1 // Auto-generated file. Do not edit!
2 // Template: src/f32-hswish/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 <emmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vbinary.h>
16
17
xnn_f32_hswish_ukernel__sse_x4(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict static1])18 void xnn_f32_hswish_ukernel__sse_x4(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_hswish_params params[restrict static 1])
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26
27 const __m128 vsixth = _mm_load_ps(params->sse.sixth);
28 const __m128 vhalf = _mm_load_ps(params->sse.half);
29 const __m128 vone = _mm_load_ps(params->sse.one);
30 const __m128 vzero = _mm_setzero_ps();
31
32 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
33 const __m128 vx0123 = _mm_loadu_ps(x);
34 x += 4;
35
36 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
37
38 vacc0123 = _mm_add_ps(vacc0123, vhalf);
39
40 vacc0123 = _mm_max_ps(vacc0123, vzero);
41
42 vacc0123 = _mm_min_ps(vacc0123, vone);
43
44 vacc0123 = _mm_mul_ps(vacc0123, vx0123);
45
46 _mm_storeu_ps(y, vacc0123);
47 y += 4;
48 }
49 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
50 const __m128 vx0123 = _mm_loadu_ps(x);
51 x += 4;
52 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
53 vacc0123 = _mm_add_ps(vacc0123, vhalf);
54 vacc0123 = _mm_max_ps(vacc0123, vzero);
55 vacc0123 = _mm_min_ps(vacc0123, vone);
56 vacc0123 = _mm_mul_ps(vacc0123, vx0123);
57 _mm_storeu_ps(y, vacc0123);
58 y += 4;
59 }
60 if XNN_UNLIKELY(n != 0) {
61 const __m128 vx0123 = _mm_loadu_ps(x);
62 __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
63 vacc0123 = _mm_add_ps(vacc0123, vhalf);
64 vacc0123 = _mm_max_ps(vacc0123, vzero);
65 vacc0123 = _mm_min_ps(vacc0123, vone);
66 vacc0123 = _mm_mul_ps(vacc0123, vx0123);
67
68 if (n & (2 * sizeof(float))) {
69 _mm_storel_pi((__m64*) y, vacc0123);
70 vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
71 y += 2;
72 }
73 if (n & (1 * sizeof(float))) {
74 _mm_store_ss(y, vacc0123);
75 }
76 }
77 }
78