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_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_x86_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(&params->scalar.prescale);
30   const v128_t valpha = wasm_v32x4_load_splat(&params->scalar.alpha);
31   const v128_t vbeta = wasm_v32x4_load_splat(&params->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_mul(vx0123, vprescale);
55     const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
56     const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
57     const v128_t vzCDEF = wasm_f32x4_mul(vxCDEF, vprescale);
58     const v128_t vzGHIJ = wasm_f32x4_mul(vxGHIJ, vprescale);
59     const v128_t vzKLMN = wasm_f32x4_mul(vxKLMN, vprescale);
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     const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
84     v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
85     const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
86     v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
87     const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
88     v128_t vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_hi), vzCDEF);
89     const v128_t vsatmCDEF = wasm_f32x4_le(vzCDEF, vsat_cutoff);
90     v128_t vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_hi), vzGHIJ);
91     const v128_t vsatmGHIJ = wasm_f32x4_le(vzGHIJ, vsat_cutoff);
92     v128_t vtKLMN = wasm_f32x4_add(wasm_f32x4_mul(vnKLMN, vminus_ln2_hi), vzKLMN);
93     const v128_t vsatmKLMN = wasm_f32x4_le(vzKLMN, vsat_cutoff);
94 
95     vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
96     vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
97     vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
98     vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
99     vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
100     vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
101     vtCDEF = wasm_f32x4_add(wasm_f32x4_mul(vnCDEF, vminus_ln2_lo), vtCDEF);
102     vsCDEF = wasm_v128_andnot(vsCDEF, vsatmCDEF);
103     vtGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vnGHIJ, vminus_ln2_lo), vtGHIJ);
104     vsGHIJ = wasm_v128_andnot(vsGHIJ, vsatmGHIJ);
105     vtKLMN = wasm_f32x4_add(wasm_f32x4_mul(vnKLMN, vminus_ln2_lo), vtKLMN);
106     vsKLMN = wasm_v128_andnot(vsKLMN, vsatmKLMN);
107 
108     vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
109     v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
110     vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
111     v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
112     vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
113     v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
114     vtCDEF = wasm_v128_andnot(vtCDEF, vsatmCDEF);
115     v128_t vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtCDEF), vc5);
116     vtGHIJ = wasm_v128_andnot(vtGHIJ, vsatmGHIJ);
117     v128_t vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtGHIJ), vc5);
118     vtKLMN = wasm_v128_andnot(vtKLMN, vsatmKLMN);
119     v128_t vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vc6, vtKLMN), vc5);
120 
121     vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
122     vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
123     vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
124     vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc4);
125     vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc4);
126     vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc4);
127 
128     vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
129     vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
130     vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
131     vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc3);
132     vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc3);
133     vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc3);
134 
135     vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
136     vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
137     vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
138     vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vc2);
139     vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vc2);
140     vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vc2);
141 
142     vp0123 = wasm_f32x4_mul(vp0123, vt0123);
143     vp4567 = wasm_f32x4_mul(vp4567, vt4567);
144     vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
145     vpCDEF = wasm_f32x4_mul(vpCDEF, vtCDEF);
146     vpGHIJ = wasm_f32x4_mul(vpGHIJ, vtGHIJ);
147     vpKLMN = wasm_f32x4_mul(vpKLMN, vtKLMN);
148 
149     vt0123 = wasm_f32x4_mul(vt0123, vs0123);
150     vs0123 = wasm_f32x4_sub(vs0123, vone);
151     vt4567 = wasm_f32x4_mul(vt4567, vs4567);
152     vs4567 = wasm_f32x4_sub(vs4567, vone);
153     vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
154     vs89AB = wasm_f32x4_sub(vs89AB, vone);
155     vtCDEF = wasm_f32x4_mul(vtCDEF, vsCDEF);
156     vsCDEF = wasm_f32x4_sub(vsCDEF, vone);
157     vtGHIJ = wasm_f32x4_mul(vtGHIJ, vsGHIJ);
158     vsGHIJ = wasm_f32x4_sub(vsGHIJ, vone);
159     vtKLMN = wasm_f32x4_mul(vtKLMN, vsKLMN);
