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_nr1fma_x64(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_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 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 >= 64 * sizeof(float); n -= 64 * 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 const __m512 vx3 = _mm512_loadu_ps(x + 48);
42 x += 64;
43
44 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
45 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
46 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
47 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
48
49 __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
50 __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
51 __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
52 __m512 vn3 = _mm512_mul_ps(vz3, vlog2e);
53
54 vn0 = _mm512_roundscale_ps(vn0, 0);
55 vn1 = _mm512_roundscale_ps(vn1, 0);
56 vn2 = _mm512_roundscale_ps(vn2, 0);
57 vn3 = _mm512_roundscale_ps(vn3, 0);
58
59 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
60 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
61 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
62 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
63
64 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
65 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
66 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
67 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
68
69 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
70 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
71 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
72 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
73
74 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
75 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
76 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
77 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
78
79 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
80 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
81 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
82 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
83
84 vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
85 vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
86 vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
87 vp3 = _mm512_fmadd_ps(vp3, vt3, vone);
88
89 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
90 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
91 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
92 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
93
94 const __m512 vd0 = _mm512_add_ps(ve0, vone);
95 const __m512 vd1 = _mm512_add_ps(ve1, vone);
96 const __m512 vd2 = _mm512_add_ps(ve2, vone);
97 const __m512 vd3 = _mm512_add_ps(ve3, vone);
98
99 __m512 vr0 = _mm512_rcp14_ps(vd0);
100 __m512 vr1 = _mm512_rcp14_ps(vd1);
101 __m512 vr2 = _mm512_rcp14_ps(vd2);
102 __m512 vr3 = _mm512_rcp14_ps(vd3);
103
104 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
105 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
106 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
107 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
108
109 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
110 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
111 __m512 vf2 = _mm512_mul_ps(ve2, vr2);
112 __m512 vf3 = _mm512_mul_ps(ve3, vr3);
113
114
115 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
116 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
117 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
118 vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
119
120 _mm512_storeu_ps(y, vf0);
121 _mm512_storeu_ps(y + 16, vf1);
122 _mm512_storeu_ps(y + 32, vf2);
123 _mm512_storeu_ps(y + 48, vf3);
124 y += 64;
125 }
126 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
127 const __m512 vx = _mm512_loadu_ps(x);
128 x += 16;
129
130 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
131
132 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
133
134 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
135
136 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
137 vp = _mm512_fmadd_ps(vp, vt, vc3);
138 vp = _mm512_fmadd_ps(vp, vt, vc2);
139 vp = _mm512_fmadd_ps(vp, vt, vc1);
140 vp = _mm512_fmadd_ps(vp, vt, vone);
141
142 const __m512 ve = _mm512_scalef_ps(vp, vn);
143 const __m512 vd = _mm512_add_ps(ve, vone);
144
145 __m512 vr = _mm512_rcp14_ps(vd);
146 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
147
148 __m512 vf = _mm512_mul_ps(ve, vr);
149
150 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
151
152 _mm512_storeu_ps(y, vf);
153 y += 16;
154 }
155 if XNN_UNLIKELY(n != 0) {
156 assert(n >= 1 * sizeof(float));
157 assert(n <= 15 * sizeof(float));
158
159 // Prepare mask for valid 32-bit elements (depends on n).
160 n >>= 2 /* log2(sizeof(float)) */;
161 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
162
163 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
164 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
165
166 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
167
168 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
169
170 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
171 vp = _mm512_fmadd_ps(vp, vt, vc3);
172 vp = _mm512_fmadd_ps(vp, vt, vc2);
173 vp = _mm512_fmadd_ps(vp, vt, vc1);
174 vp = _mm512_fmadd_ps(vp, vt, vone);
175
176 const __m512 ve = _mm512_scalef_ps(vp, vn);
177 const __m512 vd = _mm512_add_ps(ve, vone);
178
179 __m512 vr = _mm512_rcp14_ps(vd);
180 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
181
182 __m512 vf = _mm512_mul_ps(ve, vr);
183
184 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
185
186 _mm512_mask_storeu_ps(y, vmask, vf);
187 }
188 }
189