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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
39 const __m512 vx0 = _mm512_loadu_ps(x);
40 const __m512 vx1 = _mm512_loadu_ps(x + 16);
41 x += 32;
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
46 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
47 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
48
49 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
50 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
51
52 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
53 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
54
55 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
56 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
57
58 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
59 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
60
61 __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
62 __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
63
64 vt0 = _mm512_mul_ps(vt0, vl0);
65 vt1 = _mm512_mul_ps(vt1, vl1);
66
67 vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
68 vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
69
70 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
71 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
72
73 const __m512 vd0 = _mm512_add_ps(ve0, vone);
74 const __m512 vd1 = _mm512_add_ps(ve1, vone);
75
76 __m512 vr0 = _mm512_rcp14_ps(vd0);
77 __m512 vr1 = _mm512_rcp14_ps(vd1);
78
79 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
80 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
81
82 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
83 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
84
85
86 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
87 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
88
89 _mm512_storeu_ps(y, vf0);
90 _mm512_storeu_ps(y + 16, vf1);
91 y += 32;
92 }
93 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
94 const __m512 vx = _mm512_loadu_ps(x);
95 x += 16;
96
97 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
98
99 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
100 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
101 vn = _mm512_sub_ps(vn, vmagic_bias);
102
103 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
104 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
105
106 __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
107 vt = _mm512_mul_ps(vt, vl);
108 vp = _mm512_fmadd_ps(vt, vp, vl);
109
110 const __m512 ve = _mm512_scalef_ps(vp, vn);
111 const __m512 vd = _mm512_add_ps(ve, vone);
112
113 __m512 vr = _mm512_rcp14_ps(vd);
114 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
115
116 __m512 vf = _mm512_mul_ps(ve, vr);
117
118 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
119
120 _mm512_storeu_ps(y, vf);
121 y += 16;
122 }
123 if XNN_UNLIKELY(n != 0) {
124 assert(n >= 1 * sizeof(float));
125 assert(n <= 15 * sizeof(float));
126
127 // Prepare mask for valid 32-bit elements (depends on n).
128 n >>= 2 /* log2(sizeof(float)) */;
129 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
130
131 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
132 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
133
134 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
135 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
136 vn = _mm512_sub_ps(vn, vmagic_bias);
137
138 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
139 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
140
141 __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
142 vt = _mm512_mul_ps(vt, vl);
143 vp = _mm512_fmadd_ps(vt, vp, vl);
144
145 const __m512 ve = _mm512_scalef_ps(vp, vn);
146 const __m512 vd = _mm512_add_ps(ve, vone);
147
148 __m512 vr = _mm512_rcp14_ps(vd);
149 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
150
151 __m512 vf = _mm512_mul_ps(ve, vr);
152
153 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
154
155 _mm512_mask_storeu_ps(y, vmask, vf);
156 }
157 }
158