• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vop-sse.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 <xmmintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17 
18 
xnn_f32_vadd_minmax_ukernel__sse_x8(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vadd_minmax_ukernel__sse_x8(
20     size_t n,
21     const float* a,
22     const float* b,
23     float* y,
24     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28   assert(a != NULL);
29   assert(b != NULL);
30   assert(y != NULL);
31 
32   const __m128 vy_min = _mm_load_ps(params->sse.min);
33   const __m128 vy_max = _mm_load_ps(params->sse.max);
34 
35   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
36     const __m128 va0123 = _mm_loadu_ps(a);
37     const __m128 va4567 = _mm_loadu_ps(a + 4);
38     a += 8;
39 
40     const __m128 vb0123 = _mm_loadu_ps(b);
41     const __m128 vb4567 = _mm_loadu_ps(b + 4);
42     b += 8;
43 
44     __m128 vy0123 = _mm_add_ps(va0123, vb0123);
45     __m128 vy4567 = _mm_add_ps(va4567, vb4567);
46 
47 
48     vy0123 = _mm_max_ps(vy0123, vy_min);
49     vy4567 = _mm_max_ps(vy4567, vy_min);
50 
51     vy0123 = _mm_min_ps(vy0123, vy_max);
52     vy4567 = _mm_min_ps(vy4567, vy_max);
53 
54     _mm_storeu_ps(y, vy0123);
55     _mm_storeu_ps(y + 4, vy4567);
56     y += 8;
57   }
58   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
59     const __m128 va0123 = _mm_loadu_ps(a);
60     a += 4;
61 
62     const __m128 vb0123 = _mm_loadu_ps(b);
63     b += 4;
64 
65     __m128 vy0123 = _mm_add_ps(va0123, vb0123);
66     vy0123 = _mm_max_ps(vy0123, vy_min);
67     vy0123 = _mm_min_ps(vy0123, vy_max);
68     _mm_storeu_ps(y, vy0123);
69     y += 4;
70   }
71   if XNN_UNLIKELY(n != 0) {
72     const __m128 va0123 = _mm_loadu_ps(a);
73     const __m128 vb0123 = _mm_loadu_ps(b);
74 
75     __m128 vy0123 = _mm_add_ps(va0123, vb0123);
76     vy0123 = _mm_max_ps(vy0123, vy_min);
77     vy0123 = _mm_min_ps(vy0123, vy_max);
78     if (n & (2 * sizeof(float))) {
79       _mm_storel_pi((__m64*) y, vy0123);
80       vy0123 = _mm_movehl_ps(vy0123, vy0123);
81       y += 2;
82     }
83     if (n & (1 * sizeof(float))) {
84       _mm_store_ss(y, vy0123);
85     }
86   }
87 }
88