1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/avx512f-p5-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_p5_scalef_div_x48(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x48(
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 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
29 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
30 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
31 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
32 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
34 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
35 const __m512 vone = _mm512_set1_ps(1.0f);
36
37 for (; n >= 48 * sizeof(float); n -= 48 * sizeof(float)) {
38 const __m512 vx0 = _mm512_loadu_ps(x);
39 const __m512 vx1 = _mm512_loadu_ps(x + 16);
40 const __m512 vx2 = _mm512_loadu_ps(x + 32);
41 x += 48;
42
43 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
44 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
45 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
46
47 __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
48 __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
49 __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
50
51 vn0 = _mm512_roundscale_ps(vn0, 0);
52 vn1 = _mm512_roundscale_ps(vn1, 0);
53 vn2 = _mm512_roundscale_ps(vn2, 0);
54
55 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
56 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
57 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
58
59 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
60 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
61 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
62
63 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
64 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
65 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
66
67 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
68 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
69 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
70
71 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
72 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
73 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
74
75 vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
76 vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
77 vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
78
79 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
80 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
81 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
82
83 const __m512 vd0 = _mm512_add_ps(ve0, vone);
84 const __m512 vd1 = _mm512_add_ps(ve1, vone);
85 const __m512 vd2 = _mm512_add_ps(ve2, vone);
86
87 __m512 vf0 = _mm512_div_ps(ve0, vd0);
88 __m512 vf1 = _mm512_div_ps(ve1, vd1);
89 __m512 vf2 = _mm512_div_ps(ve2, vd2);
90
91 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
92 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
93 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
94
95 _mm512_storeu_ps(y, vf0);
96 _mm512_storeu_ps(y + 16, vf1);
97 _mm512_storeu_ps(y + 32, vf2);
98 y += 48;
99 }
100 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
101 const __m512 vx = _mm512_loadu_ps(x);
102 x += 16;
103
104 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
105
106 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
107
108 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
109
110 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
111 vp = _mm512_fmadd_ps(vp, vt, vc3);
112 vp = _mm512_fmadd_ps(vp, vt, vc2);
113 vp = _mm512_fmadd_ps(vp, vt, vc1);
114 vp = _mm512_fmadd_ps(vp, vt, vone);
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 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
138
139 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
140
141 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
142 vp = _mm512_fmadd_ps(vp, vt, vc3);
143 vp = _mm512_fmadd_ps(vp, vt, vc2);
144 vp = _mm512_fmadd_ps(vp, vt, vc1);
145 vp = _mm512_fmadd_ps(vp, vt, vone);
146
147 const __m512 ve = _mm512_scalef_ps(vp, vn);
148 const __m512 vd = _mm512_add_ps(ve, vone);
149
150 __m512 vf = _mm512_div_ps(ve, vd);
151
152 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
153
154 _mm512_mask_storeu_ps(y, vmask, vf);
155 }
156 }
157