1 // Auto-generated file. Do not edit!
2 // Template: src/f32-hswish/wasmsimd.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/hswish.h>
16
17
xnn_f32_hswish_ukernel__wasmsimd_x16(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_hswish_ukernel__wasmsimd_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const v128_t vsixth = wasm_v32x4_load_splat(¶ms->scalar.sixth);
28 const v128_t vthree = wasm_v32x4_load_splat(¶ms->scalar.three);
29 const v128_t vsix = wasm_v32x4_load_splat(¶ms->scalar.six);
30 const v128_t vzero = wasm_f32x4_splat(0.0f);
31
32 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 v128_t vx0123 = wasm_v128_load(x);
34 v128_t vx4567 = wasm_v128_load(x + 4);
35 v128_t vx89AB = wasm_v128_load(x + 8);
36 v128_t vxCDEF = wasm_v128_load(x + 12);
37 x += 16;
38
39 v128_t vacc0123 = wasm_f32x4_add(vx0123, vthree);
40 vx0123 = wasm_f32x4_mul(vx0123, vsixth);
41 v128_t vacc4567 = wasm_f32x4_add(vx4567, vthree);
42 vx4567 = wasm_f32x4_mul(vx4567, vsixth);
43 v128_t vacc89AB = wasm_f32x4_add(vx89AB, vthree);
44 vx89AB = wasm_f32x4_mul(vx89AB, vsixth);
45 v128_t vaccCDEF = wasm_f32x4_add(vxCDEF, vthree);
46 vxCDEF = wasm_f32x4_mul(vxCDEF, vsixth);
47
48 vacc0123 = wasm_i32x4_max(vacc0123, vzero);
49 vacc4567 = wasm_i32x4_max(vacc4567, vzero);
50 vacc89AB = wasm_i32x4_max(vacc89AB, vzero);
51 vaccCDEF = wasm_i32x4_max(vaccCDEF, vzero);
52
53 vacc0123 = wasm_i32x4_min(vacc0123, vsix);
54 vacc4567 = wasm_i32x4_min(vacc4567, vsix);
55 vacc89AB = wasm_i32x4_min(vacc89AB, vsix);
56 vaccCDEF = wasm_i32x4_min(vaccCDEF, vsix);
57
58 vacc0123 = wasm_f32x4_mul(vacc0123, vx0123);
59 vacc4567 = wasm_f32x4_mul(vacc4567, vx4567);
60 vacc89AB = wasm_f32x4_mul(vacc89AB, vx89AB);
61 vaccCDEF = wasm_f32x4_mul(vaccCDEF, vxCDEF);
62
63 wasm_v128_store(y, vacc0123);
64 wasm_v128_store(y + 4, vacc4567);
65 wasm_v128_store(y + 8, vacc89AB);
66 wasm_v128_store(y + 12, vaccCDEF);
67 y += 16;
68 }
69 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
70 v128_t vx = wasm_v128_load(x);
71 x += 4;
72
73 v128_t vacc = wasm_f32x4_add(vx, vthree);
74 vx = wasm_f32x4_mul(vx, vsixth);
75 vacc = wasm_i32x4_max(vacc, vzero);
76 vacc = wasm_i32x4_min(vacc, vsix);
77 vacc = wasm_f32x4_mul(vacc, vx);
78
79 wasm_v128_store(y, vacc);
80 y += 4;
81 }
82 if XNN_UNLIKELY(n != 0) {
83 v128_t vx = wasm_v128_load(x);
84
85 v128_t vacc = wasm_f32x4_add(vx, vthree);
86 vx = wasm_f32x4_mul(vx, vsixth);
87 vacc = wasm_i32x4_max(vacc, vzero);
88 vacc = wasm_i32x4_min(vacc, vsix);
89 vacc = wasm_f32x4_mul(vacc, vx);
90
91 if (n & (2 * sizeof(float))) {
92 *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
93 vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
94 y += 2;
95 }
96 if (n & (1 * sizeof(float))) {
97 *y = wasm_f32x4_extract_lane(vacc, 0);
98 }
99 }
100 }
101