1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrelu/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/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_vrelu_ukernel__sse_x8(size_t n,const float * x,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vrelu_ukernel__sse_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vzero = _mm_setzero_ps();
30
31 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
32 __m128 vacc0123 = _mm_loadu_ps(x);
33 __m128 vacc4567 = _mm_loadu_ps(x + 4);
34 x += 8;
35
36 vacc0123 = _mm_max_ps(vacc0123, vzero);
37 vacc4567 = _mm_max_ps(vacc4567, vzero);
38
39 _mm_storeu_ps(y, vacc0123);
40 _mm_storeu_ps(y + 4, vacc4567);
41 y += 8;
42 }
43 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
44 __m128 vacc = _mm_loadu_ps(x);
45 x += 4;
46
47 vacc = _mm_max_ps(vacc, vzero);
48
49 _mm_storeu_ps(y, vacc);
50 y += 4;
51 }
52 if XNN_UNLIKELY(n != 0) {
53 __m128 vacc = _mm_loadu_ps(x);
54 vacc = _mm_max_ps(vacc, vzero);
55
56 if (n & (2 * sizeof(float))) {
57 _mm_storel_pi((__m64*) y, vacc);
58 vacc = _mm_movehl_ps(vacc, vacc);
59 y += 2;
60 }
61 if (n & (1 * sizeof(float))) {
62 _mm_store_ss(y, vacc);
63 }
64 }
65 }
66