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_div_x112(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x112(
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 >= 112 * sizeof(float); n -= 112 * 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     x += 112;
49 
50     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
51     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
52     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
53     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
54     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
55     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
56     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), 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 
66     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
67     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
68     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
69     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
70     const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
71     const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable);
72     const __m512 vl6 = _mm512_permutexvar_ps(_mm512_castps_si512(vn6), vtable);
73 
74     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
75     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
76     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
77     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
78     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
79     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
80     vn6 = _mm512_sub_ps(vn6, vmagic_bias);
81 
82     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
83     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
84     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
85     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
86     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
87     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
88     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
89 
90     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
91     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
92     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
93     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
94     __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
95     __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2);
96     __m512 vp6 = _mm512_fmadd_ps(vt6, vc3, vc2);
97 
98     vp0 = _mm512_mul_ps(vp0, vt0);
99     vp1 = _mm512_mul_ps(vp1, vt1);
100     vp2 = _mm512_mul_ps(vp2, vt2);
101     vp3 = _mm512_mul_ps(vp3, vt3);
102     vp4 = _mm512_mul_ps(vp4, vt4);
103     vp5 = _mm512_mul_ps(vp5, vt5);
104     vp6 = _mm512_mul_ps(vp6, vt6);
105 
106     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
107     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
108     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
109     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
110     vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
111     vp5 = _mm512_fmadd_ps(vt5, vp5, vt5);
112     vp6 = _mm512_fmadd_ps(vt6, vp6, vt6);
113 
114     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
115     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
116     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
117     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
118     vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
119     vp5 = _mm512_fmadd_ps(vl5, vp5, vl5);
120     vp6 = _mm512_fmadd_ps(vl6, vp6, vl6);
121 
122     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
123     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
124     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
125     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
126     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
127     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
128     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
129 
130     const __m512 vd0 = _mm512_add_ps(ve0, vone);
131     const __m512 vd1 = _mm512_add_ps(ve1, vone);
132     const __m512 vd2 = _mm512_add_ps(ve2, vone);
133     const __m512 vd3 = _mm512_add_ps(ve3, vone);
134     const __m512 vd4 = _mm512_add_ps(ve4, vone);
135     const __m512 vd5 = _mm512_add_ps(ve5, vone);
136     const __m512 vd6 = _mm512_add_ps(ve6, vone);
137 
138     __m512 vf0 = _mm512_div_ps(ve0, vd0);
139     __m512 vf1 = _mm512_div_ps(ve1, vd1);
140     __m512 vf2 = _mm512_div_ps(ve2, vd2);
141     __m512 vf3 = _mm512_div_ps(ve3, vd3);
142     __m512 vf4 = _mm512_div_ps(ve4, vd4);
143     __m512 vf5 = _mm512_div_ps(ve5, vd5);
144     __m512 vf6 = _mm512_div_ps(ve6, vd6);
145 
146     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
147     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
148     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
149     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
150     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
151     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
152     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
153 
154     _mm512_storeu_ps(y, vf0);
155     _mm512_storeu_ps(y + 16, vf1);
156     _mm512_storeu_ps(y + 32, vf2);
157     _mm512_storeu_ps(y + 48, vf3);
158     _mm512_storeu_ps(y + 64, vf4);
159     _mm512_storeu_ps(y + 80, vf5);
160     _mm512_storeu_ps(y + 96, vf6);
161     y += 112;
162   }
163   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
164     const __m512 vx = _mm512_loadu_ps(x);
165     x += 16;
166 
167     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
168 
169     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
170     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
171     vn = _mm512_sub_ps(vn, vmagic_bias);
172 
173     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
174 
175     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
176     vp = _mm512_mul_ps(vp, vt);
177     vp = _mm512_fmadd_ps(vt, vp, vt);
178     vp = _mm512_fmadd_ps(vl, vp, vl);
179 
180     const __m512 ve = _mm512_scalef_ps(vp, vn);
181     const __m512 vd = _mm512_add_ps(ve, vone);
182 
183     __m512 vf = _mm512_div_ps(ve, vd);
184 
185     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
186 
187     _mm512_storeu_ps(y, vf);
188     y += 16;
189   }
190   if XNN_UNLIKELY(n != 0) {
191     assert(n >= 1 * sizeof(float));
192     assert(n <= 15 * sizeof(float));
193 
194     // Prepare mask for valid 32-bit elements (depends on n).
195     n >>= 2 /* log2(sizeof(float)) */;
196     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
197 
198     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
199     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
200 
201     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
202     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
203     vn = _mm512_sub_ps(vn, vmagic_bias);
204 
205     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
206 
207     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
208     vp = _mm512_mul_ps(vp, vt);
209     vp = _mm512_fmadd_ps(vt, vp, vt);
210     vp = _mm512_fmadd_ps(vl, vp, vl);
211 
212     const __m512 ve = _mm512_scalef_ps(vp, vn);
213     const __m512 vd = _mm512_add_ps(ve, vone);
214 
215     __m512 vf = _mm512_div_ps(ve, vd);
216 
217     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
218 
219     _mm512_mask_storeu_ps(y, vmask, vf);
220   }
221 }
222