• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vopc-psimd.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 <psimd.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vbinary.h>
16 
17 
xnn_f32_vminc_ukernel__psimd_x4(size_t n,const float * a,const float * b,float * y,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_vminc_ukernel__psimd_x4(
19     size_t n,
20     const float* a,
21     const float* b,
22     float* y,
23     const union xnn_f32_output_params params[restrict static 1])
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27 
28   const psimd_f32 vy_min = psimd_load_splat_f32(&params->scalar.min);
29   const psimd_f32 vy_max = psimd_load_splat_f32(&params->scalar.max);
30 
31   const psimd_f32 vb = psimd_load_splat_f32(b);
32   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33     const psimd_f32 va0123 = psimd_load_f32(a);
34     a += 4;
35 
36     psimd_f32 vy0123 = psimd_min_f32(va0123, vb);
37 
38     vy0123 = psimd_max_f32(vy0123, vy_min);
39 
40     vy0123 = psimd_min_f32(vy0123, vy_max);
41 
42     psimd_store_f32(y, vy0123);
43     y += 4;
44   }
45   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
46     const psimd_f32 va0123 = psimd_load_f32(a);
47     a += 4;
48 
49     psimd_f32 vy0123 = psimd_min_f32(va0123, vb);
50     vy0123 = psimd_max_f32(vy0123, vy_min);
51     vy0123 = psimd_min_f32(vy0123, vy_max);
52     psimd_store_f32(y, vy0123);
53     y += 4;
54   }
55   if XNN_UNLIKELY(n != 0) {
56     const psimd_f32 va0123 = psimd_load_f32(a);
57 
58     psimd_f32 vy0123 = psimd_min_f32(va0123, vb);
59     vy0123 = psimd_max_f32(vy0123, vy_min);
60     vy0123 = psimd_min_f32(vy0123, vy_max);
61     if (n & (2 * sizeof(float))) {
62       psimd_store2_f32(y, vy0123);
63       vy0123 = psimd_concat_hi_f32(vy0123, vy0123);
64       y += 2;
65     }
66     if (n & (1 * sizeof(float))) {
67       psimd_store1_f32(y, vy0123);
68     }
69   }
70 }
71