1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx-rr2-p5.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/vunary.h>
16
17
xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x16(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx_rr2_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p5.log2e);
29 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_hi);
30 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_lo);
31 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p5.c5);
32 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p5.c4);
33 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p5.c3);
34 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p5.c2);
35 const __m256 vc1 = _mm256_load_ps(params->avx_rr2_p5.c1);
36 const __m256 vone = _mm256_load_ps(params->avx_rr2_p5.one);
37 const __m256 vtwo = _mm256_load_ps(params->avx_rr2_p5.two);
38 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx_rr2_p5.denorm_cutoff);
39
40 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
41 const __m256 vx0 = _mm256_loadu_ps(x);
42 const __m256 vx1 = _mm256_loadu_ps(x + 8);
43 x += 16;
44
45 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
46 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
47
48 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
49 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
50
51 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
52 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
53 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
54 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
55 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
56 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
57
58 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
59 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
60
61 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
62 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
63
64 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
65 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
66
67 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
68 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
69
70 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
71 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
72
73 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
74 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
75
76 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
77 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
78
79 vt0 = _mm256_mul_ps(vt0, vs0);
80 vt1 = _mm256_mul_ps(vt1, vs1);
81
82 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
83 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
84
85 const __m256 vd0 = _mm256_add_ps(ve0, vone);
86 const __m256 vd1 = _mm256_add_ps(ve1, vone);
87
88 __m256 vr0 = _mm256_rcp_ps(vd0);
89 __m256 vr1 = _mm256_rcp_ps(vd1);
90
91 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
92 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
93 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
94 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
95
96 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
97 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
98
99 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
100 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
101
102 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
103 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
104
105 _mm256_storeu_ps(y, vf0);
106 _mm256_storeu_ps(y + 8, vf1);
107 y += 16;
108 }
109 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
110 const __m256 vx = _mm256_loadu_ps(x);
111 x += 8;
112
113 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
114
115 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
116
117 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
118 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
119 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
120
121 vn = _mm256_sub_ps(vn, vmagic_bias);
122
123 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
124 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
125
126 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
127 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
128 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
129 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
130
131 vt = _mm256_mul_ps(vt, vs);
132 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
133
134 const __m256 vd = _mm256_add_ps(ve, vone);
135 __m256 vr = _mm256_rcp_ps(vd);
136 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
137 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
138 __m256 vf = _mm256_mul_ps(ve, vr);
139
140 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
141 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
142
143 _mm256_storeu_ps(y, vf);
144 y += 8;
145 }
146 if XNN_UNLIKELY(n != 0) {
147 assert(n >= 1 * sizeof(float));
148 assert(n <= 7 * sizeof(float));
149 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
150
151 const __m256 vx = _mm256_maskload_ps(x, vmask);
152
153 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
154
155 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
156 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
157 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
158 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
159
160 vn = _mm256_sub_ps(vn, vmagic_bias);
161
162 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
163 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
164
165 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
166 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
167 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
168 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
169
170 vt = _mm256_mul_ps(vt, vs);
171 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
172
173 const __m256 vd = _mm256_add_ps(ve, vone);
174 __m256 vr = _mm256_rcp_ps(vd);
175 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
176 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
177 __m256 vf = _mm256_mul_ps(ve, vr);
178
179 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
180 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
181
182 __m128 vf_lo = _mm256_castps256_ps128(vf);
183 if (n & (4 * sizeof(float))) {
184 _mm_storeu_ps(y, vf_lo);
185 vf_lo = _mm256_extractf128_ps(vf, 1);
186 y += 4;
187 }
188 if (n & (2 * sizeof(float))) {
189 _mm_storel_pi((__m64*) y, vf_lo);
190 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
191 y += 2;
192 }
193 if (n & (1 * sizeof(float))) {
194 _mm_store_ss(y, vf_lo);
195 }
196 }
197 }
198