• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vunary/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/math.h>
16 #include <xnnpack/vunary.h>
17 
18 
xnn_f32_vneg_ukernel__sse_x4(size_t n,const float * x,float * y,const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vneg_ukernel__sse_x4(
20     size_t n,
21     const float* x,
22     float* y,
23     const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27   assert(x != NULL);
28   assert(y != NULL);
29 
30   const __m128 vsign_mask = _mm_load_ps(params->sse.sign_mask);
31   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
32     const __m128 vx0123 = _mm_loadu_ps(x);
33     x += 4;
34 
35     const __m128 vy0123 = _mm_xor_ps(vx0123, vsign_mask);
36 
37     _mm_storeu_ps(y, vy0123);
38     y += 4;
39   }
40   if XNN_UNLIKELY(n != 0) {
41     const __m128 vx = _mm_loadu_ps(x);
42     __m128 vy = _mm_xor_ps(vx, vsign_mask);
43     if (n & (2 * sizeof(float))) {
44       _mm_storel_pi((__m64*) y, vy);
45       vy = _mm_movehl_ps(vy, vy);
46       y += 2;
47     }
48     if (n & (1 * sizeof(float))) {
49       _mm_store_ss(y, vy);
50     }
51   }
52 }
53