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 <smmintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16 
17 
xnn_f32_vlrelu_ukernel__sse41_x4(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__sse41_x4(
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   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
29     const __m128 vx0123 = _mm_loadu_ps(x);
30     x += 4;
31 
32     __m128 vacc0123 = _mm_mul_ps(vx0123, vslope);
33 
34     vacc0123 = _mm_blendv_ps(vx0123, vacc0123, vx0123);
35 
36     _mm_storeu_ps(y, vacc0123);
37     y += 4;
38   }
39   if XNN_UNLIKELY(n != 0) {
40     const __m128 vx = _mm_loadu_ps(x);
41 
42     __m128 vacc = _mm_mul_ps(vx, vslope);
43     vacc = _mm_blendv_ps(vx, vacc, vx);
44 
45     if (n & (2 * sizeof(float))) {
46       _mm_storel_pi((__m64*) y, vacc);
47       vacc = _mm_movehl_ps(vacc, vacc);
48       y += 2;
49     }
50     if (n & (1 * sizeof(float))) {
51       _mm_store_ss(y, vacc);
52     }
53   }
54 }
55