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_x48(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_x48(
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 >= 48 * sizeof(float); n -= 48 * sizeof(float)) {
37 const __m512 vx0 = _mm512_loadu_ps(x);
38 const __m512 vx1 = _mm512_loadu_ps(x + 16);
39 const __m512 vx2 = _mm512_loadu_ps(x + 32);
40 x += 48;
41
42 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
43 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
44 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
45
46 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
47 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
48 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
49
50 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
51 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
52 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
53
54 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
55 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
56 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
57
58 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
59 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
60 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
61
62 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
63 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
64 __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
65
66 vp0 = _mm512_mul_ps(vp0, vt0);
67 vp1 = _mm512_mul_ps(vp1, vt1);
68 vp2 = _mm512_mul_ps(vp2, vt2);
69
70 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
71 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
72 vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
73
74 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
75 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
76 vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
77
78 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
79 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
80 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
81
82 const __m512 vd0 = _mm512_add_ps(ve0, vone);
83 const __m512 vd1 = _mm512_add_ps(ve1, vone);
84 const __m512 vd2 = _mm512_add_ps(ve2, vone);
85
86 __m512 vf0 = _mm512_div_ps(ve0, vd0);
87 __m512 vf1 = _mm512_div_ps(ve1, vd1);
88 __m512 vf2 = _mm512_div_ps(ve2, vd2);
89
90 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
91 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
92 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
93
94 _mm512_storeu_ps(y, vf0);
95 _mm512_storeu_ps(y + 16, vf1);
96 _mm512_storeu_ps(y + 32, vf2);
97 y += 48;
98 }
99 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
100 const __m512 vx = _mm512_loadu_ps(x);
101 x += 16;
102
103 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
104
105 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
106 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
107 vn = _mm512_sub_ps(vn, vmagic_bias);
108
109 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
110
111 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
112 vp = _mm512_mul_ps(vp, vt);
113 vp = _mm512_fmadd_ps(vt, vp, vt);
114 vp = _mm512_fmadd_ps(vl, vp, vl);
115
116 const __m512 ve = _mm512_scalef_ps(vp, vn);
117 const __m512 vd = _mm512_add_ps(ve, vone);
118
119 __m512 vf = _mm512_div_ps(ve, vd);
120
121 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
122
123 _mm512_storeu_ps(y, vf);
124 y += 16;
125 }
126 if XNN_UNLIKELY(n != 0) {
127 assert(n >= 1 * sizeof(float));
128 assert(n <= 15 * sizeof(float));
129
130 // Prepare mask for valid 32-bit elements (depends on n).
131 n >>= 2 /* log2(sizeof(float)) */;
132 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
133
134 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
135 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
136
137 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
138 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
139 vn = _mm512_sub_ps(vn, vmagic_bias);
140
141 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
142
143 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
144 vp = _mm512_mul_ps(vp, vt);
145 vp = _mm512_fmadd_ps(vt, vp, vt);
146 vp = _mm512_fmadd_ps(vl, vp, vl);
147
148 const __m512 ve = _mm512_scalef_ps(vp, vn);
149 const __m512 vd = _mm512_add_ps(ve, vone);
150
151 __m512 vf = _mm512_div_ps(ve, vd);
152
153 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
154
155 _mm512_mask_storeu_ps(y, vmask, vf);
156 }
157 }
158