1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/wasmsimd-rr2-p5-div.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/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x4(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n % sizeof(float) == 0);
25
26 const v128_t vmagic_bias = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.magic_bias);
27 const v128_t vminus_log2e = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.minus_log2e);
28 const v128_t vln2_hi = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.ln2_hi);
29 const v128_t vln2_lo = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.ln2_lo);
30 const v128_t vc5 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c5);
31 const v128_t vc4 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c4);
32 const v128_t vc3 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c3);
33 const v128_t vc2 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c2);
34 const v128_t vc1 = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.c1);
35 const v128_t vone = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.one);
36 const v128_t vdenorm_cutoff = wasm_v128_load64_splat(params->wasmsimd_rr2_p5.denorm_cutoff);
37
38 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
39 const v128_t vx = wasm_v128_load(x);
40 x += 4;
41
42 const v128_t vz = wasm_f32x4_abs(vx);
43
44 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
45 const v128_t vs = wasm_i32x4_shl(vn, 23);
46 vn = wasm_f32x4_sub(vn, vmagic_bias);
47
48 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
49 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
50
51 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
52 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
53 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
54 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
55
56 vt = wasm_f32x4_mul(vt, vs);
57 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
58 const v128_t vd = wasm_f32x4_add(ve, vone);
59
60 v128_t vf = wasm_f32x4_div(ve, vd);
61 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
62 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
63
64 wasm_v128_store(y, vf);
65 y += 4;
66 }
67 if XNN_UNLIKELY(n != 0) {
68 const v128_t vx = wasm_v128_load(x);
69
70 const v128_t vz = wasm_f32x4_abs(vx);
71
72 v128_t vn = wasm_f32x4_add(vmagic_bias, wasm_f32x4_mul(vz, vminus_log2e));
73 const v128_t vs = wasm_i32x4_shl(vn, 23);
74 vn = wasm_f32x4_sub(vn, vmagic_bias);
75
76 v128_t vt = wasm_f32x4_add(vz, wasm_f32x4_mul(vn, vln2_hi));
77 vt = wasm_f32x4_add(vt, wasm_f32x4_mul(vn, vln2_lo));
78
79 v128_t vp = wasm_f32x4_add(vc4, wasm_f32x4_mul(vt, vc5));
80 vp = wasm_f32x4_add(vc3, wasm_f32x4_mul(vt, vp));
81 vp = wasm_f32x4_add(vc2, wasm_f32x4_mul(vt, vp));
82 vp = wasm_f32x4_add(vc1, wasm_f32x4_mul(vt, vp));
83
84 vt = wasm_f32x4_mul(vt, vs);
85 const v128_t ve = wasm_f32x4_add(vs, wasm_f32x4_mul(vt, vp));
86 const v128_t vd = wasm_f32x4_add(ve, vone);
87
88 v128_t vf = wasm_f32x4_div(ve, vd);
89 vf = wasm_v128_andnot(vf, wasm_f32x4_gt(vz, vdenorm_cutoff));
90 vf = wasm_v128_bitselect(vf, wasm_f32x4_sub(vone, vf), wasm_i32x4_shr(vx, 31));
91
92 if (n & (2 * sizeof(float))) {
93 *((double*) y) = wasm_f64x2_extract_lane(vf, 0);
94 vf = wasm_v32x4_shuffle(vf, vf, 2, 3, 2, 3);
95 y += 2;
96 }
97 if (n & (1 * sizeof(float))) {
98 *y = wasm_f32x4_extract_lane(vf, 0);
99 }
100 }
101 }
102