1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx512f-rr1-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_vsigmoid_ukernel__avx512f_rr1_lut16_p3_perm_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_lut16_p3_perm_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_lut16_p3.sign_mask);
28   const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias);
29   const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e);
30   const __m512 vtable = _mm512_load_ps(params->avx512_rr1_lut16_p3.table);
31   const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2);
32   const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3);
33   const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2);
34   const __m512 vone = _mm512_set1_ps(params->avx512_rr1_lut16_p3.one);
35 
36   for (; n >= 112 * sizeof(float); n -= 112 * sizeof(float)) {
37     const __m512 vx0 = _mm512_loadu_ps(x);
38     const __m512 vx1 = _mm512_loadu_ps(x + 16);
39     const __m512 vx2 = _mm512_loadu_ps(x + 32);
40     const __m512 vx3 = _mm512_loadu_ps(x + 48);
41     const __m512 vx4 = _mm512_loadu_ps(x + 64);
42     const __m512 vx5 = _mm512_loadu_ps(x + 80);
43     const __m512 vx6 = _mm512_loadu_ps(x + 96);
44     x += 112;
45 
46     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
47     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
48     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
49     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
50     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
51     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
52     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
53 
54     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
55     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
56     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
57     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
58     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
59     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
60     __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias);
61 
62     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
63     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
64     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
65     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
66     const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
67     const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable);
68     const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable);
69 
70     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
71     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
72     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
73     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
74     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
75     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
76     vn6 = _mm512_sub_ps(vn6, vmagic_bias);
77 
78     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
79     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
80     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
81     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
82     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
83     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
84     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
85 
86     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
87     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
88     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
89     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
90     __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
91     __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2);
92     __m512 vp6 = _mm512_fmadd_ps(vt6, vc3, vc2);
93 
94     vp0 = _mm512_mul_ps(vp0, vt0);
95     vp1 = _mm512_mul_ps(vp1, vt1);
96     vp2 = _mm512_mul_ps(vp2, vt2);
97     vp3 = _mm512_mul_ps(vp3, vt3);
98     vp4 = _mm512_mul_ps(vp4, vt4);
99     vp5 = _mm512_mul_ps(vp5, vt5);
100     vp6 = _mm512_mul_ps(vp6, vt6);
101 
102     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
103     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
104     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
105     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
106     vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
107     vp5 = _mm512_fmadd_ps(vt5, vp5, vt5);
108     vp6 = _mm512_fmadd_ps(vt6, vp6, vt6);
109 
110     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
111     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
112     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
113     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
114     vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
115     vp5 = _mm512_fmadd_ps(vl5, vp5, vl5);
116     vp6 = _mm512_fmadd_ps(vl6, vp6, vl6);
117 
118     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
119     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
120     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
121     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
122     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
123     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
124     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
125 
126     const __m512 vd0 = _mm512_add_ps(ve0, vone);
127     const __m512 vd1 = _mm512_add_ps(ve1, vone);
128     const __m512 vd2 = _mm512_add_ps(ve2, vone);
129     const __m512 vd3 = _mm512_add_ps(ve3, vone);
130     const __m512 vd4 = _mm512_add_ps(ve4, vone);
131     const __m512 vd5 = _mm512_add_ps(ve5, vone);
132     const __m512 vd6 = _mm512_add_ps(ve6, vone);
133 
134     __m512 vr0 = _mm512_rcp14_ps(vd0);
135     __m512 vr1 = _mm512_rcp14_ps(vd1);
136     __m512 vr2 = _mm512_rcp14_ps(vd2);
137     __m512 vr3 = _mm512_rcp14_ps(vd3);
138     __m512 vr4 = _mm512_rcp14_ps(vd4);
139     __m512 vr5 = _mm512_rcp14_ps(vd5);
140     __m512 vr6 = _mm512_rcp14_ps(vd6);
141 
142     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
143     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
144     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
145     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
146     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
147     vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
148     vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
149 
150     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
151     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
152     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
153     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
154     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
155     __m512 vf5 = _mm512_mul_ps(ve5, vr5);
156     __m512 vf6 = _mm512_mul_ps(ve6, vr6);
157 
158 
159     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
160     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
161     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
162     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
163     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
164     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
165     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
166 
167     _mm512_storeu_ps(y, vf0);
168     _mm512_storeu_ps(y + 16, vf1);
169     _mm512_storeu_ps(y + 32, vf2);
170     _mm512_storeu_ps(y + 48, vf3);
171     _mm512_storeu_ps(y + 64, vf4);
172     _mm512_storeu_ps(y + 80, vf5);
173     _mm512_storeu_ps(y + 96, vf6);
174     y += 112;
175   }
176   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
177     const __m512 vx = _mm512_loadu_ps(x);
178     x += 16;
179 
180     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
181 
182     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
183     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
184     vn = _mm512_sub_ps(vn, vmagic_bias);
185 
186     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
187 
188     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
189     vp = _mm512_mul_ps(vp, vt);
190     vp = _mm512_fmadd_ps(vt, vp, vt);
191     vp = _mm512_fmadd_ps(vl, vp, vl);
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     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
218     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
219     vn = _mm512_sub_ps(vn, vmagic_bias);
220 
221     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
222 
223     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
224     vp = _mm512_mul_ps(vp, vt);
225     vp = _mm512_fmadd_ps(vt, vp, vt);
226     vp = _mm512_fmadd_ps(vl, vp, vl);
227 
228     const __m512 ve = _mm512_scalef_ps(vp, vn);
229     const __m512 vd = _mm512_add_ps(ve, vone);
230 
231     __m512 vr = _mm512_rcp14_ps(vd);
232     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
233 
234     __m512 vf = _mm512_mul_ps(ve, vr);
235 
236     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
237 
238     _mm512_mask_storeu_ps(y, vmask, vf);
239   }
240 }
241