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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * 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 const v128_t vxGHIJ = wasm_v128_load(x + 16);
44 x += 20;
45
46 const v128_t vz0123 = wasm_f32x4_abs(vx0123);
47 const v128_t vz4567 = wasm_f32x4_abs(vx4567);
48 const v128_t vz89AB = wasm_f32x4_abs(vx89AB);
49 const v128_t vzCDEF = wasm_f32x4_abs(vxCDEF);
50 const v128_t vzGHIJ = wasm_f32x4_abs(vxGHIJ);
51
52 v128_t vn0123 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz0123, vminus_log2e));
53 v128_t vn4567 = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz4567, vminus_log2e));
54 v128_t vn89AB = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz89AB, vminus_log2e));
55 v128_t vnCDEF = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzCDEF, vminus_log2e));
56 v128_t vnGHIJ = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vzGHIJ, vminus_log2e));
57
58 const v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
59 const v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
60 const v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
61 const v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
62 const v128_t vsGHIJ = wasm_i32x4_shl(vnGHIJ, 23);
63
64 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
65 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
66 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
67 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
68 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
69
70 v128_t vt0123 = wasm_f32x4_add(vz0123, wasm_f32x4_mul(vn0123, vln2_hi));
71 v128_t vt4567 = wasm_f32x4_add(vz4567, wasm_f32x4_mul(vn4567, vln2_hi));
72 v128_t vt89AB = wasm_f32x4_add(vz89AB, wasm_f32x4_mul(vn89AB, vln2_hi));
73 v128_t vtCDEF = wasm_f32x4_add(vzCDEF, wasm_f32x4_mul(vnCDEF, vln2_hi));
74 v128_t vtGHIJ = wasm_f32x4_add(vzGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_hi));
75
76 vt0123 = wasm_f32x4_add(vt0123, wasm_f32x4_mul(vn0123, vln2_lo));
77 vt4567 = wasm_f32x4_add(vt4567, wasm_f32x4_mul(vn4567, vln2_lo));
78 vt89AB = wasm_f32x4_add(vt89AB, wasm_f32x4_mul(vn89AB, vln2_lo));
79 vtCDEF = wasm_f32x4_add(vtCDEF, wasm_f32x4_mul(vnCDEF, vln2_lo));
80 vtGHIJ = wasm_f32x4_add(vtGHIJ, wasm_f32x4_mul(vnGHIJ, vln2_lo));
81
82 v128_t vp0123 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt0123, vc5));
83 v128_t vp4567 = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt4567, vc5));
84 v128_t vp89AB = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt89AB, vc5));
85 v128_t vpCDEF = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtCDEF, vc5));
86 v128_t vpGHIJ = wasm_f32x4_add(vc4, wasm_f32x4_mul(vtGHIJ, vc5));
87
88 vp0123 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt0123, vp0123));
89 vp4567 = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt4567, vp4567));
90 vp89AB = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt89AB, vp89AB));
91 vpCDEF = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtCDEF, vpCDEF));
92 vpGHIJ = wasm_f32x4_add(vc3, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
93
94 vp0123 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt0123, vp0123));
95 vp4567 = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt4567, vp4567));
96 vp89AB = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt89AB, vp89AB));
97 vpCDEF = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtCDEF, vpCDEF));
98 vpGHIJ = wasm_f32x4_add(vc2, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
99
100 vp0123 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt0123, vp0123));
101 vp4567 = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt4567, vp4567));
102 vp89AB = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt89AB, vp89AB));
103 vpCDEF = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtCDEF, vpCDEF));
104 vpGHIJ = wasm_f32x4_add(vc1, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
105
106 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
107 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
108 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
109 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
110 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
111
