_bias);
129 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
130
131 const float32x4_t vp = vmulq_f32(vt, vc1);
132
133 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
134 const float32x4_t vd = vaddq_f32(vy, vone);
135
136 float32x4_t vr = vrecpeq_f32(vd);
137 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
138 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
139
140 float32x4_t vf = vmulq_f32(vy, vr);
141 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
142 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
143 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
144
145 vst1q_f32(y, vf); y += 4;
146 }
147 if XNN_UNLIKELY(n != 0) {
148 const float32x4_t vx = vld1q_f32(x);
149
150 const float32x4_t vz = vabsq_f32(vx);
151
152 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
153 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
154
155 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
156 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
157 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
158 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
159 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
160 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
161 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
162 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
163
164 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
165 vn = vsubq_f32(vn, vmagic_bias);
166 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
167
168 const float32x4_t vp = vmulq_f32(vt, vc1);
169
170 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
171 const float32x4_t vd = vaddq_f32(vy, vone);
172
173 float32x4_t vr = vrecpeq_f32(vd);
174 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
175 vr = vfmaq_f32(vr, vr, vfmsq_f32(vone, vr, vd));
176
177 float32x4_t vf = vmulq_f32(vy, vr);
178 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
179 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
180 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
181
182 float32x2_t vf_lo = vget_low_f32(vf);
183 if (n & (2 * sizeof(float))) {
184 vst1_f32(y, vf_lo); y += 2;
185 vf_lo = vget_high_f32(vf);
186 }
187 if (n & (1 * sizeof(float))) {
188 vst1_lane_f32(y, vf_lo, 0);
189 }
190 }
191 }
192