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_x4(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_x4(
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 >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
45 v128_t vx = wasm_v128_load(x);
46 x += 4;
47
48 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
49
50 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
51 v128_t vs = wasm_i32x4_shl(vn, 23);
52 vn = wasm_f32x4_sub(vn, vmagic_bias);
53
54 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
55 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
56 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
57 vs = wasm_v128_andnot(vs, vsatm);
58 vt = wasm_v128_andnot(vt, vsatm);
59
60 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
61 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
62 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
63 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
64 vp = wasm_f32x4_mul(vp, vt);
65
66 vt = wasm_f32x4_mul(vt, vs);
67 vs = wasm_f32x4_sub(vs, vone);
68 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
69 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
70
71 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
72 vx = wasm_f32x4_mul(vx, vbeta);
73 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
74
75 wasm_v128_store(y, vy);
76 y += 4;
77 }
78 if XNN_UNLIKELY(n != 0) {
79 v128_t vx = wasm_v128_load(x);
80
81 const v128_t vz = wasm_f32x4_mul(vx, vprescale);
82
83 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
84 v128_t vs = wasm_i32x4_shl(vn, 23);
85 vn = wasm_f32x4_sub(vn, vmagic_bias);
86
87 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
88 const v128_t vsatm = wasm_f32x4_le(vz, vsat_cutoff);
89 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
90 vs = wasm_v128_andnot(vs, vsatm);
91 vt = wasm_v128_andnot(vt, vsatm);
92
93 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
94 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
95 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
96 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
97 vp = wasm_f32x4_mul(vp, vt);
98
99 vt = wasm_f32x4_mul(vt, vs);
100 vs = wasm_f32x4_sub(vs, vone);
101 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
102 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
103
104 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
105 vx = wasm_f32x4_mul(vx, vbeta);
106 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
107
108 if (n & (2 * sizeof(float))) {
109 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
110 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
111 y += 2;
112 }
113 if (n & (1 * sizeof(float))) {
114 *y = wasm_f32x4_extract_lane(vy, 0);
115 }
116 }
117 }
118