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_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_arm_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_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
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 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
56
57 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
58 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
59 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
60 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
61 vp = wasm_f32x4_mul(vp, vt);
62
63 vt = wasm_f32x4_mul(vt, vs);
64 vs = wasm_f32x4_sub(vs, vone);
65 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
66 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
67
68 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
69 vx = wasm_f32x4_mul(vx, vbeta);
70 const v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
71
72 wasm_v128_store(y, vy);
73 y += 4;
74 }
75 if XNN_UNLIKELY(n != 0) {
76 v128_t vx = wasm_v128_load(x);
77
78 const v128_t vz = wasm_f32x4_max(wasm_f32x4_mul(vx, vprescale), vsat_cutoff);
79
80 v128_t vn = wasm_f32x4_add(wasm_f32x4_mul(vz, vlog2e), vmagic_bias);
81 v128_t vs = wasm_i32x4_shl(vn, 23);
82 vn = wasm_f32x4_sub(vn, vmagic_bias);
83
84 v128_t vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_hi), vz);
85 vt = wasm_f32x4_add(wasm_f32x4_mul(vn, vminus_ln2_lo), vt);
86
87 v128_t vp = wasm_f32x4_add(wasm_f32x4_mul(vc6, vt), vc5);
88 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc4);
89 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc3);
90 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vc2);
91 vp = wasm_f32x4_mul(vp, vt);
92
93 vt = wasm_f32x4_mul(vt, vs);
94 vs = wasm_f32x4_sub(vs, vone);
95 vp = wasm_f32x4_add(wasm_f32x4_mul(vp, vt), vt);
96 const v128_t ve = wasm_f32x4_mul(wasm_f32x4_add(vp, vs), valpha);
97
98 const v128_t vsignm = wasm_i32x4_shr(vx, 31);
99 vx = wasm_f32x4_mul(vx, vbeta);
100 v128_t vy = wasm_v128_bitselect(ve, vx, vsignm);
101
102 if (n & (2 * sizeof(float))) {
103 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
104 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
105 y += 2;
106 }
107 if (n & (1 * sizeof(float))) {
108 *y = wasm_f32x4_extract_lane(vy, 0);
109 }
110 }
111 }
112