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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 v128_t vxKLMN = wasm_v128_load(x + 20);
52 x += 24;
53
54 const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff);
55 const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff);
56 const v128_t vz89AB = wasm_f32x4_max(wasm_f32x4_mul(vx89AB, vprescale), vsat_cutoff);
57 const v128_t vzCDEF = wasm_f32x4_max(wasm_f32x4_mul(vxCDEF, vprescale), vsat_cutoff);
58 const v128_t vzGHIJ = wasm_f32x4_max(wasm_f32x4_mul(vxGHIJ, vprescale), vsat_cutoff);
59 const v128_t vzKLMN = wasm_f32x4_max(wasm_f32x4_mul(vxKLMN, vprescale), vsat_cutoff);
60
61 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
62 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
63 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
64 v128_t vnCDEF = wasm_f32x4_add(wasm_f32x4_mul(vzCDEF, vlog2e), vmagic_bias);
65 v128_t vnGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vzGHIJ, vlog2e), vmagic_bias);
66 v128_t vnKLMN = wasm_f32x4_add(wasm_f32x4_mul(vzKLMN, vlog2e), vmagic_bias);
67
68 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
69 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
70 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
71 v128_t vsCDEF = wasm_i32x4_shl(vnCDEF, 23);
72 v128_t vsGHIJ = wasm_i32x4_shl(vnGHIJ, 23);
73 v128_t vsKLMN = wasm_i32x4_shl(vnKLMN, 23);
74
75 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
76 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
77 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
78 vnCDEF = wasm_f32x4_sub(vnCDEF, vmagic_bias);
79 vnGHIJ = wasm_f32x4_sub(vnGHIJ, vmagic_bias);
80 vnKLMN = wasm_f32x4_sub(vnKLMN, vmagic_bias);
81
82 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
83 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
84 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
85 v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
86 v128_t vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi), vzGHIJ);
87 v128_t vtKLMN = wasm_f32x4_add(wasm_f32x4_mul(vnKLMN, vminus_ln2_hi), vzKLMN);
88
89 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
90 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
91 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
92 vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
93 vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo), vtGHIJ);
94 vtKLMN = wasm_f32x4_add(wasm_f32x4_mul(vnKLMN, vminus_ln2_lo), vtKLMN);
95
96 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
97 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
98 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
99 v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5);
100 v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5);
101 v128_t vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtKLMN), vc5);
102
103 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
104 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
105 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
106 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4);
107 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4);
108 vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc4);
109
110 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
111 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
112 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
113 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3);
114 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3);
115 vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc3);
116
117 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
118 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
119 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
120 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2);
121 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2);
122 vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc2);
123
124 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
125 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
126 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
127 vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
128 vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ);
129 vpKLMN = wasm_f32x4_mul(vpKLMN, vtKLMN);
130
131 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
132 vs0123 = wasm_f32x4_sub(vs0123, vone);
133 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
134 vs4567 = wasm_f32x4_sub(vs4567, vone);
135 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
136 vs89AB = wasm_f32x4_sub(vs89AB, vone);
137 vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
138 vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
139 vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
140 vsGHIJ = wasm_f32x4_sub(vsGHIJ, vone);
141 vtKLMN = wasm_f32x4_mul(vtKLMN, vsKLMN);
142 vsKLMN = wasm_f32x4_sub(vsKLMN, vone);
143
144 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
145 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
146 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
147 vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
148 vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ);
149 vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vtKLMN);
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 const v128_t veKLMN = wasm_f32x4_mul(wasm_f32x4_add(vpKLMN, vsKLMN), valpha);
157
158 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
159 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
160 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
161 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
162 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
163 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
164 const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
165 vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
166 const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31);
167 vxGHIJ = wasm_f32x4_mul(vxGHIJ, vbeta);
168 const v128_t vsignmKLMN = wasm_i32x4_shr(vxKLMN, 31);
169 vxKLMN = wasm_f32x4_mul(vxKLMN, vbeta);
170
171 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
172 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
173 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
174 const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
175 const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ);
176 const v128_t vyKLMN = wasm_v128_bitselect(veKLMN, vxKLMN, vsignmKLMN);
177
178 wasm_v128_store(y, vy0123);
179 wasm_v128_store(y + 4, vy4567);
180 wasm_v128_store(y + 8, vy89AB);
181 wasm_v128_store(y + 12, vyCDEF);
182 wasm_v128_store(y + 16, vyGHIJ);
183 wasm_v128_store(y + 20, vyKLMN);
184 y += 24;
185 }
186 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
187 v128_t vx = wasm_v128_load(x);
188 x += 4;
189
190 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
191
192 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
193 v128_t vs = wasm_i32x4_shl(vn, 23);
194 vn = wasm_f32x4_sub(vn, vmagic_bias);
195
196 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
197 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
198
199 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
200 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
201 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
202 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
203 vp = wasm_f32x4_mul(vp, vt);
204
205 vt = wasm_f32x4_mul(vt, vs);
206 vs = wasm_f32x4_sub(vs, vone);
207 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
208 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
209
210 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
211 vx = wasm_f32x4_mul(vx, vbeta);
212 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
213
214 wasm_v128_store(y, vy);
215 y += 4;
216 }
217 if XNN_UNLIKELY(n != 0) {
218 v128_t vx = wasm_v128_load(x);
219
220 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
221
222 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
223 v128_t vs = wasm_i32x4_shl(vn, 23);
224 vn = wasm_f32x4_sub(vn, vmagic_bias);
225
226 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
227 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
228
229 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
230 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
231 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
232 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
233 vp = wasm_f32x4_mul(vp, vt);
234
235 vt = wasm_f32x4_mul(vt, vs);
236 vs = wasm_f32x4_sub(vs, vone);
237 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
238 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
239
240 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
241 vx = wasm_f32x4_mul(vx, vbeta);
242 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
243
244 if (n & (2 * sizeof(float))) {
245 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
246 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
247 y += 2;
248 }
249 if (n & (1 * sizeof(float))) {
250 *y = wasm_f32x4_extract_lane(vy, 0);
251 }
252 }
253 }
254