112 const v128_t ve0123 = wasm_f32x4_add(vs0123, wasm_f32x4_mul(vt0123, vp0123));
113 const v128_t ve4567 = wasm_f32x4_add(vs4567, wasm_f32x4_mul(vt4567, vp4567));
114 const v128_t ve89AB = wasm_f32x4_add(vs89AB, wasm_f32x4_mul(vt89AB, vp89AB));
115 const v128_t veCDEF = wasm_f32x4_add(vsCDEF, wasm_f32x4_mul(vtCDEF, vpCDEF));
116 const v128_t veGHIJ = wasm_f32x4_add(vsGHIJ, wasm_f32x4_mul(vtGHIJ, vpGHIJ));
117
118 const v128_t vd0123 = wasm_f32x4_add(ve0123, vone);
119 const v128_t vd4567 = wasm_f32x4_add(ve4567, vone);
120 const v128_t vd89AB = wasm_f32x4_add(ve89AB, vone);
121 const v128_t vdCDEF = wasm_f32x4_add(veCDEF, vone);
122 const v128_t vdGHIJ = wasm_f32x4_add(veGHIJ, vone);
123
124 v128_t vf0123 = wasm_f32x4_div(ve0123, vd0123);
125 v128_t vf4567 = wasm_f32x4_div(ve4567, vd4567);
126 v128_t vf89AB = wasm_f32x4_div(ve89AB, vd89AB);
127 v128_t vfCDEF = wasm_f32x4_div(veCDEF, vdCDEF);
128 v128_t vfGHIJ = wasm_f32x4_div(veGHIJ, vdGHIJ);
129
130 vf0123 = wasm_v128_andnot(vf0123, wasm_f32x4_gt(vz0123, vdenorm_cutoff));
131 vf4567 = wasm_v128_andnot(vf4567, wasm_f32x4_gt(vz4567, vdenorm_cutoff));
132 vf89AB = wasm_v128_andnot(vf89AB, wasm_f32x4_gt(vz89AB, vdenorm_cutoff));
133 vfCDEF = wasm_v128_andnot(vfCDEF, wasm_f32x4_gt(vzCDEF, vdenorm_cutoff));
134 vfGHIJ = wasm_v128_andnot(vfGHIJ, wasm_f32x4_gt(vzGHIJ, vdenorm_cutoff));
135
136 vf0123 = wasm_v128_bitselect(vf0123, wasm_f32x4_sub(vone, vf0123), wasm_i32x4_shr(vx0123, 31));
137 vf4567 = wasm_v128_bitselect(vf4567, wasm_f32x4_sub(vone, vf4567), wasm_i32x4_shr(vx4567, 31));
138 vf89AB = wasm_v128_bitselect(vf89AB, wasm_f32x4_sub(vone, vf89AB), wasm_i32x4_shr(vx89AB, 31));
139 vfCDEF = wasm_v128_bitselect(vfCDEF, wasm_f32x4_sub(vone, vfCDEF), wasm_i32x4_shr(vxCDEF, 31));
140 vfGHIJ = wasm_v128_bitselect(vfGHIJ, wasm_f32x4_sub(vone, vfGHIJ), wasm_i32x4_shr(vxGHIJ, 31));
141
142 wasm_v128_store(y, vf0123);
143 wasm_v128_store(y + 4, vf4567);
144 wasm_v128_store(y + 8, vf89AB);
145 wasm_v128_store(y + 12, vfCDEF);
146 wasm_v128_store(y + 16, vfGHIJ);
147 y += 20;
148 }
149 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
150 const v128_t vx = wasm_v128_load(x);
151 x += 4;
152
153 const v128_t vz = wasm_f32x4_abs(vx);
154
155 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
156 const v128_t vs = wasm_i32x4_shl(vn, 23);
157 vn = wasm_f32x4_sub(vn, vmagic_bias);
158
159 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
160 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
161
162 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
163 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
164 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
165 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
166
167 vt = wasm_f32x4_mul(vt, vs);
168 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
169 const v128_t vd = wasm_f32x4_add(ve, vone);
170
171 v128_t vf = wasm_f32x4_div(ve, vd);
172 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
173 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
174
175 wasm_v128_store(y, vf);
176 y += 4;
177 }
178 if XNN_UNLIKELY(n != 0) {
179 const v128_t vx = wasm_v128_load(x);
180
181 const v128_t vz = wasm_f32x4_abs(vx);
182
183 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
184 const v128_t vs = wasm_i32x4_shl(vn, 23);
185 vn = wasm_f32x4_sub(vn, vmagic_bias);
186
187 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
188 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
189
190 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
191 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
192 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
193 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
194
195 vt = wasm_f32x4_mul(vt, vs);
196 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
197 const v128_t vd = wasm_f32x4_add(ve, vone);
198
199 v128_t vf = wasm_f32x4_div(ve, vd);
200 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
201 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
202
203 if (n & (2 * sizeof(float))) {
204 *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
205 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
206 y += 2;
207 }
208 if (n & (1 * sizeof(float))) {
209 *y = wasm_f32x4_extract_lane(vf, 0);
210 }
211 }
212 }
213