• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vopc-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/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17 
18 
xnn_f32_vdivc_minmax_ukernel__sse_x4(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vdivc_minmax_ukernel__sse_x4(
20     size_t n,
21     const float* a,
22     const float* b,
23     float* y,
24     const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28   assert(a != NULL);
29   assert(b != NULL);
30   assert(y != NULL);
31 
32   const __m128 vy_min = _mm_load_ps(params->sse.min);
33   const __m128 vy_max = _mm_load_ps(params->sse.max);
34 
35   const __m128 vb = _mm_load1_ps(b);
36   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
37     const __m128 va0123 = _mm_loadu_ps(a);
38     a += 4;
39 
40     __m128 vy0123 = _mm_div_ps(va0123, vb);
41 
42 
43     vy0123 = _mm_max_ps(vy0123, vy_min);
44 
45     vy0123 = _mm_min_ps(vy0123, vy_max);
46 
47     _mm_storeu_ps(y, vy0123);
48     y += 4;
49   }
50   if XNN_UNLIKELY(n != 0) {
51     const __m128 va0123 = _mm_loadu_ps(a);
52 
53     __m128 vy0123 = _mm_div_ps(va0123, vb);
54     vy0123 = _mm_max_ps(vy0123, vy_min);
55     vy0123 = _mm_min_ps(vy0123, vy_max);
56     if (n & (2 * sizeof(float))) {
57       _mm_storel_pi((__m64*) y, vy0123);
58       vy0123 = _mm_movehl_ps(vy0123, vy0123);
59       y += 2;
60     }
61     if (n & (1 * sizeof(float))) {
62       _mm_store_ss(y, vy0123);
63     }
64   }
65 }
66