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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
32 const __m128 vx0123 = _mm_loadu_ps(x);
33 const __m128 vx4567 = _mm_loadu_ps(x + 4);
34 x += 8;
35
36 const __m128 vy0123 = _mm_xor_ps(vx0123, vsign_mask);
37 const __m128 vy4567 = _mm_xor_ps(vx4567, vsign_mask);
38
39 _mm_storeu_ps(y, vy0123);
40 _mm_storeu_ps(y + 4, vy4567);
41 y += 8;
42 }
43 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
44 const __m128 vx = _mm_loadu_ps(x);
45 x += 4;
46 const __m128 vy = _mm_xor_ps(vx, vsign_mask);
47 _mm_storeu_ps(y, vy);
48 y += 4;
49 }
50 if XNN_UNLIKELY(n != 0) {
51 const __m128 vx = _mm_loadu_ps(x);
52 __m128 vy = _mm_xor_ps(vx, vsign_mask);
53 if (n & (2 * sizeof(float))) {
54 _mm_storel_pi((__m64*) y, vy);
55 vy = _mm_movehl_ps(vy, vy);
56 y += 2;
57 }
58 if (n & (1 * sizeof(float))) {
59 _mm_store_ss(y, vy);
60 }
61 }
62 }
63