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_x128(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_nr1fma_x128(
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 >= 128 * sizeof(float); n -= 128 * 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 const __m512 vx4 = _mm512_loadu_ps(x + 64);
46 const __m512 vx5 = _mm512_loadu_ps(x + 80);
47 const __m512 vx6 = _mm512_loadu_ps(x + 96);
48 const __m512 vx7 = _mm512_loadu_ps(x + 112);
49 x += 128;
50
51 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
52 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
53 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
54 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
55 const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
56 const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
57 const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
58 const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask));
59
60 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
61 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
62 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
63 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
64 __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
65 __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
66 __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias);
67 __m512 vn7 = _mm512_fmadd_ps(vz7, vlog2e, vmagic_bias);
68
69 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
70 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
71 const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
72 const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
73 const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
74 const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable);
75 const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable);
76 const __m512 vl7 = _mm512_permutexvar_ps(_mm512_castps_si512(vn7), vtable);
77
78 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
79 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
80 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
81 vn3 = _mm512_sub_ps(vn3, vmagic_bias);
82 vn4 = _mm512_sub_ps(vn4, vmagic_bias);
83 vn5 = _mm512_sub_ps(vn5, vmagic_bias);
84 vn6 = _mm512_sub_ps(vn6, vmagic_bias);
85 vn7 = _mm512_sub_ps(vn7, vmagic_bias);
86
87 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
88 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
89 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
90 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
91 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
92 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
94 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2, vz7);
95
96 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
97 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
98 __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
99 __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
100 __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
101 __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2);
102 __m512 vp6 = _mm512_fmadd_ps(vt6, vc3, vc2);
103 __m512 vp7 = _mm512_fmadd_ps(vt7, vc3, vc2);
104
105 vp0 = _mm512_mul_ps(vp0, vt0);
106 vp1 = _mm512_mul_ps(vp1, vt1);
107 vp2 = _mm512_mul_ps(vp2, vt2);
108 vp3 = _mm512_mul_ps(vp3, vt3);
109 vp4 = _mm512_mul_ps(vp4, vt4);
110 vp5 = _mm512_mul_ps(vp5, vt5);
111 vp6 = _mm512_mul_ps(vp6, vt6);
112 vp7 = _mm512_mul_ps(vp7, vt7);
113
114 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
115 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
116 vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
117 vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
118 vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
119 vp5 = _mm512_fmadd_ps(vt5, vp5, vt5);
120 vp6 = _mm512_fmadd_ps(vt6, vp6, vt6);
121 vp7 = _mm512_fmadd_ps(vt7, vp7, vt7);
122
123 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
124 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
125 vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
126 vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
127 vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
128 vp5 = _mm512_fmadd_ps(vl5, vp5, vl5);
129 vp6 = _mm512_fmadd_ps(vl6, vp6, vl6);
130 vp7 = _mm512_fmadd_ps(vl7, vp7, vl7);
131
132 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
133 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
134 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
135 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
136 const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
137 const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
138 const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
139 const __m512 ve7 = _mm512_scalef_ps(vp7, vn7);
140
141 const __m512 vd0 = _mm512_add_ps(ve0, vone);
142 const __m512 vd1 = _mm512_add_ps(ve1, vone);
143 const __m512 vd2 = _mm512_add_ps(ve2, vone);
144 const __m512 vd3 = _mm512_add_ps(ve3, vone);
145 const __m512 vd4 = _mm512_add_ps(ve4, vone);
146 const __m512 vd5 = _mm512_add_ps(ve5, vone);
147 const __m512 vd6 = _mm512_add_ps(ve6, vone);
148 const __m512 vd7 = _mm512_add_ps(ve7, vone);
149
150 __m512 vr0 = _mm512_rcp14_ps(vd0);
151 __m512 vr1 = _mm512_rcp14_ps(vd1);
152 __m512 vr2 = _mm512_rcp14_ps(vd2);
153 __m512 vr3 = _mm512_rcp14_ps(vd3);
154 __m512 vr4 = _mm512_rcp14_ps(vd4);
155 __m512 vr5 = _mm512_rcp14_ps(vd5);
156 __m512 vr6 = _mm512_rcp14_ps(vd6);
157 __m512 vr7 = _mm512_rcp14_ps(vd7);
158
159 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
160 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
161 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
162 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
163 vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
164 vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
165 vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
166 vr7 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr7, vd7, vone), vr7, vr7);
167
168 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
169 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
170 __m512 vf2 = _mm512_mul_ps(ve2, vr2);
171 __m512 vf3 = _mm512_mul_ps(ve3, vr3);
172 __m512 vf4 = _mm512_mul_ps(ve4, vr4);
173 __m512 vf5 = _mm512_mul_ps(ve5, vr5);
174 __m512 vf6 = _mm512_mul_ps(ve6, vr6);
175 __m512 vf7 = _mm512_mul_ps(ve7, vr7);
176
177
178 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
179 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
180 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
181 vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
182 vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
183 vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
184 vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
185 vf7 = _mm512_mask_sub_ps(vf7, _mm512_testn_epi32_mask(_mm512_castps_si512(vx7), vsign_mask), vone, vf7);
186
187 _mm512_storeu_ps(y, vf0);
188 _mm512_storeu_ps(y + 16, vf1);
189 _mm512_storeu_ps(y + 32, vf2);
190 _mm512_storeu_ps(y + 48, vf3);
191 _mm512_storeu_ps(y + 64, vf4);
192 _mm512_storeu_ps(y + 80, vf5);
193 _mm512_storeu_ps(y + 96, vf6);
194 _mm512_storeu_ps(y + 112, vf7);
195 y += 128;
196 }
197 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
198 const __m512 vx = _mm512_loadu_ps(x);
199 x += 16;
200
201 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
202
203 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
204 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
205 vn = _mm512_sub_ps(vn, vmagic_bias);
206
207 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
208
209 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
210 vp = _mm512_mul_ps(vp, vt);
211 vp = _mm512_fmadd_ps(vt, vp, vt);
212 vp = _mm512_fmadd_ps(vl, vp, vl);
213
214 const __m512 ve = _mm512_scalef_ps(vp, vn);
215 const __m512 vd = _mm512_add_ps(ve, vone);
216
217 __m512 vr = _mm512_rcp14_ps(vd);
218 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
219
220 __m512 vf = _mm512_mul_ps(ve, vr);
221
222 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
223
224 _mm512_storeu_ps(y, vf);
225 y += 16;
226 }
227 if XNN_UNLIKELY(n != 0) {
228 assert(n >= 1 * sizeof(float));
229 assert(n <= 15 * sizeof(float));
230
231 // Prepare mask for valid 32-bit elements (depends on n).
232 n >>= 2 /* log2(sizeof(float)) */;
233 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
234
235 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
236 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
237
238 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
239 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
240 vn = _mm512_sub_ps(vn, vmagic_bias);
241
242 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
243
244 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
245 vp = _mm512_mul_ps(vp, vt);
246 vp = _mm512_fmadd_ps(vt, vp, vt);
247 vp = _mm512_fmadd_ps(vl, vp, vl);
248
249 const __m512 ve = _mm512_scalef_ps(vp, vn);
250 const __m512 vd = _mm512_add_ps(ve, vone);
251
252 __m512 vr = _mm512_rcp14_ps(vd);
253 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
254
255 __m512 vf = _mm512_mul_ps(ve, vr);
256
257 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
258
259 _mm512_mask_storeu_ps(y, vmask, vf);
260 }
261 }
262