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_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_rr2_lut32_p2_perm2_scalef_nr1fma_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_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 >= 64 * sizeof(float); n -= 64 * 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 x += 64;
44
45 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
46 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
47 const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
48 const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
49
50 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
51 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
52 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
53 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
54
55 const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
56 const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
57 const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
58 const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
59
60 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
61 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
62 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
63 vn3 = _mm512_sub_ps(vn3, vmagic_bias);
64
65 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
66 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
67 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
68 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
69
70 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
71 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
72 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
73 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
74
75 __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
76 __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
77 __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
78 __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
79
80 vt0 = _mm512_mul_ps(vt0, vl0);
81 vt1 = _mm512_mul_ps(vt1, vl1);
82 vt2 = _mm512_mul_ps(vt2, vl2);
83 vt3 = _mm512_mul_ps(vt3, vl3);
84
85 vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
86 vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
87 vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
88 vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
89
90 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
91 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
92 const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
93 const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
94
95 const __m512 vd0 = _mm512_add_ps(ve0, vone);
96 const __m512 vd1 = _mm512_add_ps(ve1, vone);
97 const __m512 vd2 = _mm512_add_ps(ve2, vone);
98 const __m512 vd3 = _mm512_add_ps(ve3, vone);
99
100 __m512 vr0 = _mm512_rcp14_ps(vd0);
101 __m512 vr1 = _mm512_rcp14_ps(vd1);
102 __m512 vr2 = _mm512_rcp14_ps(vd2);
103 __m512 vr3 = _mm512_rcp14_ps(vd3);
104
105 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
106 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
107 vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
108 vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
109
110 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
111 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
112 __m512 vf2 = _mm512_mul_ps(ve2, vr2);
113 __m512 vf3 = _mm512_mul_ps(ve3, vr3);
114
115
116 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
117 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
118 vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
119 vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
120
121 _mm512_storeu_ps(y, vf0);
122 _mm512_storeu_ps(y + 16, vf1);
123 _mm512_storeu_ps(y + 32, vf2);
124 _mm512_storeu_ps(y + 48, vf3);
125 y += 64;
126 }
127 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
128 const __m512 vx = _mm512_loadu_ps(x);
129 x += 16;
130
131 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
132
133 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
134 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
135 vn = _mm512_sub_ps(vn, vmagic_bias);
136
137 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
138 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
139
140 __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
141 vt = _mm512_mul_ps(vt, vl);
142 vp = _mm512_fmadd_ps(vt, vp, vl);
143
144 const __m512 ve = _mm512_scalef_ps(vp, vn);
145 const __m512 vd = _mm512_add_ps(ve, vone);
146
147 __m512 vr = _mm512_rcp14_ps(vd);
148 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
149
150 __m512 vf = _mm512_mul_ps(ve, vr);
151
152 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
153
154 _mm512_storeu_ps(y, vf);
155 y += 16;
156 }
157 if XNN_UNLIKELY(n != 0) {
158 assert(n >= 1 * sizeof(float));
159 assert(n <= 15 * sizeof(float));
160
161 // Prepare mask for valid 32-bit elements (depends on n).
162 n >>= 2 /* log2(sizeof(float)) */;
163 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
164
165 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
166 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
167
168 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
169 const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
170 vn = _mm512_sub_ps(vn, vmagic_bias);
171
172 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
173 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
174
175 __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
176 vt = _mm512_mul_ps(vt, vl);
177 vp = _mm512_fmadd_ps(vt, vp, vl);
178
179 const __m512 ve = _mm512_scalef_ps(vp, vn);
180 const __m512 vd = _mm512_add_ps(ve, vone);
181
182 __m512 vr = _mm512_rcp14_ps(vd);
183 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
184
185 __m512 vf = _mm512_mul_ps(ve, vr);
186
187 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
188
189 _mm512_mask_storeu_ps(y, vmask, vf);
190 }
191 }
192