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_x86_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_x86_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_mul(vx0123, vprescale);
53 const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
54 const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
55 const v128_t vzCDEF = wasm_f32x4_mul(vxCDEF, vprescale);
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 const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
74 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
75 const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
76 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
77 const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
78 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
79 const v128_t vsatmCDEF = wasm_f32x4_le(vzCDEF, vsat_cutoff);
80
81 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
82 vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
83 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
84 vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
85 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
86 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
87 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
88 vsCDEF = wasm_v128_andnot(vsCDEF, vsatmCDEF);
89
90 vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
91 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
92 vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
93 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
94 vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
95 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
96 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF);
97 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5);
98
99 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
100 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
101 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
102 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4);
103
104 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
105 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
106 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
107 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3);
108
109 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
110 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
111 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
112 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2);
113
114 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
115 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
116 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
117 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
118
119 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
120 vs0123 = wasm_f32x4_sub(vs0123, vone);
121 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
122 vs4567 = wasm_f32x4_sub(vs4567, vone);
123 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
124 vs89AB = wasm_f32x4_sub(vs89AB, vone);
125 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
126 vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
127
128 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
129 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
130 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
131 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
132
133 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
134 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
135 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
136 const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
137
138 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
139 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
140 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
141 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
142 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
143 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
144 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
145 vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
146
147 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
148 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
149 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
150 const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
151
152 wasm_v128_store(y, vy0123);
153 wasm_v128_store(y + 4, vy4567);
154 wasm_v128_store(y + 8, vy89AB);
155 wasm_v128_store(y + 12, vyCDEF);
156 y += 16;
157 }
158 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
159 v128_t vx = wasm_v128_load(x);
160 x += 4;
161
162 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
163
164 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
165 v128_t vs = wasm_i32x4_shl(vn, 23);
166 vn = wasm_f32x4_sub(vn, vmagic_bias);
167
168 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
169 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
170 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
171 vs = wasm_v128_andnot(vs, vsatm);
172 vt = wasm_v128_andnot(vt, vsatm);
173
174 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
175 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
176 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
177 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
178 vp = wasm_f32x4_mul(vp, vt);
179
180 vt = wasm_f32x4_mul(vt, vs);
181 vs = wasm_f32x4_sub(vs, vone);
182 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
183 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
184
185 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
186 vx = wasm_f32x4_mul(vx, vbeta);
187 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
188
189 wasm_v128_store(y, vy);
190 y += 4;
191 }
192 if XNN_UNLIKELY(n != 0) {
193 v128_t vx = wasm_v128_load(x);
194
195 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
196
197 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
198 v128_t vs = wasm_i32x4_shl(vn, 23);
199 vn = wasm_f32x4_sub(vn, vmagic_bias);
200
201 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
202 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
203 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
204 vs = wasm_v128_andnot(vs, vsatm);
205 vt = wasm_v128_andnot(vt, vsatm);
206
207 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
208 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
209 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
210 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
211 vp = wasm_f32x4_mul(vp, vt);
212
213 vt = wasm_f32x4_mul(vt, vs);
214 vs = wasm_f32x4_sub(vs, vone);
215 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
216 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
217
218 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
219 vx = wasm_f32x4_mul(vx, vbeta);
220 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
221
222 if (n & (2 * sizeof(float))) {
223 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
224 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
225 y += 2;
226 }
227 if (n & (1 * sizeof(float))) {
228 *y = wasm_f32x4_extract_lane(vy, 0);
229 }
230 }
231 }
232