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_x96(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_x96(
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 >= 96 * sizeof(float); n -= 96 * 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     x += 96;
44 
45     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
46     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
47     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
48     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
49     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
50     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
51 
52     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
53     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
54     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
55     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
56     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
57     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
58 
59     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
60     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
61     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
62     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
63     const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
64     const __m512 vl5 = _mm512_permutexvar_ps(_mm512_castps_si512(vn5), vtable);
65 
66     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
67     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
68     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
69     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
70     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
71     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
72 
73     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
74     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
75     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
76     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
77     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
78     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
79 
80     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
81     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
82     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
83     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
84     __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
85     __m512 vp5 = _mm512_fmadd_ps(vt5, vc3, vc2);
86 
87     vp0 = _mm512_mul_ps(vp0, vt0);
88     vp1 = _mm512_mul_ps(vp1, vt1);
89     vp2 = _mm512_mul_ps(vp2, vt2);
90     vp3 = _mm512_mul_ps(vp3, vt3);
91     vp4 = _mm512_mul_ps(vp4, vt4);
92     vp5 = _mm512_mul_ps(vp5, vt5);
93 
94     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
95     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
96     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
97     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
98     vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
99     vp5 = _mm512_fmadd_ps(vt5, vp5, vt5);
100 
101     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
102     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
103     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
104     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
105     vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
106     vp5 = _mm512_fmadd_ps(vl5, vp5, vl5);
107 
108     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
109     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
110     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
111     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
112     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
113     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
114 
115     const __m512 vd0 = _mm512_add_ps(ve0, vone);
116     const __m512 vd1 = _mm512_add_ps(ve1, vone);
117     const __m512 vd2 = _mm512_add_ps(ve2, vone);
118     const __m512 vd3 = _mm512_add_ps(ve3, vone);
119     const __m512 vd4 = _mm512_add_ps(ve4, vone);
120     const __m512 vd5 = _mm512_add_ps(ve5, vone);
121 
122     __m512 vr0 = _mm512_rcp14_ps(vd0);
123     __m512 vr1 = _mm512_rcp14_ps(vd1);
124     __m512 vr2 = _mm512_rcp14_ps(vd2);
125     __m512 vr3 = _mm512_rcp14_ps(vd3);
126     __m512 vr4 = _mm512_rcp14_ps(vd4);
127     __m512 vr5 = _mm512_rcp14_ps(vd5);
128 
129     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
130     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
131     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
132     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
133     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
134     vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
135 
136     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
137     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
138     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
139     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
140     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
141     __m512 vf5 = _mm512_mul_ps(ve5, vr5);
142 
143 
144     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
145     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
146     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
147     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
148     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
149     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
150 
151     _mm512_storeu_ps(y, vf0);
152     _mm512_storeu_ps(y + 16, vf1);
153     _mm512_storeu_ps(y + 32, vf2);
154     _mm512_storeu_ps(y + 48, vf3);
155     _mm512_storeu_ps(y + 64, vf4);
156     _mm512_storeu_ps(y + 80, vf5);
157     y += 96;
158   }
159   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
160     const __m512 vx = _mm512_loadu_ps(x);
161     x += 16;
162 
163     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
164 
165     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
166     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
167     vn = _mm512_sub_ps(vn, vmagic_bias);
168 
169     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
170 
171     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
172     vp = _mm512_mul_ps(vp, vt);
173     vp = _mm512_fmadd_ps(vt, vp, vt);
174     vp = _mm512_fmadd_ps(vl, vp, vl);
175 
176     const __m512 ve = _mm512_scalef_ps(vp, vn);
177     const __m512 vd = _mm512_add_ps(ve, vone);
178 
179     __m512 vr = _mm512_rcp14_ps(vd);
180     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
181 
182     __m512 vf = _mm512_mul_ps(ve, vr);
183 
184     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
185 
186     _mm512_storeu_ps(y, vf);
187     y += 16;
188   }
189   if XNN_UNLIKELY(n != 0) {
190     assert(n >= 1 * sizeof(float));
191     assert(n <= 15 * sizeof(float));
192 
193     // Prepare mask for valid 32-bit elements (depends on n).
194     n >>= 2 /* log2(sizeof(float)) */;
195     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
196 
197     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
198     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
199 
200     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
201     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
202     vn = _mm512_sub_ps(vn, vmagic_bias);
203 
204     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
205 
206     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
207     vp = _mm512_mul_ps(vp, vt);
208     vp = _mm512_fmadd_ps(vt, vp, vt);
209     vp = _mm512_fmadd_ps(vl, vp, vl);
210 
211     const __m512 ve = _mm512_scalef_ps(vp, vn);
212     const __m512 vd = _mm512_add_ps(ve, vone);
213 
214     __m512 vr = _mm512_rcp14_ps(vd);
215     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
216 
217     __m512 vf = _mm512_mul_ps(ve, vr);
218 
219     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
220 
221     _mm512_mask_storeu_ps(y, vmask, vf);
222   }
223 }
224