• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-clamp/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/clamp.h>
15 #include <xnnpack/common.h>
16 
17 
xnn_f32_clamp_ukernel__wasmsimd_x86_x4(size_t n,const float * x,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_clamp_ukernel__wasmsimd_x86_x4(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26   assert(x != NULL);
27   assert(y != NULL);
28 
29   const v128_t vy_min = wasm_v32x4_load_splat(&params->scalar.min);
30   const v128_t vy_max = wasm_v32x4_load_splat(&params->scalar.max);
31 
32   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33     v128_t vacc = wasm_v128_load(x);
34     x += 4;
35 
36     const v128_t vmaskmin = wasm_f32x4_lt(vacc, vy_min);
37     const v128_t vmaskmax = wasm_f32x4_le(vy_max, vacc);
38     vacc = wasm_v128_bitselect(vy_min, vacc, vmaskmin);
39     vacc = wasm_v128_bitselect(vy_max, vacc, vmaskmax);
40 
41     wasm_v128_store(y, vacc);
42     y += 4;
43   }
44   if XNN_UNLIKELY(n != 0) {
45     v128_t vacc = wasm_v128_load(x);
46 
47     const v128_t vmaskmin = wasm_f32x4_lt(vacc, vy_min);
48     const v128_t vmaskmax = wasm_f32x4_le(vy_max, vacc);
49     vacc = wasm_v128_bitselect(vy_min, vacc, vmaskmin);
50     vacc = wasm_v128_bitselect(vy_max, vacc, vmaskmax);
51 
52     if (n & (2 * sizeof(float))) {
53       *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
54       vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
55       y += 2;
56     }
57     if (n & (1 * sizeof(float))) {
58       *y = wasm_f32x4_extract_lane(vacc, 0);
59     }
60   }
61 }
62