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_x64(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x64(
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 >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
41 const __m512 vx0 = _mm512_loadu_ps(x);
42 const __m512 vx1 = _mm512_loadu_ps(x + 16);
43 const __m512 vx2 = _mm512_loadu_ps(x + 32);
44 const __m512 vx3 = _mm512_loadu_ps(x + 48);
45 x += 64;
46
47 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
48 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
49 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
50 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
51
52 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
53 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
54 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
55 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
56
57 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
58 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
59 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
60 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
61
62 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
63 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
64 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
65 vn3 = _mm512_sub_ps(vn3, vmagic_bias);
66
67 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
68 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
69 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
70 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
71
72 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
73 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
74 __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
75 __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
76
77 vp0 = _mm512_mul_ps(vp0, vt0);
78 vp1 = _mm512_mul_ps(vp1, vt1);
79 vp2 = _mm512_mul_ps(vp2, vt2);
80 vp3 = _mm512_mul_ps(vp3, vt3);
81
82 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
83 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
84 vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
85 vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
86
87 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
88 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
89 vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
90 vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
91
92 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
93 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
94 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
95 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
96
97 const __m512 vd0 = _mm512_add_ps(ve0, vone);
98 const __m512 vd1 = _mm512_add_ps(ve1, vone);
99 const __m512 vd2 = _mm512_add_ps(ve2, vone);
100 const __m512 vd3 = _mm512_add_ps(ve3, vone);
101
102 __m512 vf0 = _mm512_div_ps(ve0, vd0);
103 __m512 vf1 = _mm512_div_ps(ve1, vd1);
104 __m512 vf2 = _mm512_div_ps(ve2, vd2);
105 __m512 vf3 = _mm512_div_ps(ve3, vd3);
106
107 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
108 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
109 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
110 vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
111
112 _mm512_storeu_ps(y, vf0);
113 _mm512_storeu_ps(y + 16, vf1);
114 _mm512_storeu_ps(y + 32, vf2);
115 _mm512_storeu_ps(y + 48, vf3);
116 y += 64;
117 }
118 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
119 const __m512 vx = _mm512_loadu_ps(x);
120 x += 16;
121
122 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
123
124 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
125 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
126 vn = _mm512_sub_ps(vn, vmagic_bias);
127
128 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
129
130 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
131 vp = _mm512_mul_ps(vp, vt);
132 vp = _mm512_fmadd_ps(vt, vp, vt);
133 vp = _mm512_fmadd_ps(vl, vp, vl);
134
135 const __m512 ve = _mm512_scalef_ps(vp, vn);
136 const __m512 vd = _mm512_add_ps(ve, vone);
137
138 __m512 vf = _mm512_div_ps(ve, vd);
139
140 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
141
142 _mm512_storeu_ps(y, vf);
143 y += 16;
144 }
145 if XNN_UNLIKELY(n != 0) {
146 assert(n >= 1 * sizeof(float));
147 assert(n <= 15 * sizeof(float));
148
149 // Prepare mask for valid 32-bit elements (depends on n).
150 n >>= 2 /* log2(sizeof(float)) */;
151 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
152
153 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
154 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
155
156 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
157 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
158 vn = _mm512_sub_ps(vn, vmagic_bias);
159
160 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
161
162 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
163 vp = _mm512_mul_ps(vp, vt);
164 vp = _mm512_fmadd_ps(vt, vp, vt);
165 vp = _mm512_fmadd_ps(vl, vp, vl);
166
167 const __m512 ve = _mm512_scalef_ps(vp, vn);
168 const __m512 vd = _mm512_add_ps(ve, vone);
169
170 __m512 vf = _mm512_div_ps(ve, vd);
171
172 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
173
174 _mm512_mask_storeu_ps(y, vmask, vf);
175 }
176 }
177