160     vsKLMN = wasm_f32x4_sub(vsKLMN, vone);
161 
162     vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
163     vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
164     vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
165     vpCDEF = wasm_f32x4_add(wasm_f32x4_mul(vpCDEF, vtCDEF), vtCDEF);
166     vpGHIJ = wasm_f32x4_add(wasm_f32x4_mul(vpGHIJ, vtGHIJ), vtGHIJ);
167     vpKLMN = wasm_f32x4_add(wasm_f32x4_mul(vpKLMN, vtKLMN), vtKLMN);
168 
169     const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
170     const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
171     const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
172     const v128_t veCDEF = wasm_f32x4_mul(wasm_f32x4_add(vpCDEF, vsCDEF), valpha);
173     const v128_t veGHIJ = wasm_f32x4_mul(wasm_f32x4_add(vpGHIJ, vsGHIJ), valpha);
174     const v128_t veKLMN = wasm_f32x4_mul(wasm_f32x4_add(vpKLMN, vsKLMN), valpha);
175 
176     const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
177     vx0123 = wasm_f32x4_mul(vx0123, vbeta);
178     const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
179     vx4567 = wasm_f32x4_mul(vx4567, vbeta);
180     const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
181     vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
182     const v128_t vsignmCDEF = wasm_i32x4_shr(vxCDEF, 31);
183     vxCDEF = wasm_f32x4_mul(vxCDEF, vbeta);
184     const v128_t vsignmGHIJ = wasm_i32x4_shr(vxGHIJ, 31);
185     vxGHIJ = wasm_f32x4_mul(vxGHIJ, vbeta);
186     const v128_t vsignmKLMN = wasm_i32x4_shr(vxKLMN, 31);
187     vxKLMN = wasm_f32x4_mul(vxKLMN, vbeta);
188 
189     const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
190     const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
191     const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
192     const v128_t vyCDEF = wasm_v128_bitselect(veCDEF, vxCDEF, vsignmCDEF);
193     const v128_t vyGHIJ = wasm_v128_bitselect(veGHIJ, vxGHIJ, vsignmGHIJ);
194     const v128_t vyKLMN = wasm_v128_bitselect(veKLMN, vxKLMN, vsignmKLMN);
195 
196     wasm_v128_store(y, vy0123);
197     wasm_v128_store(y + 4, vy4567);
198     wasm_v128_store(y + 8, vy89AB);
199     wasm_v128_store(y + 12, vyCDEF);
200     wasm_v128_store(y + 16, vyGHIJ);
201     wasm_v128_store(y + 20, vyKLMN);
202     y += 24;
203   }
204   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
205     v128_t vx = wasm_v128_load(x);
206     x += 4;
207 
208     const v128_t vz = wasm_f32x4_mul(vx, vprescale);
209 
210     v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
211     v128_t vs = wasm_i32x4_shl(vn, 23);
212     vn = wasm_f32x4_sub(vn, vmagic_bias);
213 
214     v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
215     const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
216     vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
217     vs = wasm_v128_andnot(vs, vsatm);
218     vt = wasm_v128_andnot(vt, vsatm);
219 
220     v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
221     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
222     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
223     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
224     vp = wasm_f32x4_mul(vp, vt);
225 
226     vt = wasm_f32x4_mul(vt, vs);
227     vs = wasm_f32x4_sub(vs, vone);
228     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
229     const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
230 
231     const v128_t vsignm = wasm_i32x4_shr(vx, 31);
232     vx = wasm_f32x4_mul(vx, vbeta);
233     const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
234 
235     wasm_v128_store(y, vy);
236     y += 4;
237   }
238   if XNN_UNLIKELY(n != 0) {
239     v128_t vx = wasm_v128_load(x);
240 
241     const v128_t vz = wasm_f32x4_mul(vx, vprescale);
242 
243     v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
244     v128_t vs = wasm_i32x4_shl(vn, 23);
245     vn = wasm_f32x4_sub(vn, vmagic_bias);
246 
247     v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
248     const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
249     vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
250     vs = wasm_v128_andnot(vs, vsatm);
251     vt = wasm_v128_andnot(vt, vsatm);
252 
253     v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
254     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
255     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
256     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
257     vp = wasm_f32x4_mul(vp, vt);
258 
259     vt = wasm_f32x4_mul(vt, vs);
260     vs = wasm_f32x4_sub(vs, vone);
261     vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
262     const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
263 
264     const v128_t vsignm = wasm_i32x4_shr(vx, 31);
265     vx = wasm_f32x4_mul(vx, vbeta);
266     v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
267 
268     if (n & (2 * sizeof(float))) {
269       *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
270       vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
271       y += 2;
272     }
273     if (n & (1 * sizeof(float))) {
274       *y = wasm_f32x4_extract_lane(vy, 0);
275     }
276   }
277 }
278