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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * 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 v128_t vxGHIJ = wasm_v128_load(x + 16);
51 x += 20;
52
53 const v128_t vz0123 = wasm_f32x4_mul(vx0123, vprescale);
54 const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
55 const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
56 const v128_t vzCDEF = wasm_f32x4_mul(vxCDEF, vprescale);
57 const v128_t vzGHIJ = wasm_f32x4_mul(vxGHIJ, vprescale);
58
59 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
60 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
61 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
62 v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias);
63 v128_t vnGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vzGHIJ, vlog2e), vmagic_bias);
64
65 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
66 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
67 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
68 v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
69 v128_t vsGHIJ = wasm_i32x4_shl(vnGHIJ, 23);
70
71 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
72 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
73 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
74 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
75 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
76
77 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
78 const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
79 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
80 const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
81 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
82 const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
83 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
84 const v128_t vsatmCDEF = wasm_f32x4_le(vzCDEF, vsat_cutoff);
85 v128_t vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi), vzGHIJ);
86 const v128_t vsatmGHIJ = wasm_f32x4_le(vzGHIJ, vsat_cutoff);
87
88 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
89 vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
90 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
91 vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
92 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
93 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
94 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
95 vsCDEF = wasm_v128_andnot(vsCDEF, vsatmCDEF);
96 vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo), vtGHIJ);
97 vsGHIJ = wasm_v128_andnot(vsGHIJ, vsatmGHIJ);
98
99 vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
100 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
101 vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
102 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
103 vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
104 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
105 vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF);
106 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5);
107 vtGHIJ = wasm_v128_andnot(vtGHIJ, vsatmGHIJ);
108 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5);
109
110 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
111 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
112 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
113 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4);
114 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4);
115
116 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
117 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
118 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
119 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3);
120 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3);
121
122 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
123 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
124 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
125 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2);
126 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2);
127
128 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
129 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
130 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
131 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
132 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ);
133
134 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
135 vs0123 = wasm_f32x4_sub(vs0123, vone);
136 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
137 vs4567 = wasm_f32x4_sub(vs4567, vone);
138 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
139 vs89AB = wasm_f32x4_sub(vs89AB, vone);
140 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
141 vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
142 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
143 vsGHIJ = wasm_f32x4_sub(vsGHIJ, vone);
144
145 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
146 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
147 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
148 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
149 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ);
150
151 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
152 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
153 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
154 const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
155 const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha);
156
157 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
158 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
159 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
160 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
161 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
162 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
163 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
164 vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
165 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31);
166 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vbeta);
167
168 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
169 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
170 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
171 const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
172 const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ);
173
174 wasm_v128_store(y, vy0123);
175 wasm_v128_store(y + 4, vy4567);
176 wasm_v128_store(y + 8, vy89AB);
177 wasm_v128_store(y + 12, vyCDEF);
178 wasm_v128_store(y + 16, vyGHIJ);
179 y += 20;
180 }
181 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
182 v128_t vx = wasm_v128_load(x);
183 x += 4;
184
185 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
186
187 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
188 v128_t vs = wasm_i32x4_shl(vn, 23);
189 vn = wasm_f32x4_sub(vn, vmagic_bias);
190
191 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
192 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
193 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
194 vs = wasm_v128_andnot(vs, vsatm);
195 vt = wasm_v128_andnot(vt, vsatm);
196
197 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
198 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
199 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
200 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
201 vp = wasm_f32x4_mul(vp, vt);
202
203 vt = wasm_f32x4_mul(vt, vs);
204 vs = wasm_f32x4_sub(vs, vone);
205 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
206 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
207
208 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
209 vx = wasm_f32x4_mul(vx, vbeta);
210 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
211
212 wasm_v128_store(y, vy);
213 y += 4;
214 }
215 if XNN_UNLIKELY(n != 0) {
216 v128_t vx = wasm_v128_load(x);
217
218 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
219
220 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
221 v128_t vs = wasm_i32x4_shl(vn, 23);
222 vn = wasm_f32x4_sub(vn, vmagic_bias);
223
224 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
225 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
226 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
227 vs = wasm_v128_andnot(vs, vsatm);
228 vt = wasm_v128_andnot(vt, vsatm);
229
230 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
231 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
232 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
233 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
234 vp = wasm_f32x4_mul(vp, vt);
235
236 vt = wasm_f32x4_mul(vt, vs);
237 vs = wasm_f32x4_sub(vs, vone);
238 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
239 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
240
241 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
242 vx = wasm_f32x4_mul(vx, vbeta);
243 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
244
245 if (n & (2 * sizeof(float))) {
246 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
247 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
248 y += 2;
249 }
250 if (n & (1 * sizeof(float))) {
251 *y = wasm_f32x4_extract_lane(vy, 0);
252 }
253 }
254 }
255