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_arm_x8(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_arm_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
33     v128_t vacc0123 = wasm_v128_load(x);
34     v128_t vacc4567 = wasm_v128_load(x + 4);
35     x += 8;
36 
37     vacc0123 = wasm_f32x4_max(vacc0123, vy_min);
38     vacc4567 = wasm_f32x4_max(vacc4567, vy_min);
39 
40     vacc0123 = wasm_f32x4_min(vacc0123, vy_max);
41     vacc4567 = wasm_f32x4_min(vacc4567, vy_max);
42 
43     wasm_v128_store(y, vacc0123);
44     wasm_v128_store(y + 4, vacc4567);
45     y += 8;
46   }
47   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
48     v128_t vacc = wasm_v128_load(x);
49     x += 4;
50 
51     vacc = wasm_f32x4_max(vacc, vy_min);
52     vacc = wasm_f32x4_min(vacc, vy_max);
53 
54     wasm_v128_store(y, vacc);
55     y += 4;
56   }
57   if XNN_UNLIKELY(n != 0) {
58     v128_t vacc = wasm_v128_load(x);
59 
60     vacc = wasm_f32x4_max(vacc, vy_min);
61     vacc = wasm_f32x4_min(vacc, vy_max);
62 
63     if (n & (2 * sizeof(float))) {
64       *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
65       vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
66       y += 2;
67     }
68     if (n & (1 * sizeof(float))) {
69       *y = wasm_f32x4_extract_lane(vacc, 0);
70     }
71   }
72 }
73