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_div_x8(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_div_x8(
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 >= 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 vf0123 = vdivq_f32(vy0123, vd0123);
87 float32x4_t vf4567 = vdivq_f32(vy4567, vd4567);
88
89 vf0123 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf0123), vcagtq_f32(vx0123, vdenorm_cutoff)));
90 vf4567 = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf4567), vcagtq_f32(vx4567, vdenorm_cutoff)));
91
92 const uint32x4_t vm0123 = vcltq_f32(vx0123, vmovq_n_f32(0.0f));
93 const uint32x4_t vm4567 = vcltq_f32(vx4567, vmovq_n_f32(0.0f));
94
95 vf0123 = vbslq_f32(vm0123, vf0123, vsubq_f32(vone, vf0123));
96 vf4567 = vbslq_f32(vm4567, vf4567, vsubq_f32(vone, vf4567));
97
98 vst1q_f32(y, vf0123); y += 4;
99 vst1q_f32(y, vf4567); y += 4;
100 }
101 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
102 const float32x4_t vx = vld1q_f32(x); x += 4;
103
104 const float32x4_t vz = vabsq_f32(vx);
105
106 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
107 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
108
109 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
110 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
111 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
112 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
113 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
114 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
115 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
116 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
117
118 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
119 vn = vsubq_f32(vn, vmagic_bias);
120 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
121
122 const float32x4_t vp = vmulq_f32(vt, vc1);
123
124 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
125 const float32x4_t vd = vaddq_f32(vy, vone);
126
127 float32x4_t vf = vdivq_f32(vy, vd);
128 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
129 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
130 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
131
132 vst1q_f32(y, vf); y += 4;
133 }
134 if XNN_UNLIKELY(n != 0) {
135 const float32x4_t vx = vld1q_f32(x);
136
137 const float32x4_t vz = vabsq_f32(vx);
138
139 float32x4_t vn = vfmaq_f32(vmagic_bias, vz, vminus_log2e);
140 const int32x4_t ve = vshlq_n_s32(vreinterpretq_s32_f32(vn), 12);
141
142 const uint64x2_t vidx = vreinterpretq_u64_s32(vandq_s32(vreinterpretq_s32_f32(vn), vindex_mask));
143 const uint64_t vidx_lo = vgetq_lane_u64(vidx, 0);
144 const uint64_t vidx_hi = vgetq_lane_u64(vidx, 1);
145 float32x2_t vl_lo = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_lo]);
146 float32x2_t vl_hi = vld1_dup_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) vidx_hi]);
147 vl_lo = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_lo >> 32)], vl_lo, 1);
148 vl_hi = vld1_lane_f32(&xnn_table_exp2minus_k_over_2048[(uint32_t) (vidx_hi >> 32)], vl_hi, 1);
149 const float32x4_t vl = vcombine_f32(vl_lo, vl_hi);
150
151 const float32x4_t vs = vreinterpretq_f32_s32(vaddq_s32(vreinterpretq_s32_f32(vl), ve));
152 vn = vsubq_f32(vn, vmagic_bias);
153 float32x4_t vt = vfmaq_f32(vz, vn, vln2);
154
155 const float32x4_t vp = vmulq_f32(vt, vc1);
156
157 const float32x4_t vy = vfmaq_f32(vs, vs, vp);
158 const float32x4_t vd = vaddq_f32(vy, vone);
159
160 float32x4_t vf = vdivq_f32(vy, vd);
161 vf = vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(vf), vcagtq_f32(vx, vdenorm_cutoff)));
162 const uint32x4_t vm = vcltq_f32(vx, vmovq_n_f32(0.0f));
163 vf = vbslq_f32(vm, vf, vsubq_f32(vone, vf));
164
165 float32x2_t vf_lo = vget_low_f32(vf);
166 if (n & (2 * sizeof(float))) {
167 vst1_f32(y, vf_lo); y += 2;
168 vf_lo = vget_high_f32(vf);
169 }
170 if (n & (1 * sizeof(float))) {
171 vst1_lane_f32(y, vf_lo, 0);
172 }
173 }
174 }
175