1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/avx512f-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_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x96(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_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(0x1.800000p18f);
29   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p0f);
30   const __m512 vtable_hi = _mm512_set_ps(
31     0x1.F50766p+0f, 0x1.EA4AFAp+0f, 0x1.DFC974p+0f, 0x1.D5818Ep+0f,
32     0x1.CB720Ep+0f, 0x1.C199BEp+0f, 0x1.B7F770p+0f, 0x1.AE89FAp+0f,
33     0x1.A5503Cp+0f, 0x1.9C4918p+0f, 0x1.93737Cp+0f, 0x1.8ACE54p+0f,
34     0x1.82589Ap+0f, 0x1.7A1148p+0f, 0x1.71F75Ep+0f, 0x1.6A09E6p+0f);
35   const __m512 vtable_lo = _mm512_set_ps(
36     0x1.6247ECp+0f, 0x1.5AB07Ep+0f, 0x1.5342B6p+0f, 0x1.4BFDAEp+0f,
37     0x1.44E086p+0f, 0x1.3DEA64p+0f, 0x1.371A74p+0f, 0x1.306FE0p+0f,
38     0x1.29E9E0p+0f, 0x1.2387A6p+0f, 0x1.1D4874p+0f, 0x1.172B84p+0f,
39     0x1.11301Ep+0f, 0x1.0B5586p+0f, 0x1.059B0Ep+0f, 0x1.000000p+0f);
40   const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
41   const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
42   const __m512 vc2 = _mm512_set1_ps(0x1.000000p-1f);
43   const __m512 vc1 = _mm512_set1_ps(0x1.0000F6p-0f);
44   const __m512 vone = _mm512_set1_ps(1.0f);
45 
46   for (; n >= 96 * sizeof(float); n -= 96 * sizeof(float)) {
47     const __m512 vx0 = _mm512_loadu_ps(x);
48     const __m512 vx1 = _mm512_loadu_ps(x + 16);
49     const __m512 vx2 = _mm512_loadu_ps(x + 32);
50     const __m512 vx3 = _mm512_loadu_ps(x + 48);
51     const __m512 vx4 = _mm512_loadu_ps(x + 64);
52     const __m512 vx5 = _mm512_loadu_ps(x + 80);
53     x += 96;
54 
55     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
56     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
57     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
58     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
59     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
60     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
61 
62     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
63     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
64     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
65     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
66     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
67     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
68 
69     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
70     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
71     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
72     const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
73     const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi);
74     const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), 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 
83     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
84     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
85     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
86     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
87     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vz4);
88     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vz5);
89 
90     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
91     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
92     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
93     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
94     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
95     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
96 
97     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
98     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
99     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
100     __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
101     __m512 vp4 = _mm512_fmadd_ps(vt4, vc2, vc1);
102     __m512 vp5 = _mm512_fmadd_ps(vt5, vc2, vc1);
103 
104     vt0 = _mm512_mul_ps(vt0, vl0);
105     vt1 = _mm512_mul_ps(vt1, vl1);
106     vt2 = _mm512_mul_ps(vt2, vl2);
107     vt3 = _mm512_mul_ps(vt3, vl3);
108     vt4 = _mm512_mul_ps(vt4, vl4);
109     vt5 = _mm512_mul_ps(vt5, vl5);
110 
111     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
112     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
113     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
114     vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
115     vp4 = _mm512_fmadd_ps(vt4, vp4, vl4);
116     vp5 = _mm512_fmadd_ps(vt5, vp5, vl5);
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 
125     const __m512 vd0 = _mm512_add_ps(ve0, vone);
126     const __m512 vd1 = _mm512_add_ps(ve1, vone);
127     const __m512 vd2 = _mm512_add_ps(ve2, vone);
128     const __m512 vd3 = _mm512_add_ps(ve3, vone);
129     const __m512 vd4 = _mm512_add_ps(ve4, vone);
130     const __m512 vd5 = _mm512_add_ps(ve5, vone);
131 
132     __m512 vf0 = _mm512_div_ps(ve0, vd0);
133     __m512 vf1 = _mm512_div_ps(ve1, vd1);
134     __m512 vf2 = _mm512_div_ps(ve2, vd2);
135     __m512 vf3 = _mm512_div_ps(ve3, vd3);
136     __m512 vf4 = _mm512_div_ps(ve4, vd4);
137     __m512 vf5 = _mm512_div_ps(ve5, vd5);
138 
139     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
140     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
141     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
142     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
143     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
144     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
145 
146     _mm512_storeu_ps(y, vf0);
147     _mm512_storeu_ps(y + 16, vf1);
148     _mm512_storeu_ps(y + 32, vf2);
149     _mm512_storeu_ps(y + 48, vf3);
150     _mm512_storeu_ps(y + 64, vf4);
151     _mm512_storeu_ps(y + 80, vf5);
152     y += 96;
153   }
154   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
155     const __m512 vx = _mm512_loadu_ps(x);
156     x += 16;
157 
158     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
159 
160     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
161     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
162     vn = _mm512_sub_ps(vn, vmagic_bias);
163 
164     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
165     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
166 
167     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
168     vt = _mm512_mul_ps(vt, vl);
169     vp = _mm512_fmadd_ps(vt, vp, vl);
170 
171     const __m512 ve = _mm512_scalef_ps(vp, vn);
172     const __m512 vd = _mm512_add_ps(ve, vone);
173 
174     __m512 vf = _mm512_div_ps(ve, vd);
175 
176     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
177 
178     _mm512_storeu_ps(y, vf);
179     y += 16;
180   }
181   if XNN_UNLIKELY(n != 0) {
182     assert(n >= 1 * sizeof(float));
183     assert(n <= 15 * sizeof(float));
184 
185     // Prepare mask for valid 32-bit elements (depends on n).
186     n >>= 2 /* log2(sizeof(float)) */;
187     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
188 
189     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
190     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
191 
192     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
193     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
194     vn = _mm512_sub_ps(vn, vmagic_bias);
195 
196     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
197     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
198 
199     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
200     vt = _mm512_mul_ps(vt, vl);
201     vp = _mm512_fmadd_ps(vt, vp, vl);
202 
203     const __m512 ve = _mm512_scalef_ps(vp, vn);
204     const __m512 vd = _mm512_add_ps(ve, vone);
205 
206     __m512 vf = _mm512_div_ps(ve, vd);
207 
208     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
209 
210     _mm512_mask_storeu_ps(y, vmask, vf);
211   }
212 }
213