blob: 09c79293b34efc2e19d0ac3076f673d269b465ab [file] [log] [blame]
Marat Dukhan40a672f2019-11-25 03:08:22 -08001// 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>
Marat Dukhan8aaf1862020-04-28 10:26:13 -070015#include <xnnpack/intrinsics-polyfill.h>
Marat Dukhan40a672f2019-11-25 03:08:22 -080016#include <xnnpack/vbinary.h>
17
18
Marat Dukhan91cd2b72020-04-09 23:57:31 -070019void xnn_f32_vsubc_minmax_ukernel__sse_x8(
Marat Dukhan40a672f2019-11-25 03:08:22 -080020 size_t n,
21 const float* a,
22 const float* b,
23 float* y,
Marat Dukhanb2217dd2020-05-28 17:30:28 -070024 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
Marat Dukhan40a672f2019-11-25 03:08:22 -080025{
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
Frank Barchard0822dde2020-07-04 12:47:24 -070028 assert(a != NULL);
29 assert(b != NULL);
30 assert(y != NULL);
Marat Dukhan40a672f2019-11-25 03:08:22 -080031
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
37 const __m128 va0123 = _mm_loadu_ps(a);
38 const __m128 va4567 = _mm_loadu_ps(a + 4);
39 a += 8;
40
41 __m128 vy0123 = _mm_sub_ps(va0123, vb);
42 __m128 vy4567 = _mm_sub_ps(va4567, vb);
43
Marat Dukhan13bafb02020-06-05 00:43:11 -070044
Marat Dukhan40a672f2019-11-25 03:08:22 -080045 vy0123 = _mm_max_ps(vy0123, vy_min);
46 vy4567 = _mm_max_ps(vy4567, vy_min);
47
48 vy0123 = _mm_min_ps(vy0123, vy_max);
49 vy4567 = _mm_min_ps(vy4567, vy_max);
50
51 _mm_storeu_ps(y, vy0123);
52 _mm_storeu_ps(y + 4, vy4567);
53 y += 8;
54 }
55 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
56 const __m128 va0123 = _mm_loadu_ps(a);
57 a += 4;
58
59 __m128 vy0123 = _mm_sub_ps(va0123, vb);
60 vy0123 = _mm_max_ps(vy0123, vy_min);
61 vy0123 = _mm_min_ps(vy0123, vy_max);
62 _mm_storeu_ps(y, vy0123);
63 y += 4;
64 }
65 if XNN_UNLIKELY(n != 0) {
Marat Dukhanb2217dd2020-05-28 17:30:28 -070066 const __m128 va0123 = _mm_loadu_ps(a);
Marat Dukhan40a672f2019-11-25 03:08:22 -080067
68 __m128 vy0123 = _mm_sub_ps(va0123, vb);
69 vy0123 = _mm_max_ps(vy0123, vy_min);
70 vy0123 = _mm_min_ps(vy0123, vy_max);
71 if (n & (2 * sizeof(float))) {
72 _mm_storel_pi((__m64*) y, vy0123);
73 vy0123 = _mm_movehl_ps(vy0123, vy0123);
74 y += 2;
75 }
76 if (n & (1 * sizeof(float))) {
77 _mm_store_ss(y, vy0123);
78 }
79 }
80}