• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x48(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_x48(
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 >= 48 * sizeof(float); n -= 48 * 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     x += 48;
43 
44     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
45     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
46     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), 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 
52     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
53     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
54     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
55 
56     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
57     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
58     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
59 
60     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
61     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
62     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
63 
64     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
65     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
66     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
67 
68     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
69     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
70     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
71 
72     vt0 = _mm512_mul_ps(vt0, vl0);
73     vt1 = _mm512_mul_ps(vt1, vl1);
74     vt2 = _mm512_mul_ps(vt2, vl2);
75 
76     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
77     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
78     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
79 
80     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
81     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
82     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
83 
84     const __m512 vd0 = _mm512_add_ps(ve0, vone);
85     const __m512 vd1 = _mm512_add_ps(ve1, vone);
86     const __m512 vd2 = _mm512_add_ps(ve2, vone);
87 
88     __m512 vf0 = _mm512_div_ps(ve0, vd0);
89     __m512 vf1 = _mm512_div_ps(ve1, vd1);
90     __m512 vf2 = _mm512_div_ps(ve2, vd2);
91 
92     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
93     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
94     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
95 
96     _mm512_storeu_ps(y, vf0);
97     _mm512_storeu_ps(y + 16, vf1);
98     _mm512_storeu_ps(y + 32, vf2);
99     y += 48;
100   }
101   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
102     const __m512 vx = _mm512_loadu_ps(x);
103     x += 16;
104 
105     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
106 
107     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
108     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
109     vn = _mm512_sub_ps(vn, vmagic_bias);
110 
111     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
112     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
113 
114     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
115     vt = _mm512_mul_ps(vt, vl);
116     vp = _mm512_fmadd_ps(vt, vp, vl);
117 
118     const __m512 ve = _mm512_scalef_ps(vp, vn);
119     const __m512 vd = _mm512_add_ps(ve, vone);
120 
121     __m512 vf = _mm512_div_ps(ve, vd);
122 
123     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
124 
125     _mm512_storeu_ps(y, vf);
126     y += 16;
127   }
128   if XNN_UNLIKELY(n != 0) {
129     assert(n >= 1 * sizeof(float));
130     assert(n <= 15 * sizeof(float));
131 
132     // Prepare mask for valid 32-bit elements (depends on n).
133     n >>= 2 /* log2(sizeof(float)) */;
134     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
135 
136     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
137     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
138 
139     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
140     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
141     vn = _mm512_sub_ps(vn, vmagic_bias);
142 
143     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
144     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
145 
146     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
147     vt = _mm512_mul_ps(vt, vl);
148     vp = _mm512_fmadd_ps(vt, vp, vl);
149 
150     const __m512 ve = _mm512_scalef_ps(vp, vn);
151     const __m512 vd = _mm512_add_ps(ve, vone);
152 
153     __m512 vf = _mm512_div_ps(ve, vd);
154 
155     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
156 
157     _mm512_mask_storeu_ps(y, vmask, vf);
158   }
159 }
160