1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx512f-rr1-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_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x32(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n % sizeof(float) == 0);
26
27 const __m512i vsign_mask = _mm512_set1_epi32((int) params->avx512_rr1_lut16_p3.sign_mask);
28 const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias);
29 const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e);
30 const __m512 vtable = _mm512_load_ps(params->avx512_rr1_lut16_p3.table);
31 const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2);
32 const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3);
33 const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2);
34 const __m512 vone = _mm512_set1_ps(params->avx512_rr1_lut16_p3.one);
35
36 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
37 const __m512 vx0 = _mm512_loadu_ps(x);
38 const __m512 vx1 = _mm512_loadu_ps(x + 16);
39 x += 32;
40
41 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
42 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
43
44 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
45 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
46
47 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
48 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
49
50 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
51 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
52
53 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
54 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
55
56 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
57 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
58
59 vp0 = _mm512_mul_ps(vp0, vt0);
60 vp1 = _mm512_mul_ps(vp1, vt1);
61
62 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
63 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
64
65 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
66 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
67
68 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
69 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
70
71 const __m512 vd0 = _mm512_add_ps(ve0, vone);
72 const __m512 vd1 = _mm512_add_ps(ve1, vone);
73
74 __m512 vf0 = _mm512_div_ps(ve0, vd0);
75 __m512 vf1 = _mm512_div_ps(ve1, vd1);
76
77 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
78 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
79
80 _mm512_storeu_ps(y, vf0);
81 _mm512_storeu_ps(y + 16, vf1);
82 y += 32;
83 }
84 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
85 const __m512 vx = _mm512_loadu_ps(x);
86 x += 16;
87
88 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
89
90 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
91 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
92 vn = _mm512_sub_ps(vn, vmagic_bias);
93
94 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
95
96 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
97 vp = _mm512_mul_ps(vp, vt);
98 vp = _mm512_fmadd_ps(vt, vp, vt);
99 vp = _mm512_fmadd_ps(vl, vp, vl);
100
101 const __m512 ve = _mm512_scalef_ps(vp, vn);
102 const __m512 vd = _mm512_add_ps(ve, vone);
103
104 __m512 vf = _mm512_div_ps(ve, vd);
105
106 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
107
108 _mm512_storeu_ps(y, vf);
109 y += 16;
110 }
111 if XNN_UNLIKELY(n != 0) {
112 assert(n >= 1 * sizeof(float));
113 assert(n <= 15 * sizeof(float));
114
115 // Prepare mask for valid 32-bit elements (depends on n).
116 n >>= 2 /* log2(sizeof(float)) */;
117 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
118
119 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
120 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
121
122 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
123 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
124 vn = _mm512_sub_ps(vn, vmagic_bias);
125
126 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
127
128 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
129 vp = _mm512_mul_ps(vp, vt);
130 vp = _mm512_fmadd_ps(vt, vp, vt);
131 vp = _mm512_fmadd_ps(vl, vp, vl);
132
133 const __m512 ve = _mm512_scalef_ps(vp, vn);
134 const __m512 vd = _mm512_add_ps(ve, vone);
135
136 __m512 vf = _mm512_div_ps(ve, vd);
137
138 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
139
140 _mm512_mask_storeu_ps(y, vmask, vf);
141 }
142 }
143