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_x12(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_x12(
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_OOB_READS
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_v128_load64_splat(params->wasmsimd_rr2_p6.prescale);
30 const v128_t valpha = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.alpha);
31 const v128_t vbeta = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.beta);
32 const v128_t vsat_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.sat_cutoff);
33 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.magic_bias);
34 const v128_t vlog2e = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.log2e);
35 const v128_t vminus_ln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.minus_ln2_hi);
36 const v128_t vminus_ln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.minus_ln2_lo);
37 const v128_t vc6 = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.c6);
38 const v128_t vc5 = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.c5);
39 const v128_t vc4 = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.c4);
40 const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.c3);
41 const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.c2);
42 const v128_t vone = wasm_v128_load64_splat(params->wasmsimd_rr2_p6.one);
43
44 for (; n >= 12 * sizeof(float); n -= 12 * sizeof(float)) {
45 v128_t vx0123 = wasm_v128_load(x);
46 v128_t vx4567 = wasm_v128_load(x + 4);
47 v128_t vx89AB = wasm_v128_load(x + 8);
48 x += 12;
49
50 const v128_t vz0123 = wasm_f32x4_mul(vx0123, vprescale);
51 const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
52 const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
53
54 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
55 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
56 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
57
58 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
59 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
60 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
61
62 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
63 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
64 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
65
66 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
67 const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
68 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
69 const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
70 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
71 const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
72
73 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
74 vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
75 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
76 vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
77 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
78 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
79
80 vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
81 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
82 vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
83 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
84 vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
85 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), 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
91 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
92 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
93 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
94
95 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
96 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
97 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
98
99 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
100 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
101 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
102
103 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
104 vs0123 = wasm_f32x4_sub(vs0123, vone);
105 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
106 vs4567 = wasm_f32x4_sub(vs4567, vone);
107 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
108 vs89AB = wasm_f32x4_sub(vs89AB, vone);
109
110 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
111 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
112 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
113
114 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
115 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
116 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
117
118 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
119 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
120 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
121 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
122 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
123 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
124
125 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
126 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
127 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
128
129 wasm_v128_store(y, vy0123);
130 wasm_v128_store(y + 4, vy4567);
131 wasm_v128_store(y + 8, vy89AB);
132 y += 12;
133 }
134 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
135 v128_t vx = wasm_v128_load(x);
136 x += 4;
137
138 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
139
140 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
141 v128_t vs = wasm_i32x4_shl(vn, 23);
142 vn = wasm_f32x4_sub(vn, vmagic_bias);
143
144 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
145 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
146 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
147 vs = wasm_v128_andnot(vs, vsatm);
148 vt = wasm_v128_andnot(vt, vsatm);
149
150 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
151 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
152 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
153 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
154 vp = wasm_f32x4_mul(vp, vt);
155
156 vt = wasm_f32x4_mul(vt, vs);
157 vs = wasm_f32x4_sub(vs, vone);
158 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
159 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
160
161 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
162 vx = wasm_f32x4_mul(vx, vbeta);
163 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
164
165 wasm_v128_store(y, vy);
166 y += 4;
167 }
168 if XNN_UNLIKELY(n != 0) {
169 v128_t vx = wasm_v128_load(x);
170
171 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
172
173 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
174 v128_t vs = wasm_i32x4_shl(vn, 23);
175 vn = wasm_f32x4_sub(vn, vmagic_bias);
176
177 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
178 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
179 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
180 vs = wasm_v128_andnot(vs, vsatm);
181 vt = wasm_v128_andnot(vt, vsatm);
182
183 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
184 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
185 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
186 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
187 vp = wasm_f32x4_mul(vp, vt);
188
189 vt = wasm_f32x4_mul(vt, vs);
190 vs = wasm_f32x4_sub(vs, vone);
191 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
192 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
193
194 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
195 vx = wasm_f32x4_mul(vx, vbeta);
196 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
197
198 if (n & (2 * sizeof(float))) {
199 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
200 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
201 y += 2;
202 }
203 if (n & (1 * sizeof(float))) {
204 *y = wasm_f32x4_extract_lane(vy, 0);
205 }
206 }
207 }
208