1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5
6 #include <assert.h>
7
8 #include <emmintrin.h>
9
10 #include <xnnpack/clamp.h>
11
12
xnn_f32_clamp_ukernel__sse(size_t n,const float * x,float * y,const union xnn_f32_output_params params[restrict static1])13 void xnn_f32_clamp_ukernel__sse(
14 size_t n,
15 const float* x,
16 float* y,
17 const union xnn_f32_output_params params[restrict static 1])
18 {
19 assert(n != 0);
20 assert(n % sizeof(float) == 0);
21
22 const __m128 voutput_max = _mm_load_ps(params->sse.max);
23 const __m128 voutput_min = _mm_load_ps(params->sse.min);
24
25 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
26 const __m128 vx = _mm_loadu_ps(x);
27 x += 4;
28
29 const __m128 vy = _mm_min_ps(_mm_max_ps(vx, voutput_min), voutput_max);
30
31 _mm_storeu_ps(y, vy);
32 y += 4;
33 }
34 if (n != 0) {
35 const __m128 vx = _mm_loadu_ps(x);
36
37 __m128 vy = _mm_min_ps(_mm_max_ps(vx, voutput_min), voutput_max);
38
39 if (n & 2 * sizeof(float)) {
40 _mm_storel_pi((__m64*) y, vy);
41 vy = _mm_movehl_ps(vy, vy);
42 y += 2;
43 }
44 if (n & 1 * sizeof(float)) {
45 _mm_store_ss(y, vy);
46 }
47 }
48 }
49