1 // Auto-generated file. Do not edit!
2 // Template: src/f32-clamp/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/clamp.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_clamp_ukernel__sse_x8(size_t n,const float * x,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_clamp_ukernel__sse_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vy_min = _mm_load_ps(params->sse.min);
30 const __m128 vy_max = _mm_load_ps(params->sse.max);
31
32 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
33 __m128 vacc0123 = _mm_loadu_ps(x);
34 __m128 vacc4567 = _mm_loadu_ps(x + 4);
35 x += 8;
36
37 vacc0123 = _mm_max_ps(vacc0123, vy_min);
38 vacc4567 = _mm_max_ps(vacc4567, vy_min);
39
40 vacc0123 = _mm_min_ps(vacc0123, vy_max);
41 vacc4567 = _mm_min_ps(vacc4567, vy_max);
42
43 _mm_storeu_ps(y, vacc0123);
44 _mm_storeu_ps(y + 4, vacc4567);
45 y += 8;
46 }
47 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
48 __m128 vacc = _mm_loadu_ps(x);
49 x += 4;
50
51 vacc = _mm_max_ps(vacc, vy_min);
52 vacc = _mm_min_ps(vacc, vy_max);
53
54 _mm_storeu_ps(y, vacc);
55 y += 4;
56 }
57 if XNN_UNLIKELY(n != 0) {
58 __m128 vacc = _mm_loadu_ps(x);
59 vacc = _mm_max_ps(vacc, vy_min);
60 vacc = _mm_min_ps(vacc, vy_max);
61
62 if (n & (2 * sizeof(float))) {
63 _mm_storel_pi((__m64*) y, vacc);
64 vacc = _mm_movehl_ps(vacc, vacc);
65 y += 2;
66 }
67 if (n & (1 * sizeof(float))) {
68 _mm_store_ss(y, vacc);
69 }
70 }
71 }
72