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/vbinary.h>
16
17
xnn_f32_vsub_ukernel__sse_x8(size_t n,const float * a,const float * b,float * y,const union xnn_f32_output_params params[restrict static1])18 void xnn_f32_vsub_ukernel__sse_x8(
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 __m128 vy_min = _mm_load_ps(params->sse.min);
29 const __m128 vy_max = _mm_load_ps(params->sse.max);
30
31 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
32 const __m128 va0123 = _mm_loadu_ps(a);
33 const __m128 va4567 = _mm_loadu_ps(a + 4);
34 a += 8;
35
36 const __m128 vb0123 = _mm_loadu_ps(b);
37 const __m128 vb4567 = _mm_loadu_ps(b + 4);
38 b += 8;
39
40 __m128 vy0123 = _mm_sub_ps(va0123, vb0123);
41 __m128 vy4567 = _mm_sub_ps(va4567, vb4567);
42
43 vy0123 = _mm_max_ps(vy0123, vy_min);
44 vy4567 = _mm_max_ps(vy4567, vy_min);
45
46 vy0123 = _mm_min_ps(vy0123, vy_max);
47 vy4567 = _mm_min_ps(vy4567, vy_max);
48
49 _mm_storeu_ps(y, vy0123);
50 _mm_storeu_ps(y + 4, vy4567);
51 y += 8;
52 }
53 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
54 const __m128 va0123 = _mm_loadu_ps(a);
55 a += 4;
56
57 const __m128 vb0123 = _mm_loadu_ps(b);
58 b += 4;
59
60 __m128 vy0123 = _mm_sub_ps(va0123, vb0123);
61 vy0123 = _mm_max_ps(vy0123, vy_min);
62 vy0123 = _mm_min_ps(vy0123, vy_max);
63 _mm_storeu_ps(y, vy0123);
64 y += 4;
65 }
66 if XNN_UNLIKELY(n != 0) {
67 const __m128 va0123 = _mm_loadu_ps(a);
68 const __m128 vb0123 = _mm_loadu_ps(b);
69
70 __m128 vy0123 = _mm_sub_ps(va0123, vb0123);
71 vy0123 = _mm_max_ps(vy0123, vy_min);
72 vy0123 = _mm_min_ps(vy0123, vy_max);
73 if (n & (2 * sizeof(float))) {
74 _mm_storel_pi((__m64*) y, vy0123);
75 vy0123 = _mm_movehl_ps(vy0123, vy0123);
76 y += 2;
77 }
78 if (n & (1 * sizeof(float))) {
79 _mm_store_ss(y, vy0123);
80 }
81 }
82 }
83