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_vsub_minmax_ukernel__avx512f_x32(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsub_minmax_ukernel__avx512f_x32(
20 size_t n,
21 const float* a,
22 const float* b,
23 float* y,
24 const union xnn_f32_minmax_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 const __m512 vy_min = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.min));
33 const __m512 vy_max = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.max));
34
35 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
36 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
37 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16);
38 a += 32;
39
40 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b);
41 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16);
42 b += 32;
43
44 __m512 vy0123456789ABCDEF = _mm512_sub_ps(va0123456789ABCDEF, vb0123456789ABCDEF);
45 __m512 vyGHIJKLMNOPQRSTUV = _mm512_sub_ps(vaGHIJKLMNOPQRSTUV, vbGHIJKLMNOPQRSTUV);
46
47
48 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min);
49 vyGHIJKLMNOPQRSTUV = _mm512_max_ps(vyGHIJKLMNOPQRSTUV, vy_min);
50
51 vy0123456789ABCDEF = _mm512_min_ps(vy0123456789ABCDEF, vy_max);
52 vyGHIJKLMNOPQRSTUV = _mm512_min_ps(vyGHIJKLMNOPQRSTUV, vy_max);
53
54 _mm512_storeu_ps(y, vy0123456789ABCDEF);
55 _mm512_storeu_ps(y + 16, vyGHIJKLMNOPQRSTUV);
56 y += 32;
57 }
58 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
59 const __m512 va = _mm512_loadu_ps(a);
60 a += 16;
61
62 const __m512 vb = _mm512_loadu_ps(b);
63 b += 16;
64
65 __m512 vy = _mm512_sub_ps(va, vb);
66 vy = _mm512_max_ps(vy, vy_min);
67 vy = _mm512_min_ps(vy, vy_max);
68 _mm512_storeu_ps(y, vy);
69 y += 16;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 assert(n >= 1 * sizeof(float));
73 assert(n <= 15 * sizeof(float));
74 // Prepare mask for valid 32-bit elements (depends on n).
75 n >>= 2 /* log2(sizeof(float)) */;
76 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
77
78 const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
79 const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
80
81 __m512 vy = _mm512_sub_ps(va, vb);
82 vy = _mm512_max_ps(vy, vy_min);
83 vy = _mm512_min_ps(vy, vy_max);
84 _mm512_mask_storeu_ps(y, vmask, vy);
85 }
86 }
87