1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vunary/avx512f.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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/vunary.h>
17
18
xnn_f32_vabs_ukernel__avx512f_x32(size_t n,const float * x,float * y,const union xnn_f32_abs_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vabs_ukernel__avx512f_x32(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_abs_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(x != NULL);
28 assert(y != NULL);
29
30 const __m512i vnonsign_mask = _mm512_set1_epi32((int) params->avx512.nonsign_mask);
31 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
32 const __m512i vx0123456789ABCDEF = _mm512_loadu_si512(x);
33 const __m512i vxGHIJKLMNOPQRSTUV = _mm512_loadu_si512(x + 16);
34 x += 32;
35
36 const __m512i vy0123456789ABCDEF = _mm512_and_epi32(vx0123456789ABCDEF, vnonsign_mask);
37 const __m512i vyGHIJKLMNOPQRSTUV = _mm512_and_epi32(vxGHIJKLMNOPQRSTUV, vnonsign_mask);
38
39 _mm512_storeu_si512(y, vy0123456789ABCDEF);
40 _mm512_storeu_si512(y + 16, vyGHIJKLMNOPQRSTUV);
41 y += 32;
42 }
43 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
44 const __m512i vx = _mm512_loadu_si512(x);
45 x += 16;
46
47 const __m512i vy = _mm512_and_epi32(vx, vnonsign_mask);
48
49 _mm512_storeu_si512(y, vy);
50 y += 16;
51 }
52 if XNN_UNLIKELY(n != 0) {
53 assert(n >= 1 * sizeof(float));
54 assert(n <= 15 * sizeof(float));
55 // Prepare mask for valid 32-bit elements (depends on n).
56 n >>= 2 /* log2(sizeof(float)) */;
57 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
58
59 const __m512i vx = _mm512_maskz_loadu_epi32(vmask, x);
60 const __m512i vy = _mm512_and_epi32(vx, vnonsign_mask);
61 _mm512_mask_storeu_epi32(y, vmask, vy);
62 }
63 }
64