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_vadd_ukernel__sse_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_vadd_ukernel__sse_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 __m128 vy_min = _mm_load_ps(params->sse.min);
29 const __m128 vy_max = _mm_load_ps(params->sse.max);
30
31 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
32 const __m128 va0123 = _mm_loadu_ps(a);
33 a += 4;
34
35 const __m128 vb0123 = _mm_loadu_ps(b);
36 b += 4;
37
38 __m128 vy0123 = _mm_add_ps(va0123, vb0123);
39
40 vy0123 = _mm_max_ps(vy0123, vy_min);
41
42 vy0123 = _mm_min_ps(vy0123, vy_max);
43
44 _mm_storeu_ps(y, vy0123);
45 y += 4;
46 }
47 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
48 const __m128 va0123 = _mm_loadu_ps(a);
49 a += 4;
50
51 const __m128 vb0123 = _mm_loadu_ps(b);
52 b += 4;
53
54 __m128 vy0123 = _mm_add_ps(va0123, vb0123);
55 vy0123 = _mm_max_ps(vy0123, vy_min);
56 vy0123 = _mm_min_ps(vy0123, vy_max);
57 _mm_storeu_ps(y, vy0123);
58 y += 4;
59 }
60 if XNN_UNLIKELY(n != 0) {
61 const __m128 va0123 = _mm_loadu_ps(a);
62 const __m128 vb0123 = _mm_loadu_ps(b);
63
64 __m128 vy0123 = _mm_add_ps(va0123, vb0123);
65 vy0123 = _mm_max_ps(vy0123, vy_min);
66 vy0123 = _mm_min_ps(vy0123, vy_max);
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