1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/wasmsimd-rr2-p5-div.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <wasm_simd128.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x16(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n % sizeof(float) == 0);
25
26 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.magic_bias);
27 const v128_t vminus_log2e = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_log2e);
28 const v128_t vln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.ln2_hi);
29 const v128_t vln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.ln2_lo);
30 const v128_t vc5 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c5);
31 const v128_t vc4 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c4);
32 const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c3);
33 const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c2);
34 const v128_t vc1 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c1);
35 const v128_t vone = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.one);
36 const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.denorm_cutoff);
37
38 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
39 const v128_t vx0123 = wasm_v128_load(x);
40 const v128_t vx4567 = wasm_v128_load(x + 4);
41 const v128_t vx89AB = wasm_v128_load(x + 8);
42 const v128_t vxCDEF = wasm_v128_load(x + 12);
43 x += 16;
44
45 const v128_t vz0123 = wasm_f32x4_abs(vx0123);
46 const v128_t vz4567 = wasm_f32x4_abs(vx4567);
47 const v128_t vz89AB = wasm_f32x4_abs(vx89AB);
48 const v128_t vzCDEF = wasm_f32x4_abs(vxCDEF);
49
50 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e));
51 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e));
52 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e));
53 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e));
54
55 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
56 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
57 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
58 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
59
60 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
61 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
62 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
63 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
64
65 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi));
66 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi));
67 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi));
68 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi));
69
70 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo));
71 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo));
72 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo));
73 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo));
74
75 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt0123, vc5));
76 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt4567, vc5));
77 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt89AB, vc5));
78 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtCDEF, vc5));
79
80 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt0123, vp0123));
81 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt4567, vp4567));
82 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt89AB, vp89AB));
83 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtCDEF, vpCDEF));
84
85 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt0123, vp0123));
86 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt4567, vp4567));
87 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt89AB, vp89AB));
88 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtCDEF, vpCDEF));
89
90 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt0123, vp0123));
91 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt4567, vp4567));
92 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt89AB, vp89AB));
93 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtCDEF, vpCDEF));
94
95 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
96 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
97 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
98 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
99
100 const v128_t ve0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
101 const v128_t ve4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
102 const v128_t ve89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
103 const v128_t veCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
104
105 const v128_t vd0123 = wasm_f32x4_add(ve0123, vone);
106 const v128_t vd4567 = wasm_f32x4_add(ve4567, vone);
107 const v128_t vd89AB = wasm_f32x4_add(ve89AB, vone);
108 const v128_t vdCDEF = wasm_f32x4_add(veCDEF, vone);
109
110 v128_t vf0123 = wasm_f32x4_div(ve0123, vd0123);
111 v128_t vf4567 = wasm_f32x4_div(ve4567, vd4567);
112 v128_t vf89AB = wasm_f32x4_div(ve89AB, vd89AB);
113 v128_t vfCDEF = wasm_f32x4_div(veCDEF, vdCDEF);
114
115 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_gt(vz0123, vdenorm_cutoff));
116 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_gt(vz4567, vdenorm_cutoff));
117 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_gt(vz89AB, vdenorm_cutoff));
118 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_gt(vzCDEF, vdenorm_cutoff));
119
120 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));
121 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));
122 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));
123 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));
124
125 wasm_v128_store(y, vf0123);
126 wasm_v128_store(y + 4, vf4567);
127 wasm_v128_store(y + 8, vf89AB);
128 wasm_v128_store(y + 12, vfCDEF);
129 y += 16;
130 }
131 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
132 const v128_t vx = wasm_v128_load(x);
133 x += 4;
134
135 const v128_t vz = wasm_f32x4_abs(vx);
136
137 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
138 const v128_t vs = wasm_i32x4_shl(vn, 23);
139 vn = wasm_f32x4_sub(vn, vmagic_bias);
140
141 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
142 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
143
144 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
145 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
146 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
147 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
148
149 vt = wasm_f32x4_mul(vt, vs);
150 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
151 const v128_t vd = wasm_f32x4_add(ve, vone);
152
153 v128_t vf = wasm_f32x4_div(ve, vd);
154 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
155 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
156
157 wasm_v128_store(y, vf);
158 y += 4;
159 }
160 if XNN_UNLIKELY(n != 0) {
161 const v128_t vx = wasm_v128_load(x);
162
163 const v128_t vz = wasm_f32x4_abs(vx);
164
165 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
166 const v128_t vs = wasm_i32x4_shl(vn, 23);
167 vn = wasm_f32x4_sub(vn, vmagic_bias);
168
169 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
170 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
171
172 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
173 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
174 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
175 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
176
177 vt = wasm_f32x4_mul(vt, vs);
178 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
179 const v128_t vd = wasm_f32x4_add(ve, vone);
180
181 v128_t vf = wasm_f32x4_div(ve, vd);
182 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
183 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
184
185 if (n & (2 * sizeof(float))) {
186 *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
187 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
188 y += 2;
189 }
190 if (n & (1 * sizeof(float))) {
191 *y = wasm_f32x4_extract_lane(vf, 0);
192 }
193 }
194 }
195