1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vlrelu/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_vlrelu_ukernel__avx512f_x32(size_t n,const float * x,float * y,const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vlrelu_ukernel__avx512f_x32(
20     size_t n,
21     const float* x,
22     float* y,
23     const union xnn_f32_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(n != 0);
26   assert(n % sizeof(float) == 0);
27 
28   const __m512 vslope = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.slope));
29   const __m512 vzero = _mm512_setzero_ps();
30 
31   for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
32     __m512 vacc0123456789ABCDEF = _mm512_loadu_ps(x);
33     __m512 vaccGHIJKLMNOPQRSTUV = _mm512_loadu_ps(x + 16);
34     x += 32;
35 
36     const __mmask16 vsign0123456789ABCDEF = _mm512_cmp_ps_mask(vacc0123456789ABCDEF, vzero, _CMP_LT_OQ);
37     const __mmask16 vsignGHIJKLMNOPQRSTUV = _mm512_cmp_ps_mask(vaccGHIJKLMNOPQRSTUV, vzero, _CMP_LT_OQ);
38 
39     vacc0123456789ABCDEF = _mm512_mask_mul_ps(vacc0123456789ABCDEF, vsign0123456789ABCDEF, vacc0123456789ABCDEF, vslope);
40     vaccGHIJKLMNOPQRSTUV = _mm512_mask_mul_ps(vaccGHIJKLMNOPQRSTUV, vsignGHIJKLMNOPQRSTUV, vaccGHIJKLMNOPQRSTUV, vslope);
41 
42     _mm512_storeu_ps(y, vacc0123456789ABCDEF);
43     _mm512_storeu_ps(y + 16, vaccGHIJKLMNOPQRSTUV);
44     y += 32;
45   }
46   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
47     __m512 vacc = _mm512_loadu_ps(x);
48     x += 16;
49     const __mmask16 vsign = _mm512_cmp_ps_mask(vacc, vzero, _CMP_LT_OQ);
50     vacc = _mm512_mask_mul_ps(vacc, vsign, vacc, vslope);
51     _mm512_storeu_ps(y, vacc);
52     y += 16;
53   }
54   if XNN_UNLIKELY(n != 0) {
55     assert(n >= 1 * sizeof(float));
56     assert(n <= 15 * sizeof(float));
57     // Prepare mask for valid 32-bit elements (depends on n).
58     n >>= 2 /* log2(sizeof(float)) */;
59     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
60 
61     __m512 vacc = _mm512_maskz_loadu_ps(vmask, x);
62     const __mmask16 vsign = _mm512_mask_cmp_ps_mask(vmask, vacc, vzero, _CMP_LT_OQ);
63     vacc = _mm512_mask_mul_ps(vacc, vsign, vacc, vslope);
64     _mm512_mask_storeu_ps(y, vmask, vacc);
65   }
66 }
67