1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx512f-rr2-lut32-p2-perm2-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_rr2_lut32_p2_perm2_scalef_div_x128(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_rr2_lut32_p2_perm2_scalef_div_x128(
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_rr2_lut32_p2.sign_mask);
28   const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr2_lut32_p2.magic_bias);
29   const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr2_lut32_p2.log2e);
30   const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo);
31   const __m512 vtable_hi = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_hi);
32   const __m512 vminus_ln2_hi = _mm512_set1_ps(params->avx512_rr2_lut32_p2.minus_ln2_hi);
33   const __m512 vminus_ln2_lo = _mm512_set1_ps(params->avx512_rr2_lut32_p2.minus_ln2_lo);
34   const __m512 vc2 = _mm512_set1_ps(params->avx512_rr2_lut32_p2.c2);
35   const __m512 vc1 = _mm512_set1_ps(params->avx512_rr2_lut32_p2.c1);
36   const __m512 vone = _mm512_set1_ps(params->avx512_rr2_lut32_p2.one);
37 
38   for (; n >= 128 * sizeof(float); n -= 128 * sizeof(float)) {
39     const __m512 vx0 = _mm512_loadu_ps(x);
40     const __m512 vx1 = _mm512_loadu_ps(x + 16);
41     const __m512 vx2 = _mm512_loadu_ps(x + 32);
42     const __m512 vx3 = _mm512_loadu_ps(x + 48);
43     const __m512 vx4 = _mm512_loadu_ps(x + 64);
44     const __m512 vx5 = _mm512_loadu_ps(x + 80);
45     const __m512 vx6 = _mm512_loadu_ps(x + 96);
46     const __m512 vx7 = _mm512_loadu_ps(x + 112);
47     x += 128;
48 
49     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
50     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
51     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
52     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
53     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
54     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
55     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
56     const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask));
57 
58     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
59     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
60     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
61     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
62     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
63     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
64     __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias);
65     __m512 vn7 = _mm512_fmadd_ps(vz7, vlog2e, vmagic_bias);
66 
67     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
68     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
69     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
70     const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
71     const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi);
72     const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi);
73     const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi);
74     const __m512 vl7 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn7), vtable_hi);
75 
76     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
77     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
78     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
79     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
80     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
81     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
82     vn6 = _mm512_sub_ps(vn6, vmagic_bias);
83     vn7 = _mm512_sub_ps(vn7, vmagic_bias);
84 
85     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
86     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
87     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
88     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
89     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vz4);
90     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vz5);
91     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vz6);
92     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vz7);
93 
94     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
95     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
96     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
97     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
98     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
99     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
100     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
101     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
102 
103     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
104     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
105     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
106     __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
107     __m512 vp4 = _mm512_fmadd_ps(vt4, vc2, vc1);
108     __m512 vp5 = _mm512_fmadd_ps(vt5, vc2, vc1);
109     __m512 vp6 = _mm512_fmadd_ps(vt6, vc2, vc1);
110     __m512 vp7 = _mm512_fmadd_ps(vt7, vc2, vc1);
111 
112     vt0 = _mm512_mul_ps(vt0, vl0);
113     vt1 = _mm512_mul_ps(vt1, vl1);
114     vt2 = _mm512_mul_ps(vt2, vl2);
115     vt3 = _mm512_mul_ps(vt3, vl3);
116     vt4 = _mm512_mul_ps(vt4, vl4);
117     vt5 = _mm512_mul_ps(vt5, vl5);
118     vt6 = _mm512_mul_ps(vt6, vl6);
119     vt7 = _mm512_mul_ps(vt7, vl7);
120 
121     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
122     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
123     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
124     vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
125     vp4 = _mm512_fmadd_ps(vt4, vp4, vl4);
126     vp5 = _mm512_fmadd_ps(vt5, vp5, vl5);
127     vp6 = _mm512_fmadd_ps(vt6, vp6, vl6);
128     vp7 = _mm512_fmadd_ps(vt7, vp7, vl7);
129 
130     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
131     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
132     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
133     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
134     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
135     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
136     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
137     const __m512 ve7 = _mm512_scalef_ps(vp7, vn7);
138 
139     const __m512 vd0 = _mm512_add_ps(ve0, vone);
140     const __m512 vd1 = _mm512_add_ps(ve1, vone);
141     const __m512 vd2 = _mm512_add_ps(ve2, vone);
142     const __m512 vd3 = _mm512_add_ps(ve3, vone);
143     const __m512 vd4 = _mm512_add_ps(ve4, vone);
144     const __m512 vd5 = _mm512_add_ps(ve5, vone);
145     const __m512 vd6 = _mm512_add_ps(ve6, vone);
146     const __m512 vd7 = _mm512_add_ps(ve7, vone);
147 
148     __m512 vf0 = _mm512_div_ps(ve0, vd0);
149     __m512 vf1 = _mm512_div_ps(ve1, vd1);
150     __m512 vf2 = _mm512_div_ps(ve2, vd2);
151     __m512 vf3 = _mm512_div_ps(ve3, vd3);
152     __m512 vf4 = _mm512_div_ps(ve4, vd4);
153     __m512 vf5 = _mm512_div_ps(ve5, vd5);
154     __m512 vf6 = _mm512_div_ps(ve6, vd6);
155     __m512 vf7 = _mm512_div_ps(ve7, vd7);
156 
157     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
158     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
159     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
160     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
161     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
162     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
163     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
164     vf7 = _mm512_mask_sub_ps(vf7, _mm512_testn_epi32_mask(_mm512_castps_si512(vx7), vsign_mask), vone, vf7);
165 
166     _mm512_storeu_ps(y, vf0);
167     _mm512_storeu_ps(y + 16, vf1);
168     _mm512_storeu_ps(y + 32, vf2);
169     _mm512_storeu_ps(y + 48, vf3);
170     _mm512_storeu_ps(y + 64, vf4);
171     _mm512_storeu_ps(y + 80, vf5);
172     _mm512_storeu_ps(y + 96, vf6);
173     _mm512_storeu_ps(y + 112, vf7);
174     y += 128;
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_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
184     vn = _mm512_sub_ps(vn, vmagic_bias);
185 
186     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
187     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
188 
189     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
190     vt = _mm512_mul_ps(vt, vl);
191     vp = _mm512_fmadd_ps(vt, vp, vl);
192 
193     const __m512 ve = _mm512_scalef_ps(vp, vn);
194     const __m512 vd = _mm512_add_ps(ve, vone);
195 
196     __m512 vf = _mm512_div_ps(ve, vd);
197 
198     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
199 
200     _mm512_storeu_ps(y, vf);
201     y += 16;
202   }
203   if XNN_UNLIKELY(n != 0) {
204     assert(n >= 1 * sizeof(float));
205     assert(n <= 15 * sizeof(float));
206 
207     // Prepare mask for valid 32-bit elements (depends on n).
208     n >>= 2 /* log2(sizeof(float)) */;
209     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
210 
211     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
212     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
213 
214     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
215     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
216     vn = _mm512_sub_ps(vn, vmagic_bias);
217 
218     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
219     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
220 
221     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
222     vt = _mm512_mul_ps(vt, vl);
223     vp = _mm512_fmadd_ps(vt, vp, vl);
224 
225     const __m512 ve = _mm512_scalef_ps(vp, vn);
226     const __m512 vd = _mm512_add_ps(ve, vone);
227 
228     __m512 vf = _mm512_div_ps(ve, vd);
229 
230     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
231 
232     _mm512_mask_storeu_ps(y, vmask, vf);
233   }
234 }
235