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_div_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_rr2_lut32_p2_perm2_scalef_div_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_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 >= 96 * sizeof(float); n -= 96 * 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     x += 96;
46 
47     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
48     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
49     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
50     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
51     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
52     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
53 
54     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
55     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
56     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
57     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
58     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
59     __m512 vn5 = _mm512_fmadd_ps(vz5, vlog2e, vmagic_bias);
60 
61     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
62     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
63     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
64     const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
65     const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi);
66     const __m512 vl5 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn5), vtable_hi);
67 
68     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
69     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
70     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
71     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
72     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
73     vn5 = _mm512_sub_ps(vn5, vmagic_bias);
74 
75     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
76     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
77     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
78     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
79     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vz4);
80     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vz5);
81 
82     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
83     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
84     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
85     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
86     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
87     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
88 
89     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
90     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
91     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
92     __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
93     __m512 vp4 = _mm512_fmadd_ps(vt4, vc2, vc1);
94     __m512 vp5 = _mm512_fmadd_ps(vt5, vc2, vc1);
95 
96     vt0 = _mm512_mul_ps(vt0, vl0);
97     vt1 = _mm512_mul_ps(vt1, vl1);
98     vt2 = _mm512_mul_ps(vt2, vl2);
99     vt3 = _mm512_mul_ps(vt3, vl3);
100     vt4 = _mm512_mul_ps(vt4, vl4);
101     vt5 = _mm512_mul_ps(vt5, vl5);
102 
103     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
104     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
105     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
106     vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
107     vp4 = _mm512_fmadd_ps(vt4, vp4, vl4);
108     vp5 = _mm512_fmadd_ps(vt5, vp5, vl5);
109 
110     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
111     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
112     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
113     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
114     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
115     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
116 
117     const __m512 vd0 = _mm512_add_ps(ve0, vone);
118     const __m512 vd1 = _mm512_add_ps(ve1, vone);
119     const __m512 vd2 = _mm512_add_ps(ve2, vone);
120     const __m512 vd3 = _mm512_add_ps(ve3, vone);
121     const __m512 vd4 = _mm512_add_ps(ve4, vone);
122     const __m512 vd5 = _mm512_add_ps(ve5, vone);
123 
124     __m512 vf0 = _mm512_div_ps(ve0, vd0);
125     __m512 vf1 = _mm512_div_ps(ve1, vd1);
126     __m512 vf2 = _mm512_div_ps(ve2, vd2);
127     __m512 vf3 = _mm512_div_ps(ve3, vd3);
128     __m512 vf4 = _mm512_div_ps(ve4, vd4);
129     __m512 vf5 = _mm512_div_ps(ve5, vd5);
130 
131     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
132     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
133     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
134     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
135     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
136     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
137 
138     _mm512_storeu_ps(y, vf0);
139     _mm512_storeu_ps(y + 16, vf1);
140     _mm512_storeu_ps(y + 32, vf2);
141     _mm512_storeu_ps(y + 48, vf3);
142     _mm512_storeu_ps(y + 64, vf4);
143     _mm512_storeu_ps(y + 80, vf5);
144     y += 96;
145   }
146   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
147     const __m512 vx = _mm512_loadu_ps(x);
148     x += 16;
149 
150     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
151 
152     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
153     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
154     vn = _mm512_sub_ps(vn, vmagic_bias);
155 
156     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
157     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
158 
159     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
160     vt = _mm512_mul_ps(vt, vl);
161     vp = _mm512_fmadd_ps(vt, vp, vl);
162 
163     const __m512 ve = _mm512_scalef_ps(vp, vn);
164     const __m512 vd = _mm512_add_ps(ve, vone);
165 
166     __m512 vf = _mm512_div_ps(ve, vd);
167 
168     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
169 
170     _mm512_storeu_ps(y, vf);
171     y += 16;
172   }
173   if XNN_UNLIKELY(n != 0) {
174     assert(n >= 1 * sizeof(float));
175     assert(n <= 15 * sizeof(float));
176 
177     // Prepare mask for valid 32-bit elements (depends on n).
178     n >>= 2 /* log2(sizeof(float)) */;
179     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
180 
181     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
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 vf = _mm512_div_ps(ve, vd);
199 
200     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
201 
202     _mm512_mask_storeu_ps(y, vmask, vf);
203   }
204 }
205