1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vop-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>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17
18
xnn_f32_vsqrdiff_ukernel__sse_x4(size_t n,const float * a,const float * b,float * y,const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsqrdiff_ukernel__sse_x4(
20 size_t n,
21 const float* a,
22 const float* b,
23 float* y,
24 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28 assert(a != NULL);
29 assert(b != NULL);
30 assert(y != NULL);
31
32
33 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
34 const __m128 va0123 = _mm_loadu_ps(a);
35 a += 4;
36
37 const __m128 vb0123 = _mm_loadu_ps(b);
38 b += 4;
39
40 __m128 vy0123 = _mm_sub_ps(va0123, vb0123);
41
42 vy0123 = _mm_mul_ps(vy0123, vy0123);
43
44
45 _mm_storeu_ps(y, vy0123);
46 y += 4;
47 }
48 if XNN_UNLIKELY(n != 0) {
49 const __m128 va0123 = _mm_loadu_ps(a);
50 const __m128 vb0123 = _mm_loadu_ps(b);
51
52 __m128 vy0123 = _mm_sub_ps(va0123, vb0123);
53 vy0123 = _mm_mul_ps(vy0123, vy0123);
54 if (n & (2 * sizeof(float))) {
55 _mm_storel_pi((__m64*) y, vy0123);
56 vy0123 = _mm_movehl_ps(vy0123, vy0123);
57 y += 2;
58 }
59 if (n & (1 * sizeof(float))) {
60 _mm_store_ss(y, vy0123);
61 }
62 }
63 }
64