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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
46 v128_t vx0123 = wasm_v128_load(x);
47 v128_t vx4567 = wasm_v128_load(x + 4);
48 x += 8;
49
50 const v128_t vz0123 = wasm_f32x4_max(wasm_f32x4_mul(vx0123, vprescale), vsat_cutoff);
51 const v128_t vz4567 = wasm_f32x4_max(wasm_f32x4_mul(vx4567, vprescale), vsat_cutoff);
52
53 v128_t vn0123 = wasm_f32x4_add(wasm_f32x4_mul(vz0123, vlog2e), vmagic_bias);
54 v128_t vn4567 = wasm_f32x4_add(wasm_f32x4_mul(vz4567, vlog2e), vmagic_bias);
55
56 v128_t vs0123 = wasm_i32x4_shl(vn0123, 23);
57 v128_t vs4567 = wasm_i32x4_shl(vn4567, 23);
58
59 vn0123 = wasm_f32x4_sub(vn0123, vmagic_bias);
60 vn4567 = wasm_f32x4_sub(vn4567, vmagic_bias);
61
62 v128_t vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_hi), vz0123);
63 v128_t vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_hi), vz4567);
64
65 vt0123 = wasm_f32x4_add(wasm_f32x4_mul(vn0123, vminus_ln2_lo), vt0123);
66 vt4567 = wasm_f32x4_add(wasm_f32x4_mul(vn4567, vminus_ln2_lo), vt4567);
67
68 v128_t vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt0123), vc5);
69 v128_t vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt4567), vc5);
70
71 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc4);
72 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc4);
73
74 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc3);
75 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc3);
76
77 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vc2);
78 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vc2);
79
80 vp0123 = wasm_f32x4_mul(vp0123, vt0123);
81 vp4567 = wasm_f32x4_mul(vp4567, vt4567);
82
83 vt0123 = wasm_f32x4_mul(vt0123, vs0123);
84 vs0123 = wasm_f32x4_sub(vs0123, vone);
85 vt4567 = wasm_f32x4_mul(vt4567, vs4567);
86 vs4567 = wasm_f32x4_sub(vs4567, vone);
87
88 vp0123 = wasm_f32x4_add(wasm_f32x4_mul(vp0123, vt0123), vt0123);
89 vp4567 = wasm_f32x4_add(wasm_f32x4_mul(vp4567, vt4567), vt4567);
90
91 const v128_t ve0123 = wasm_f32x4_mul(wasm_f32x4_add(vp0123, vs0123), valpha);
92 const v128_t ve4567 = wasm_f32x4_mul(wasm_f32x4_add(vp4567, vs4567), valpha);
93
94 const v128_t vsignm0123 = wasm_i32x4_shr(vx0123, 31);
95 vx0123 = wasm_f32x4_mul(vx0123, vbeta);
96 const v128_t vsignm4567 = wasm_i32x4_shr(vx4567, 31);
97 vx4567 = wasm_f32x4_mul(vx4567, vbeta);
98
99 const v128_t vy0123 = wasm_v128_bitselect(ve0123, vx0123, vsignm0123);
100 const v128_t vy4567 = wasm_v128_bitselect(ve4567, vx4567, vsignm4567);
101
102 wasm_v128_store(y, vy0123);
103 wasm_v128_store(y + 4, vy4567);
104 y += 8;
105 }
106 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
107 v128_t vx = wasm_v128_load(x);
108 x += 4;
109
110 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
111
112 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
113 v128_t vs = wasm_i32x4_shl(vn, 23);
114 vn = wasm_f32x4_sub(vn, vmagic_bias);
115
116 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
117 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
118
119 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
120 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
121 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
122 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
123 vp = wasm_f32x4_mul(vp, vt);
124
125 vt = wasm_f32x4_mul(vt, vs);
126 vs = wasm_f32x4_sub(vs, vone);
127 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
128 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
129
130 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
131 vx = wasm_f32x4_mul(vx, vbeta);
132 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
133
134 wasm_v128_store(y, vy);
135 y += 4;
136 }
137 if XNN_UNLIKELY(n != 0) {
138 v128_t vx = wasm_v128_load(x);
139
140 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
141
142 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
143 v128_t vs = wasm_i32x4_shl(vn, 23);
144 vn = wasm_f32x4_sub(vn, vmagic_bias);
145
146 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
147 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
148
149 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
150 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
151 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
152 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
153 vp = wasm_f32x4_mul(vp, vt);
154
155 vt = wasm_f32x4_mul(vt, vs);
156 vs = wasm_f32x4_sub(vs, vone);
157 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
158 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
159
160 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
161 vx = wasm_f32x4_mul(vx, vbeta);
162 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
163
164 if (n & (2 * sizeof(float))) {
165 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
166 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
167 y += 2;
168 }
169 if (n & (1 * sizeof(float))) {
170 *y = wasm_f32x4_extract_lane(vy, 0);
171 }
172 }
173 }
174