1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrnd/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/math.h>
17 #include <xnnpack/vunary.h>
18
19
xnn_f32_vrndne_ukernel__avx512f_x16(size_t n,const float * x,float * y,const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vrndne_ukernel__avx512f_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
25 {
26 assert(n != 0);
27 assert(n % sizeof(float) == 0);
28
29 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
30 const __m512 vx0123456789ABCDEF = _mm512_loadu_ps(x);
31 x += 16;
32
33 const __m512 vy0123456789ABCDEF = _mm512_roundscale_ps(vx0123456789ABCDEF, _MM_FROUND_TO_NEAREST_INT);
34
35 _mm512_storeu_ps(y, vy0123456789ABCDEF);
36 y += 16;
37 }
38 if XNN_UNLIKELY(n != 0) {
39 assert(n >= 1 * sizeof(float));
40 assert(n <= 15 * sizeof(float));
41 // Prepare mask for valid 32-bit elements (depends on n).
42 n >>= 2 /* log2(sizeof(float)) */;
43 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
44
45 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
46 const __m512 vy = _mm512_maskz_roundscale_ps(vmask, vx, _MM_FROUND_TO_NEAREST_INT);
47 _mm512_mask_storeu_ps(y, vmask, vy);
48 }
49 }
50