• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/avx512f-lut16-p3-perm-scalef.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_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x16(
20     size_t n,
21     const float* x,
22     float* y,
23     const void* params)
24 {
25   assert(n % sizeof(float) == 0);
26 
27   const __m512i vsign_mask = _mm512_set1_epi32(0x80000000);
28   const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f);
29   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f);
30   const __m512 vtable = _mm512_set_ps(
31     0x1.EA4AFAp+0f, 0x1.D5818Ep+0f, 0x1.C199BEp+0f, 0x1.AE89FAp+0f,
32     0x1.9C4918p+0f, 0x1.8ACE54p+0f, 0x1.7A1148p+0f, 0x1.6A09E6p+0f,
33     0x1.5AB07Ep+0f, 0x1.4BFDAEp+0f, 0x1.3DEA64p+0f, 0x1.306FE0p+0f,
34     0x1.2387A6p+0f, 0x1.172B84p+0f, 0x1.0B5586p+0f, 0x1.000000p+0f);
35   const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
36   const __m512 vc3 = _mm512_set1_ps(0x1.55559Ap-3f);
37   const __m512 vc2 = _mm512_set1_ps(0x1.00021Ep-1f);
38   const __m512 vone = _mm512_set1_ps(1.0f);
39 
40   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
41     const __m512 vx = _mm512_loadu_ps(x);
42     x += 16;
43 
44     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
45 
46     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
47     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
48     vn = _mm512_sub_ps(vn, vmagic_bias);
49 
50     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
51 
52     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
53     vp = _mm512_mul_ps(vp, vt);
54     vp = _mm512_fmadd_ps(vt, vp, vt);
55     vp = _mm512_fmadd_ps(vl, vp, vl);
56 
57     const __m512 ve = _mm512_scalef_ps(vp, vn);
58     const __m512 vd = _mm512_add_ps(ve, vone);
59 
60     __m512 vf = _mm512_div_ps(ve, vd);
61 
62     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
63 
64     _mm512_storeu_ps(y, vf);
65     y += 16;
66   }
67   if XNN_UNLIKELY(n != 0) {
68     assert(n >= 1 * sizeof(float));
69     assert(n <= 15 * sizeof(float));
70 
71     // Prepare mask for valid 32-bit elements (depends on n).
72     n >>= 2 /* log2(sizeof(float)) */;
73     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
74 
75     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
76     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
77 
78     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
79     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
80     vn = _mm512_sub_ps(vn, vmagic_bias);
81 
82     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
83 
84     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
85     vp = _mm512_mul_ps(vp, vt);
86     vp = _mm512_fmadd_ps(vt, vp, vt);
87     vp = _mm512_fmadd_ps(vl, vp, vl);
88 
89     const __m512 ve = _mm512_scalef_ps(vp, vn);
90     const __m512 vd = _mm512_add_ps(ve, vone);
91 
92     __m512 vf = _mm512_div_ps(ve, vd);
93 
94     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
95 
96     _mm512_mask_storeu_ps(y, vmask, vf);
97   }
98 }
99