1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vop-avx512f.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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vbinary.h>
17
18
xnn_f32_vsqrdiff_ukernel__avx512f_x32(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__avx512f_x32(
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)])
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 >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
34 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
35 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16);
36 a += 32;
37
38 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b);
39 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16);
40 b += 32;
41
42 __m512 vy0123456789ABCDEF = _mm512_sub_ps(va0123456789ABCDEF, vb0123456789ABCDEF);
43 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV);
44
45 vy0123456789ABCDEF = _mm512_mul_ps(vy0123456789ABCDEF, vy0123456789ABCDEF);
46 vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV);
47
48
49 _mm512_storeu_ps(y, vy0123456789ABCDEF);
50 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
51 y += 32;
52 }
53 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
54 const __m512 va = _mm512_loadu_ps(a);
55 a += 16;
56
57 const __m512 vb = _mm512_loadu_ps(b);
58 b += 16;
59
60 __m512 vy = _mm512_sub_ps(va, vb);
61 vy = _mm512_mul_ps(vy, vy);
62 _mm512_storeu_ps(y, vy);
63 y += 16;
64 }
65 if XNN_UNLIKELY(n != 0) {
66 assert(n >= 1 * sizeof(float));
67 assert(n <= 15 * sizeof(float));
68 // Prepare mask for valid 32-bit elements (depends on n).
69 n >>= 2 /* log2(sizeof(float)) */;
70 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
71
72 const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
73 const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
74
75 __m512 vy = _mm512_sub_ps(va, vb);
76 vy = _mm512_mul_ps(vy, vy);
77 _mm512_mask_storeu_ps(y, vmask, vy);
78 }
79 }
80