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_nr1fma_x32(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x32(
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 >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
41 const __m512 vx0 = _mm512_loadu_ps(x);
42 const __m512 vx1 = _mm512_loadu_ps(x + 16);
43 x += 32;
44
45 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
46 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
47
48 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
49 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
50
51 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
52 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
53
54 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
55 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
56
57 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
58 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
59
60 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
61 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
62
63 vp0 = _mm512_mul_ps(vp0, vt0);
64 vp1 = _mm512_mul_ps(vp1, vt1);
65
66 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
67 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
68
69 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
70 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
71
72 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
73 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
74
75 const __m512 vd0 = _mm512_add_ps(ve0, vone);
76 const __m512 vd1 = _mm512_add_ps(ve1, vone);
77
78 __m512 vr0 = _mm512_rcp14_ps(vd0);
79 __m512 vr1 = _mm512_rcp14_ps(vd1);
80
81 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
82 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
83
84 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
85 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
86
87
88 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
89 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
90
91 _mm512_storeu_ps(y, vf0);
92 _mm512_storeu_ps(y + 16, vf1);
93 y += 32;
94 }
95 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
96 const __m512 vx = _mm512_loadu_ps(x);
97 x += 16;
98
99 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
100
101 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
102 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
103 vn = _mm512_sub_ps(vn, vmagic_bias);
104
105 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
106
107 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
108 vp = _mm512_mul_ps(vp, vt);
109 vp = _mm512_fmadd_ps(vt, vp, vt);
110 vp = _mm512_fmadd_ps(vl, vp, vl);
111
112 const __m512 ve = _mm512_scalef_ps(vp, vn);
113 const __m512 vd = _mm512_add_ps(ve, vone);
114
115 __m512 vr = _mm512_rcp14_ps(vd);
116 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
117
118 __m512 vf = _mm512_mul_ps(ve, vr);
119
120 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
121
122 _mm512_storeu_ps(y, vf);
123 y += 16;
124 }
125 if XNN_UNLIKELY(n != 0) {
126 assert(n >= 1 * sizeof(float));
127 assert(n <= 15 * sizeof(float));
128
129 // Prepare mask for valid 32-bit elements (depends on n).
130 n >>= 2 /* log2(sizeof(float)) */;
131 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
132
133 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
134 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
135
136 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
137 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
138 vn = _mm512_sub_ps(vn, vmagic_bias);
139
140 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
141
142 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
143 vp = _mm512_mul_ps(vp, vt);
144 vp = _mm512_fmadd_ps(vt, vp, vt);
145 vp = _mm512_fmadd_ps(vl, vp, vl);
146
147 const __m512 ve = _mm512_scalef_ps(vp, vn);
148 const __m512 vd = _mm512_add_ps(ve, vone);
149
150 __m512 vr = _mm512_rcp14_ps(vd);
151 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
152
153 __m512 vf = _mm512_mul_ps(ve, vr);
154
155 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
156
157 _mm512_mask_storeu_ps(y, vmask, vf);
158 }
159 }
160