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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
41 const __m256 vx0 = _mm256_loadu_ps(x);
42 const __m256 vx1 = _mm256_loadu_ps(x + 8);
43 const __m256 vx2 = _mm256_loadu_ps(x + 16);
44 x += 24;
45
46 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
47 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
48 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
49
50 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
51 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
52 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
53
54 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
55 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
56 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
57 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
58 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
59 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
60 const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
61 const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
62 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1);
63
64 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
65 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
66 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
67
68 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
69 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
70 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
71
72 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
73 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
74 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
75
76 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
77 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
78 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc5, vt2), vc4);
79
80 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
81 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
82 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
83
84 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
85 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
86 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
87
88 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
89 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
90 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc1);
91
92 vt0 = _mm256_mul_ps(vt0, vs0);
93 vt1 = _mm256_mul_ps(vt1, vs1);
94 vt2 = _mm256_mul_ps(vt2, vs2);
95
96 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
97 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
98 const __m256 ve2 = _mm256_add_ps(_mm256_mul_ps(vt2, vp2), vs2);
99
100 const __m256 vd0 = _mm256_add_ps(ve0, vone);
101 const __m256 vd1 = _mm256_add_ps(ve1, vone);
102 const __m256 vd2 = _mm256_add_ps(ve2, vone);
103
104 __m256 vr0 = _mm256_rcp_ps(vd0);
105 __m256 vr1 = _mm256_rcp_ps(vd1);
106 __m256 vr2 = _mm256_rcp_ps(vd2);
107
108 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
109 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
110 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
111 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
112 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
113 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
114
115 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
116 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
117 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
118
119 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
120 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
121 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
122
123 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
124 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
125 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
126
127 _mm256_storeu_ps(y, vf0);
128 _mm256_storeu_ps(y + 8, vf1);
129 _mm256_storeu_ps(y + 16, vf2);
130 y += 24;
131 }
132 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
133 const __m256 vx = _mm256_loadu_ps(x);
134 x += 8;
135
136 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
137
138 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
139
140 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
141 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
142 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
143
144 vn = _mm256_sub_ps(vn, vmagic_bias);
145
146 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
147 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
148
149 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
150 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
151 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
152 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
153
154 vt = _mm256_mul_ps(vt, vs);
155 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
156
157 const __m256 vd = _mm256_add_ps(ve, vone);
158 __m256 vr = _mm256_rcp_ps(vd);
159 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
160 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
161 __m256 vf = _mm256_mul_ps(ve, vr);
162
163 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
164 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
165
166 _mm256_storeu_ps(y, vf);
167 y += 8;
168 }
169 if XNN_UNLIKELY(n != 0) {
170 assert(n >= 1 * sizeof(float));
171 assert(n <= 7 * sizeof(float));
172 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
173
174 const __m256 vx = _mm256_maskload_ps(x, vmask);
175
176 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
177
178 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
179 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
180 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
181 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
182
183 vn = _mm256_sub_ps(vn, vmagic_bias);
184
185 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
186 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
187
188 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
189 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
190 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
191 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
192
193 vt = _mm256_mul_ps(vt, vs);
194 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
195
196 const __m256 vd = _mm256_add_ps(ve, vone);
197 __m256 vr = _mm256_rcp_ps(vd);
198 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
199 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
200 __m256 vf = _mm256_mul_ps(ve, vr);
201
202 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
203 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
204
205 __m128 vf_lo = _mm256_castps256_ps128(vf);
206 if (n & (4 * sizeof(float))) {
207 _mm_storeu_ps(y, vf_lo);
208 vf_lo = _mm256_extractf128_ps(vf, 1);
209 y += 4;
210 }
211 if (n & (2 * sizeof(float))) {
212 _mm_storel_pi((__m64*) y, vf_lo);
213 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
214 y += 2;
215 }
216 if (n & (1 * sizeof(float))) {
217 _mm_store_ss(y, vf_lo);
218 }
219 }
220 }
221