xref: /aosp_15_r20/external/XNNPACK/src/f32-vhswish/gen/vhswish-sse-x8.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vhswish/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/vunary.h>
16 
17 
xnn_f32_vhswish_ukernel__sse_x8(size_t n,const float * x,float * y,const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vhswish_ukernel__sse_x8(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
33     const __m128 vx0123 = _mm_loadu_ps(x);
34     const __m128 vx4567 = _mm_loadu_ps(x + 4);
35     x += 8;
36 
37     __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
38     __m128 vacc4567 = _mm_mul_ps(vx4567, vsixth);
39 
40     vacc0123 = _mm_add_ps(vacc0123, vhalf);
41     vacc4567 = _mm_add_ps(vacc4567, vhalf);
42 
43     vacc0123 = _mm_max_ps(vacc0123, vzero);
44     vacc4567 = _mm_max_ps(vacc4567, vzero);
45 
46     vacc0123 = _mm_min_ps(vacc0123, vone);
47     vacc4567 = _mm_min_ps(vacc4567, vone);
48 
49     vacc0123 = _mm_mul_ps(vacc0123, vx0123);
50     vacc4567 = _mm_mul_ps(vacc4567, vx4567);
51 
52     _mm_storeu_ps(y, vacc0123);
53     _mm_storeu_ps(y + 4, vacc4567);
54     y += 8;
55   }
56   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
57     const __m128 vx0123 = _mm_loadu_ps(x);
58     x += 4;
59     __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
60     vacc0123 = _mm_add_ps(vacc0123, vhalf);
61     vacc0123 = _mm_max_ps(vacc0123, vzero);
62     vacc0123 = _mm_min_ps(vacc0123, vone);
63     vacc0123 = _mm_mul_ps(vacc0123, vx0123);
64     _mm_storeu_ps(y, vacc0123);
65     y += 4;
66   }
67   if XNN_UNLIKELY(n != 0) {
68     const __m128 vx0123 = _mm_loadu_ps(x);
69     __m128 vacc0123 = _mm_mul_ps(vx0123, vsixth);
70     vacc0123 = _mm_add_ps(vacc0123, vhalf);
71     vacc0123 = _mm_max_ps(vacc0123, vzero);
72     vacc0123 = _mm_min_ps(vacc0123, vone);
73     vacc0123 = _mm_mul_ps(vacc0123, vx0123);
74 
75     if (n & (2 * sizeof(float))) {
76       _mm_storel_pi((__m64*) y, vacc0123);
77       vacc0123 = _mm_movehl_ps(vacc0123, vacc0123);
78       y += 2;
79     }
80     if (n & (1 * sizeof(float))) {
81       _mm_store_ss(y, vacc0123);
82     }
83   }
84 }
85