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_div_x64(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_div_x64(
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 >= 64 * sizeof(float); n -= 64 * 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     x += 64;
42 
43     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
44     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
45     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
46     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
47 
48     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
49     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
50     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
51     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
52 
53     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
54     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
55     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
56     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
57 
58     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
59     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
60     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
61     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
62 
63     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
64     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
65     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
66     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
67 
68     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
69     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
70     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
71     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
72 
73     vp0 = _mm512_mul_ps(vp0, vt0);
74     vp1 = _mm512_mul_ps(vp1, vt1);
75     vp2 = _mm512_mul_ps(vp2, vt2);
76     vp3 = _mm512_mul_ps(vp3, vt3);
77 
78     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
79     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
80     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
81     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
82 
83     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
84     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
85     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
86     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
87 
88     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
89     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
90     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
91     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
92 
93     const __m512 vd0 = _mm512_add_ps(ve0, vone);
94     const __m512 vd1 = _mm512_add_ps(ve1, vone);
95     const __m512 vd2 = _mm512_add_ps(ve2, vone);
96     const __m512 vd3 = _mm512_add_ps(ve3, vone);
97 
98     __m512 vf0 = _mm512_div_ps(ve0, vd0);
99     __m512 vf1 = _mm512_div_ps(ve1, vd1);
100     __m512 vf2 = _mm512_div_ps(ve2, vd2);
101     __m512 vf3 = _mm512_div_ps(ve3, vd3);
102 
103     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
104     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
105     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
106     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
107 
108     _mm512_storeu_ps(y, vf0);
109     _mm512_storeu_ps(y + 16, vf1);
110     _mm512_storeu_ps(y + 32, vf2);
111     _mm512_storeu_ps(y + 48, vf3);
112     y += 64;
113   }
114   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
115     const __m512 vx = _mm512_loadu_ps(x);
116     x += 16;
117 
118     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
119 
120     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
121     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
122     vn = _mm512_sub_ps(vn, vmagic_bias);
123 
124     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
125 
126     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
127     vp = _mm512_mul_ps(vp, vt);
128     vp = _mm512_fmadd_ps(vt, vp, vt);
129     vp = _mm512_fmadd_ps(vl, vp, vl);
130 
131     const __m512 ve = _mm512_scalef_ps(vp, vn);
132     const __m512 vd = _mm512_add_ps(ve, vone);
133 
134     __m512 vf = _mm512_div_ps(ve, vd);
135 
136     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
137 
138     _mm512_storeu_ps(y, vf);
139     y += 16;
140   }
141   if XNN_UNLIKELY(n != 0) {
142     assert(n >= 1 * sizeof(float));
143     assert(n <= 15 * sizeof(float));
144 
145     // Prepare mask for valid 32-bit elements (depends on n).
146     n >>= 2 /* log2(sizeof(float)) */;
147     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
148 
149     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
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_permutexvar_ps(_mm512_castps_si512(vn), vtable);
154     vn = _mm512_sub_ps(vn, vmagic_bias);
155 
156     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
157 
158     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
159     vp = _mm512_mul_ps(vp, vt);
160     vp = _mm512_fmadd_ps(vt, vp, vt);
161     vp = _mm512_fmadd_ps(vl, 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_mask_storeu_ps(y, vmask, vf);
171   }
172 }
173