1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vlrelu/sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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_vlrelu_ukernel__sse_x8(size_t n,const float * x,float * y,const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vlrelu_ukernel__sse_x8(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26 
27   const __m128 vslope = _mm_load_ps(params->sse.slope);
28   const __m128 vzero = _mm_setzero_ps();
29   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
30     __m128 vx0123 = _mm_loadu_ps(x);
31     __m128 vx4567 = _mm_loadu_ps(x + 4);
32     x += 8;
33 
34     __m128 vacc0123 = _mm_max_ps(_mm_setzero_ps(), vx0123);
35     vx0123 = _mm_min_ps(vx0123, vzero);
36     __m128 vacc4567 = _mm_max_ps(_mm_setzero_ps(), vx4567);
37     vx4567 = _mm_min_ps(vx4567, vzero);
38 
39     vacc0123 = _mm_add_ps(vacc0123, _mm_mul_ps(vx0123, vslope));
40     vacc4567 = _mm_add_ps(vacc4567, _mm_mul_ps(vx4567, vslope));
41 
42     _mm_storeu_ps(y, vacc0123);
43     _mm_storeu_ps(y + 4, vacc4567);
44     y += 8;
45   }
46   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
47     __m128 vx = _mm_loadu_ps(x);
48     x += 4;
49 
50     __m128 vacc = _mm_max_ps(_mm_setzero_ps(), vx);
51     vx = _mm_min_ps(vx, vzero);
52     vacc = _mm_add_ps(vacc, _mm_mul_ps(vx, vslope));
53 
54     _mm_storeu_ps(y, vacc);
55     y += 4;
56   }
57   if XNN_UNLIKELY(n != 0) {
58     __m128 vx = _mm_loadu_ps(x);
59 
60     __m128 vacc = _mm_max_ps(_mm_setzero_ps(), vx);
61     vx = _mm_min_ps(vx, vzero);
62     vacc = _mm_add_ps(vacc, _mm_mul_ps(vx, vslope));
63 
64     if (n & (2 * sizeof(float))) {
65       _mm_storel_pi((__m64*) y, vacc);
66       vacc = _mm_movehl_ps(vacc, vacc);
67       y += 2;
68     }
69     if (n & (1 * sizeof(float))) {
70       _mm_store_ss(y, vacc);
71     }
72   }
73 }
74