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_nr1fma_x112(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_nr1fma_x112(
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 >= 112 * sizeof(float); n -= 112 * 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     x += 112;
47 
48     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
49     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
50     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
51     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
52     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
53     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
54     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), 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     __m512 vn6 = _mm512_fmadd_ps(vz6, vlog2e, vmagic_bias);
63 
64     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
65     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
66     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
67     const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
68     const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi);
69     const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi);
70     const __m512 vl6 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn6), vtable_hi);
71 
72     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
73     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
74     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
75     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
76     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
77     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
78     vn6 = _mm512_sub_ps(vn6, vmagic_bias);
79 
80     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
81     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
82     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
83     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
84     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vz4);
85     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vz5);
86     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vz6);
87 
88     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
89     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
90     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
91     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
92     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
93     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
94     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
95 
96     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
97     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
98     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
99     __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
100     __m512 vp4 = _mm512_fmadd_ps(vt4, vc2, vc1);
101     __m512 vp5 = _mm512_fmadd_ps(vt5, vc2, vc1);
102     __m512 vp6 = _mm512_fmadd_ps(vt6, 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     vt6 = _mm512_mul_ps(vt6, vl6);
111 
112     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
113     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
114     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
115     vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
116     vp4 = _mm512_fmadd_ps(vt4, vp4, vl4);
117     vp5 = _mm512_fmadd_ps(vt5, vp5, vl5);
118     vp6 = _mm512_fmadd_ps(vt6, vp6, vl6);
119 
120     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
121     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
122     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
123     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
124     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
125     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
126     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
127 
128     const __m512 vd0 = _mm512_add_ps(ve0, vone);
129     const __m512 vd1 = _mm512_add_ps(ve1, vone);
130     const __m512 vd2 = _mm512_add_ps(ve2, vone);
131     const __m512 vd3 = _mm512_add_ps(ve3, vone);
132     const __m512 vd4 = _mm512_add_ps(ve4, vone);
133     const __m512 vd5 = _mm512_add_ps(ve5, vone);
134     const __m512 vd6 = _mm512_add_ps(ve6, vone);
135 
136     __m512 vr0 = _mm512_rcp14_ps(vd0);
137     __m512 vr1 = _mm512_rcp14_ps(vd1);
138     __m512 vr2 = _mm512_rcp14_ps(vd2);
139     __m512 vr3 = _mm512_rcp14_ps(vd3);
140     __m512 vr4 = _mm512_rcp14_ps(vd4);
141     __m512 vr5 = _mm512_rcp14_ps(vd5);
142     __m512 vr6 = _mm512_rcp14_ps(vd6);
143 
144     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
145     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
146     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
147     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
148     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
149     vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
150     vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
151 
152     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
153     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
154     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
155     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
156     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
157     __m512 vf5 = _mm512_mul_ps(ve5, vr5);
158     __m512 vf6 = _mm512_mul_ps(ve6, vr6);
159 
160 
161     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
162     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
163     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
164     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
165     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
166     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
167     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
168 
169     _mm512_storeu_ps(y, vf0);
170     _mm512_storeu_ps(y + 16, vf1);
171     _mm512_storeu_ps(y + 32, vf2);
172     _mm512_storeu_ps(y + 48, vf3);
173     _mm512_storeu_ps(y + 64, vf4);
174     _mm512_storeu_ps(y + 80, vf5);
175     _mm512_storeu_ps(y + 96, vf6);
176     y += 112;
177   }
178   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
179     const __m512 vx = _mm512_loadu_ps(x);
180     x += 16;
181 
182     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
183 
184     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
185     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
186     vn = _mm512_sub_ps(vn, vmagic_bias);
187 
188     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
189     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
190 
191     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
192     vt = _mm512_mul_ps(vt, vl);
193     vp = _mm512_fmadd_ps(vt, vp, vl);
194 
195     const __m512 ve = _mm512_scalef_ps(vp, vn);
196     const __m512 vd = _mm512_add_ps(ve, vone);
197 
198     __m512 vr = _mm512_rcp14_ps(vd);
199     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
200 
201     __m512 vf = _mm512_mul_ps(ve, vr);
202 
203     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
204 
205     _mm512_storeu_ps(y, vf);
206     y += 16;
207   }
208   if XNN_UNLIKELY(n != 0) {
209     assert(n >= 1 * sizeof(float));
210     assert(n <= 15 * sizeof(float));
211 
212     // Prepare mask for valid 32-bit elements (depends on n).
213     n >>= 2 /* log2(sizeof(float)) */;
214     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
215 
216     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
217     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
218 
219     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
220     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
221     vn = _mm512_sub_ps(vn, vmagic_bias);
222 
223     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
224     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
225 
226     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
227     vt = _mm512_mul_ps(vt, vl);
228     vp = _mm512_fmadd_ps(vt, vp, vl);
229 
230     const __m512 ve = _mm512_scalef_ps(vp, vn);
231     const __m512 vd = _mm512_add_ps(ve, vone);
232 
233     __m512 vr = _mm512_rcp14_ps(vd);
234     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
235 
236     __m512 vf = _mm512_mul_ps(ve, vr);
237 
238     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
239 
240     _mm512_mask_storeu_ps(y, vmask, vf);
241   }
242 }
243