1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx512f-rr1-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_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x112(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_p5_scalef_nr1fma_x112(
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_p5.sign_mask);
28 const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_p5.log2e);
29 const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_p5.minus_ln2);
30 const __m512 vc5 = _mm512_set1_ps(params->avx512_rr1_p5.c5);
31 const __m512 vc4 = _mm512_set1_ps(params->avx512_rr1_p5.c4);
32 const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_p5.c3);
33 const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_p5.c2);
34 const __m512 vc1 = _mm512_set1_ps(params->avx512_rr1_p5.c1);
35 const __m512 vone = _mm512_set1_ps(params->avx512_rr1_p5.one);
36
37 for (; n >= 112 * sizeof(float); n -= 112 * 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 const __m512 vx4 = _mm512_loadu_ps(x + 64);
43 const __m512 vx5 = _mm512_loadu_ps(x + 80);
44 const __m512 vx6 = _mm512_loadu_ps(x + 96);
45 x += 112;
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 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
52 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
53 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
54
55 __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
56 __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
57 __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
58 __m512 vn3 = _mm512_mul_ps(vz3, vlog2e);
59 __m512 vn4 = _mm512_mul_ps(vz4, vlog2e);
60 __m512 vn5 = _mm512_mul_ps(vz5, vlog2e);
61 __m512 vn6 = _mm512_mul_ps(vz6, vlog2e);
62
63 vn0 = _mm512_roundscale_ps(vn0, 0);
64 vn1 = _mm512_roundscale_ps(vn1, 0);
65 vn2 = _mm512_roundscale_ps(vn2, 0);
66 vn3 = _mm512_roundscale_ps(vn3, 0);
67 vn4 = _mm512_roundscale_ps(vn4, 0);
68 vn5 = _mm512_roundscale_ps(vn5, 0);
69 vn6 = _mm512_roundscale_ps(vn6, 0);
70
71 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
72 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
73 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
74 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
75 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
76 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
77 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
78
79 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
80 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
81 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
82 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
83 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
84 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
85 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
86
87 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
88 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
89 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
90 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
91 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
92 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
93 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
94
95 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
96 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
97 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
98 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
99 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
100 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
101 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
102
103 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
104 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
105 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
106 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
107 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
108 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
110
111 vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
112 vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
113 vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
114 vp3 = _mm512_fmadd_ps(vp3, vt3, vone);
115 vp4 = _mm512_fmadd_ps(vp4, vt4, vone);
116 vp5 = _mm512_fmadd_ps(vp5, vt5, vone);
117 vp6 = _mm512_fmadd_ps(vp6, vt6, vone);
118
119 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
120 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
121 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
122 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
123 const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
124 const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
125 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
126
127 const __m512 vd0 = _mm512_add_ps(ve0, vone);
128 const __m512 vd1 = _mm512_add_ps(ve1, vone);
129 const __m512 vd2 = _mm512_add_ps(ve2, vone);
130 const __m512 vd3 = _mm512_add_ps(ve3, vone);
131 const __m512 vd4 = _mm512_add_ps(ve4, vone);
132 const __m512 vd5 = _mm512_add_ps(ve5, vone);
133 const __m512 vd6 = _mm512_add_ps(ve6, vone);
134
135 __m512 vr0 = _mm512_rcp14_ps(vd0);
136 __m512 vr1 = _mm512_rcp14_ps(vd1);
137 __m512 vr2 = _mm512_rcp14_ps(vd2);
138 __m512 vr3 = _mm512_rcp14_ps(vd3);
139 __m512 vr4 = _mm512_rcp14_ps(vd4);
140 __m512 vr5 = _mm512_rcp14_ps(vd5);
141 __m512 vr6 = _mm512_rcp14_ps(vd6);
142
143 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
144 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
145 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
146 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
147 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
148 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
149 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
150
151 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
152 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
153 __m512 vf2 = _mm512_mul_ps(ve2, vr2);
154 __m512 vf3 = _mm512_mul_ps(ve3, vr3);
155 __m512 vf4 = _mm512_mul_ps(ve4, vr4);
156 __m512 vf5 = _mm512_mul_ps(ve5, vr5);
157 __m512 vf6 = _mm512_mul_ps(ve6, vr6);
158
159
160 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
161 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
162 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
163 vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
164 vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
165 vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
166 vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
167
168 _mm512_storeu_ps(y, vf0);
169 _mm512_storeu_ps(y + 16, vf1);
170 _mm512_storeu_ps(y + 32, vf2);
171 _mm512_storeu_ps(y + 48, vf3);
172 _mm512_storeu_ps(y + 64, vf4);
173 _mm512_storeu_ps(y + 80, vf5);
174 _mm512_storeu_ps(y + 96, vf6);
175 y += 112;
176 }
177 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
178 const __m512 vx = _mm512_loadu_ps(x);
179 x += 16;
180
181 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
182
183 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
184
185 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
186
187 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
188 vp = _mm512_fmadd_ps(vp, vt, vc3);
189 vp = _mm512_fmadd_ps(vp, vt, vc2);
190 vp = _mm512_fmadd_ps(vp, vt, vc1);
191 vp = _mm512_fmadd_ps(vp, vt, vone);
192
193 const __m512 ve = _mm512_scalef_ps(vp, vn);
194 const __m512 vd = _mm512_add_ps(ve, vone);
195
196 __m512 vr = _mm512_rcp14_ps(vd);
197 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
198
199 __m512 vf = _mm512_mul_ps(ve, vr);
200
201 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
202
203 _mm512_storeu_ps(y, vf);
204 y += 16;
205 }
206 if XNN_UNLIKELY(n != 0) {
207 assert(n >= 1 * sizeof(float));
208 assert(n <= 15 * sizeof(float));
209
210 // Prepare mask for valid 32-bit elements (depends on n).
211 n >>= 2 /* log2(sizeof(float)) */;
212 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
213
214 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
215 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
216
217 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
218
219 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
220
221 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
222 vp = _mm512_fmadd_ps(vp, vt, vc3);
223 vp = _mm512_fmadd_ps(vp, vt, vc2);
224 vp = _mm512_fmadd_ps(vp, vt, vc1);
225 vp = _mm512_fmadd_ps(vp, vt, vone);
226
227 const __m512 ve = _mm512_scalef_ps(vp, vn);
228 const __m512 vd = _mm512_add_ps(ve, vone);
229
230 __m512 vr = _mm512_rcp14_ps(vd);
231 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
232
233 __m512 vf = _mm512_mul_ps(ve, vr);
234
235 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
236
237 _mm512_mask_storeu_ps(y, vmask, vf);
238 }
239 }
240