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_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 >= 12 * sizeof(float); n -= 12 * 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 x += 12;
50
51 const v128_t vz0123 = wasm_f32x4_mul(vx0123, vprescale);
52 const v128_t vz4567 = wasm_f32x4_mul(vx4567, vprescale);
53 const v128_t vz89AB = wasm_f32x4_mul(vx89AB, vprescale);
54
55 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
56 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
57 v128_t vn89AB = wasm_f32x4_add(wasm_f32x4_mul(vz89AB, vlog2e), vmagic_bias);
58
59 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
60 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
61 v128_t vs89AB = wasm_i32x4_shl(vn89AB, 23);
62
63 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
64 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
65 vn89AB = wasm_f32x4_sub(vn89AB, vmagic_bias);
66
67 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
68 const v128_t vsatm0123 = wasm_f32x4_le(vz0123, vsat_cutoff);
69 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
70 const v128_t vsatm4567 = wasm_f32x4_le(vz4567, vsat_cutoff);
71 v128_t vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_hi), vz89AB);
72 const v128_t vsatm89AB = wasm_f32x4_le(vz89AB, vsat_cutoff);
73
74 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
75 vs0123 = wasm_v128_andnot(vs0123, vsatm0123);
76 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
77 vs4567 = wasm_v128_andnot(vs4567, vsatm4567);
78 vt89AB = wasm_f32x4_add(wasm_f32x4_mul(vn89AB, vminus_ln2_lo), vt89AB);
79 vs89AB = wasm_v128_andnot(vs89AB, vsatm89AB);
80
81 vt0123 = wasm_v128_andnot(vt0123, vsatm0123);
82 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
83 vt4567 = wasm_v128_andnot(vt4567, vsatm4567);
84 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
85 vt89AB = wasm_v128_andnot(vt89AB, vsatm89AB);
86 v128_t vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt89AB), vc5);
87
88 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
89 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
90 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc4);
91
92 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
93 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
94 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc3);
95
96 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
97 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
98 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vc2);
99
100 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
101 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
102 vp89AB = wasm_f32x4_mul(vp89AB, vt89AB);
103
104 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
105 vs0123 = wasm_f32x4_sub(vs0123, vone);
106 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
107 vs4567 = wasm_f32x4_sub(vs4567, vone);
108 vt89AB = wasm_f32x4_mul(vt89AB, vs89AB);
109 vs89AB = wasm_f32x4_sub(vs89AB, vone);
110
111 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
112 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
113 vp89AB = wasm_f32x4_add(wasm_f32x4_mul(vp89AB, vt89AB), vt89AB);
114
115 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
116 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
117 const v128_t ve89AB = wasm_f32x4_mul(wasm_f32x4_add(vp89AB, vs89AB), valpha);
118
119 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
120 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
121 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
122 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
123 const v128_t vsignm89AB = wasm_i32x4_shr(vx89AB, 31);
124 vx89AB = wasm_f32x4_mul(vx89AB, vbeta);
125
126 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
127 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
128 const v128_t vy89AB = wasm_v128_bitselect(ve89AB, vx89AB, vsignm89AB);
129
130 wasm_v128_store(y, vy0123);
131 wasm_v128_store(y + 4, vy4567);
132 wasm_v128_store(y + 8, vy89AB);
133 y += 12;
134 }
135 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
136 v128_t vx = wasm_v128_load(x);
137 x += 4;
138
139 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
140
141 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
142 v128_t vs = wasm_i32x4_shl(vn, 23);
143 vn = wasm_f32x4_sub(vn, vmagic_bias);
144
145 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
146 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
147 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
148 vs = wasm_v128_andnot(vs, vsatm);
149 vt = wasm_v128_andnot(vt, vsatm);
150
151 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
152 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
153 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
154 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
155 vp = wasm_f32x4_mul(vp, vt);
156
157 vt = wasm_f32x4_mul(vt, vs);
158 vs = wasm_f32x4_sub(vs, vone);
159 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
160 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
161
162 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
163 vx = wasm_f32x4_mul(vx, vbeta);
164 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
165
166 wasm_v128_store(y, vy);
167 y += 4;
168 }
169 if XNN_UNLIKELY(n != 0) {
170 v128_t vx = wasm_v128_load(x);
171
172 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
173
174 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
175 v128_t vs = wasm_i32x4_shl(vn, 23);
176 vn = wasm_f32x4_sub(vn, vmagic_bias);
177
178 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
179 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
180 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
181 vs = wasm_v128_andnot(vs, vsatm);
182 vt = wasm_v128_andnot(vt, vsatm);
183
184 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
185 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
186 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
187 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
188 vp = wasm_f32x4_mul(vp, vt);
189
190 vt = wasm_f32x4_mul(vt, vs);
191 vs = wasm_f32x4_sub(vs, vone);
192 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
193 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
194
195 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
196 vx = wasm_f32x4_mul(vx, vbeta);
197 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
198
199 if (n & (2 * sizeof(float))) {
200 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
201 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
202 y += 2;
203 }
204 if (n & (1 * sizeof(float))) {
205 *y = wasm_f32x4_extract_lane(vy, 0);
206 }
207 }
208 }
209