1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrelu/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/vunary.h>
15 #include <xnnpack/common.h>
16 #include <xnnpack/intrinsics-polyfill.h>
17
18
xnn_f32_vrelu_ukernel__avx512f_x32(size_t n,const float * x,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vrelu_ukernel__avx512f_x32(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_relu_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 __m512 vzero = _mm512_setzero_ps();
31
32 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
33 __m512 vacc0123456789ABCDEF = _mm512_loadu_ps(x);
34 __m512 vaccGHIJKLMNOPQRSTUV = _mm512_loadu_ps(x + 16);
35 x += 32;
36
37 vacc0123456789ABCDEF = _mm512_max_ps(vacc0123456789ABCDEF, vzero);
38 vaccGHIJKLMNOPQRSTUV = _mm512_max_ps(vaccGHIJKLMNOPQRSTUV, vzero);
39
40 _mm512_storeu_ps(y, vacc0123456789ABCDEF);
41 _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV);
42 y += 32;
43 }
44 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
45 __m512 vacc = _mm512_loadu_ps(x);
46 x += 16;
47
48 vacc = _mm512_max_ps(vacc, vzero);
49
50 _mm512_storeu_ps(y, vacc);
51 y += 16;
52 }
53 if XNN_UNLIKELY(n != 0) {
54 assert(n >= 1 * sizeof(float));
55 assert(n <= 15 * sizeof(float));
56 // Prepare mask for valid 32-bit elements (depends on n).
57 n >>= 2 /* log2(sizeof(float)) */;
58 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
59
60 __m512 vacc = _mm512_maskz_loadu_ps(vmask, x);
61 vacc = _mm512_max_ps(vacc, vzero);
62 _mm512_mask_storeu_ps(y, vmask, vacc);
63 }
64 }
65