• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vbinary/vopc-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_vsqrdiffc_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_vsqrdiffc_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   const __m512 vb = _mm512_set1_ps(*b);
34   for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
35     const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
36     const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16);
37     a += 32;
38 
39     __m512 vy0123456789ABCDEF = _mm512_sub_ps(va0123456789ABCDEF, vb);
40     __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vb);
41 
42     vy0123456789ABCDEF = _mm512_mul_ps(vy0123456789ABCDEF, vy0123456789ABCDEF);
43     vyGHIJKLMNOPQRSTUV = _mm512_mul_ps(vyGHIJKLMNOPQRSTUV, vyGHIJKLMNOPQRSTUV);
44 
45 
46     _mm512_storeu_ps(y, vy0123456789ABCDEF);
47     _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
48     y += 32;
49   }
50   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
51     const __m512 va = _mm512_loadu_ps(a);
52     a += 16;
53 
54     __m512 vy = _mm512_sub_ps(va, vb);
55     vy = _mm512_mul_ps(vy, vy);
56     _mm512_storeu_ps(y, vy);
57     y += 16;
58   }
59   if XNN_UNLIKELY(n != 0) {
60     assert(n >= 1 * sizeof(float));
61     assert(n <= 15 * sizeof(float));
62     // Prepare mask for valid 32-bit elements (depends on n).
63     n >>= 2 /* log2(sizeof(float)) */;
64     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
65 
66     const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
67 
68     __m512 vy = _mm512_sub_ps(va, vb);
69     vy = _mm512_mul_ps(vy, vy);
70     _mm512_mask_storeu_ps(y, vmask, vy);
71   }
72 }
73