1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/wasmsimd-rr2-p6.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/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const v128_t vprescale = wasm_v32x4_load_splat(¶ms->scalar.prescale);
30 const v128_t valpha = wasm_v32x4_load_splat(¶ms->scalar.alpha);
31 const v128_t vbeta = wasm_v32x4_load_splat(¶ms->scalar.beta);
32
33 const v128_t vsat_cutoff = wasm_f32x4_splat(-0x1.154246p+4f);
34 const v128_t vmagic_bias = wasm_f32x4_splat(0x1.8000FEp23f);
35 const v128_t vlog2e = wasm_f32x4_splat(0x1.715476p+0f);
36 const v128_t vminus_ln2_hi = wasm_f32x4_splat(-0x1.62E440p-1f);
37 const v128_t vminus_ln2_lo = wasm_f32x4_splat(0x1.0105C6p-21f);
38 const v128_t vc6 = wasm_f32x4_splat(0x1.6b7338p-10f);
39 const v128_t vc5 = wasm_f32x4_splat(0x1.12278Ep-7f);
40 const v128_t vc4 = wasm_f32x4_splat(0x1.555716p-5f);
41 const v128_t vc3 = wasm_f32x4_splat(0x1.5554B0p-3f);
42 const v128_t vc2 = wasm_f32x4_splat(0x1.FFFFFEp-2f);
43 const v128_t vone = wasm_f32x4_splat(1.0f);
44
45 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
46 v128_t vx0123 = wasm_v128_load(x);
47 v128_t vx4567 = wasm_v128_load(x + 4);
48 v128_t vx89AB = wasm_v128_load(x + 8);
49 v128_t vxCDEF = wasm_v128_load(x + 12);
50 x += 16;
51
52 const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff);
53 const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff);
54 const v128_t vz89AB = wasm_f32x4_max(wasm_f32x4_mul(vx89AB, vprescale), vsat_cutoff);
55 const v128_t vzCDEF = wasm_f32x4_max(wasm_f32x4_mul(vxCDEF, vprescale), vsat_cutoff);
56
57 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
58 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
59 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
60 v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias);
61
62 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
63 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
64 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
65 v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
66
67 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
68 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
69 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
70 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
71
72 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
73 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
74 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
75 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
76
77 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
78 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
79 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
80 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
81
82 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
83 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
84 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
85 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5);
86
87 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
88 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
89 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
90 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4);
91
92 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
93 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
94 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
95 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3);
96
97 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
98 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
99 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
100 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2);
101
102 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
103 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
104 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
105 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
106
107 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
108 vs0123 = wasm_f32x4_sub(vs0123, vone);
109 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
110 vs4567 = wasm_f32x4_sub(vs4567, vone);
111 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
112 vs89AB = wasm_f32x4_sub(vs89AB, vone);
113 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
114 vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
115
116 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
117 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
118 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
119 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
120
121 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
122 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
123 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
124 const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
125
126 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
127 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
128 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
129 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
130 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
131 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
132 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
133 vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
134
135 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
136 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
137 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
138 const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
139
140 wasm_v128_store(y, vy0123);
141 wasm_v128_store(y + 4, vy4567);
142 wasm_v128_store(y + 8, vy89AB);
143 wasm_v128_store(y + 12, vyCDEF);
144 y += 16;
145 }
146 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
147 v128_t vx = wasm_v128_load(x);
148 x += 4;
149
150 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
151
152 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
153 v128_t vs = wasm_i32x4_shl(vn, 23);
154 vn = wasm_f32x4_sub(vn, vmagic_bias);
155
156 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
157 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
158
159 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
160 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
161 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
162 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
163 vp = wasm_f32x4_mul(vp, vt);
164
165 vt = wasm_f32x4_mul(vt, vs);
166 vs = wasm_f32x4_sub(vs, vone);
167 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
168 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
169
170 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
171 vx = wasm_f32x4_mul(vx, vbeta);
172 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
173
174 wasm_v128_store(y, vy);
175 y += 4;
176 }
177 if XNN_UNLIKELY(n != 0) {
178 v128_t vx = wasm_v128_load(x);
179
180 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
181
182 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
183 v128_t vs = wasm_i32x4_shl(vn, 23);
184 vn = wasm_f32x4_sub(vn, vmagic_bias);
185
186 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
187 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
188
189 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
190 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
191 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
192 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
193 vp = wasm_f32x4_mul(vp, vt);
194
195 vt = wasm_f32x4_mul(vt, vs);
196 vs = wasm_f32x4_sub(vs, vone);
197 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
198 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
199
200 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
201 vx = wasm_f32x4_mul(vx, vbeta);
202 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
203
204 if (n & (2 * sizeof(float))) {
205 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
206 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
207 y += 2;
208 }
209 if (n & (1 * sizeof(float))) {
210 *y = wasm_f32x4_extract_lane(vy, 0);
211 }
212 }
213 }
214