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_x96(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_lut16_p3_perm_scalef_div_x96(
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 >= 96 * sizeof(float); n -= 96 * 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     x += 96;
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 
56     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
57     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
58     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
59     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
60     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
61     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
62 
63     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
64     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
65     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
66     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
67     const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
68     const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), 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 
77     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
78     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
79     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
80     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
81     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
82     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
83 
84     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
85     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
86     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
87     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
88     __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
89     __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2);
90 
91     vp0 = _mm512_mul_ps(vp0, vt0);
92     vp1 = _mm512_mul_ps(vp1, vt1);
93     vp2 = _mm512_mul_ps(vp2, vt2);
94     vp3 = _mm512_mul_ps(vp3, vt3);
95     vp4 = _mm512_mul_ps(vp4, vt4);
96     vp5 = _mm512_mul_ps(vp5, vt5);
97 
98     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
99     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
100     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
101     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
102     vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
103     vp5 = _mm512_fmadd_ps(vt5, vp5, vt5);
104 
105     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
106     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
107     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
108     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
109     vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
110     vp5 = _mm512_fmadd_ps(vl5, vp5, vl5);
111 
112     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
113     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
114     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
115     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
116     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
117     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
118 
119     const __m512 vd0 = _mm512_add_ps(ve0, vone);
120     const __m512 vd1 = _mm512_add_ps(ve1, vone);
121     const __m512 vd2 = _mm512_add_ps(ve2, vone);
122     const __m512 vd3 = _mm512_add_ps(ve3, vone);
123     const __m512 vd4 = _mm512_add_ps(ve4, vone);
124     const __m512 vd5 = _mm512_add_ps(ve5, vone);
125 
126     __m512 vf0 = _mm512_div_ps(ve0, vd0);
127     __m512 vf1 = _mm512_div_ps(ve1, vd1);
128     __m512 vf2 = _mm512_div_ps(ve2, vd2);
129     __m512 vf3 = _mm512_div_ps(ve3, vd3);
130     __m512 vf4 = _mm512_div_ps(ve4, vd4);
131     __m512 vf5 = _mm512_div_ps(ve5, vd5);
132 
133     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
134     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
135     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
136     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
137     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
138     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
139 
140     _mm512_storeu_ps(y, vf0);
141     _mm512_storeu_ps(y + 16, vf1);
142     _mm512_storeu_ps(y + 32, vf2);
143     _mm512_storeu_ps(y + 48, vf3);
144     _mm512_storeu_ps(y + 64, vf4);
145     _mm512_storeu_ps(y + 80, vf5);
146     y += 96;
147   }
148   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
149     const __m512 vx = _mm512_loadu_ps(x);
150     x += 16;
151 
152     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
153 
154     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
155     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
156     vn = _mm512_sub_ps(vn, vmagic_bias);
157 
158     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
159 
160     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
161     vp = _mm512_mul_ps(vp, vt);
162     vp = _mm512_fmadd_ps(vt, vp, vt);
163     vp = _mm512_fmadd_ps(vl, vp, vl);
164 
165     const __m512 ve = _mm512_scalef_ps(vp, vn);
166     const __m512 vd = _mm512_add_ps(ve, vone);
167 
168     __m512 vf = _mm512_div_ps(ve, vd);
169 
170     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
171 
172     _mm512_storeu_ps(y, vf);
173     y += 16;
174   }
175   if XNN_UNLIKELY(n != 0) {
176     assert(n >= 1 * sizeof(float));
177     assert(n <= 15 * sizeof(float));
178 
179     // Prepare mask for valid 32-bit elements (depends on n).
180     n >>= 2 /* log2(sizeof(float)) */;
181     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
182 
183     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
184     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
185 
186     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
187     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
188     vn = _mm512_sub_ps(vn, vmagic_bias);
189 
190     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
191 
192     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
193     vp = _mm512_mul_ps(vp, vt);
194     vp = _mm512_fmadd_ps(vt, vp, vt);
195     vp = _mm512_fmadd_ps(vl, vp, vl);
196 
197     const __m512 ve = _mm512_scalef_ps(vp, vn);
198     const __m512 vd = _mm512_add_ps(ve, vone);
199 
200     __m512 vf = _mm512_div_ps(ve, vd);
201 
202     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
203 
204     _mm512_mask_storeu_ps(y, vmask, vf);
205   }
206 }
207