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_vmax_ukernel__sse_x8(size_t n,const float * a,const float * b,float * y,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vmax_ukernel__sse_x8(
20     size_t n,
21     const float* a,
22     const float* b,
23     float* y,
24     const union xnn_f32_default_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 
33   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
34     const __m128 va0123 = _mm_loadu_ps(a);
35     const __m128 va4567 = _mm_loadu_ps(a + 4);
36     a += 8;
37 
38     const __m128 vb0123 = _mm_loadu_ps(b);
39     const __m128 vb4567 = _mm_loadu_ps(b + 4);
40     b += 8;
41 
42     __m128 vy0123 = _mm_max_ps(va0123, vb0123);
43     __m128 vy4567 = _mm_max_ps(va4567, vb4567);
44 
45 
46 
47     _mm_storeu_ps(y, vy0123);
48     _mm_storeu_ps(y + 4, vy4567);
49     y += 8;
50   }
51   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
52     const __m128 va0123 = _mm_loadu_ps(a);
53     a += 4;
54 
55     const __m128 vb0123 = _mm_loadu_ps(b);
56     b += 4;
57 
58     __m128 vy0123 = _mm_max_ps(va0123, vb0123);
59     _mm_storeu_ps(y, vy0123);
60     y += 4;
61   }
62   if XNN_UNLIKELY(n != 0) {
63     const __m128 va0123 = _mm_loadu_ps(a);
64     const __m128 vb0123 = _mm_loadu_ps(b);
65 
66     __m128 vy0123 = _mm_max_ps(va0123, vb0123);
67     if (n & (2 * sizeof(float))) {
68       _mm_storel_pi((__m64*) y, vy0123);
69       vy0123 = _mm_movehl_ps(vy0123, vy0123);
70       y += 2;
71     }
72     if (n & (1 * sizeof(float))) {
73       _mm_store_ss(y, vy0123);
74     }
75   }
76 }
77