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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a);
34 a += 16;
35
36 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b);
37 b += 16;
38
39 __m512 vy0123456789ABCDEF = _mm512_add_ps(va0123456789ABCDEF, vb0123456789ABCDEF);
40
41 vy0123456789ABCDEF = _mm512_max_ps(vy0123456789ABCDEF, vy_min);
42
43 vy0123456789ABCDEF = _mm512_min_ps(vy0123456789ABCDEF, vy_max);
44
45 _mm512_storeu_ps(y, vy0123456789ABCDEF);
46 y += 16;
47 }
48 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
49 const __m512 va = _mm512_loadu_ps(a);
50 a += 16;
51
52 const __m512 vb = _mm512_loadu_ps(b);
53 b += 16;
54
55 __m512 vy = _mm512_add_ps(va, vb);
56 vy = _mm512_max_ps(vy, vy_min);
57 vy = _mm512_min_ps(vy, vy_max);
58 _mm512_storeu_ps(y, vy);
59 y += 16;
60 }
61 if XNN_UNLIKELY(n != 0) {
62 assert(n >= 1 * sizeof(float));
63 assert(n <= 15 * sizeof(float));
64 // Prepare mask for valid 32-bit elements (depends on n).
65 n >>= 2 /* log2(sizeof(float)) */;
66 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
67
68 const __m512 va = _mm512_maskz_loadu_ps(vmask, a);
69 const __m512 vb = _mm512_maskz_loadu_ps(vmask, b);
70
71 __m512 vy = _mm512_add_ps(va, vb);
72 vy = _mm512_max_ps(vy, vy_min);
73 vy = _mm512_min_ps(vy, vy_max);
74 _mm512_mask_storeu_ps(y, vmask, vy);
75 }
76 }
77