• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_vadd_ukernel__avx512f_x32(size_t n,const float * a,const float * b,float * y,const union xnn_f32_output_params params[restrict static1])19 void xnn_f32_vadd_ukernel__avx512f_x32(
20     size_t n,
21     const float* a,
22     const float* b,
23     float* y,
24     const union xnn_f32_output_params params[restrict static 1])
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28 
29   const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
30   const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
31 
32   for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
33     const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
34     const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16);
35     a += 32;
36 
37     const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b);
38     const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16);
39     b += 32;
40 
41     __m512 vy0123456789ABCDEF = _mm512_add_ps(va0123456789ABCDEF, vb0123456789ABCDEF);
42     __m512 vyGHIJKLMNOPQRSTUV = _mm512_add_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV);
43 
44     vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min);
45     vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min);
46 
47     vy0123456789ABCDEF = _mm512_min_ps(vy0123456789ABCDEF, vy_max);
48     vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max);
49 
50     _mm512_storeu_ps(y, vy0123456789ABCDEF);
51     _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
52     y += 32;
53   }
54   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
55     const __m512 va = _mm512_loadu_ps(a);
56     a += 16;
57 
58     const __m512 vb = _mm512_loadu_ps(b);
59     b += 16;
60 
61     __m512 vy = _mm512_add_ps(va, vb);
62     vy = _mm512_max_ps(vy, vy_min);
63     vy = _mm512_min_ps(vy, vy_max);
64     _mm512_storeu_ps(y, vy);
65     y += 16;
66   }
67   if XNN_UNLIKELY(n != 0) {
68     assert(n >= 1 * sizeof(float));
69     assert(n <= 15 * sizeof(float));
70     // Prepare mask for valid 32-bit elements (depends on n).
71     n >>= 2 /* log2(sizeof(float)) */;
72     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
73 
74     const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
75     const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
76 
77     __m512 vy = _mm512_add_ps(va, vb);
78     vy = _mm512_max_ps(vy, vy_min);
79     vy = _mm512_min_ps(vy, vy_max);
80     _mm512_mask_storeu_ps(y, vmask, vy);
81   }
82 }
83