xref: /aosp_15_r20/external/XNNPACK/src/f32-vhswish/gen/vhswish-wasmsimd-x4.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vhswish/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/vunary.h>
16 
17 
xnn_f32_vhswish_ukernel__wasmsimd_x4(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vhswish_ukernel__wasmsimd_x4(
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_OOB_READS
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   const v128_t vsixth = wasm_v128_load64_splat(params->wasmsimd.sixth);
28   const v128_t vthree = wasm_v128_load64_splat(params->wasmsimd.three);
29   const v128_t vsix = wasm_v128_load64_splat(params->wasmsimd.six);
30   const v128_t vzero = wasm_i32x4_const_splat(0);
31 
32   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33     v128_t vx = wasm_v128_load(x);
34     x += 4;
35 
36     v128_t vacc = wasm_f32x4_add(vx, vthree);
37     vx = wasm_f32x4_mul(vx, vsixth);
38     vacc = wasm_i32x4_max(vacc, vzero);
39     vacc = wasm_i32x4_min(vacc, vsix);
40     vacc = wasm_f32x4_mul(vacc, vx);
41 
42     wasm_v128_store(y, vacc);
43     y += 4;
44   }
45   if XNN_UNLIKELY(n != 0) {
46     v128_t vx = wasm_v128_load(x);
47 
48     v128_t vacc = wasm_f32x4_add(vx, vthree);
49     vx = wasm_f32x4_mul(vx, vsixth);
50     vacc = wasm_i32x4_max(vacc, vzero);
51     vacc = wasm_i32x4_min(vacc, vsix);
52     vacc = wasm_f32x4_mul(vacc, vx);
53 
54     if (n & (2 * sizeof(float))) {
55       *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
56       vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
57       y += 2;
58     }
59     if (n & (1 * sizeof(float))) {
60       *y = wasm_f32x4_extract_lane(vacc, 0);
61     }
62   }
63 }
64