1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx2-rr1-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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__avx2_rr1_p5_div_x64(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__avx2_rr1_p5_div_x64(
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->avx2_rr1_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
29 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
30 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
31 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
32 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
33 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
34 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
35 const __m256 vone = _mm256_load_ps(params->avx2_rr1_p5.one);
36 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);
37
38 for (; n >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
39 const __m256 vx0 = _mm256_loadu_ps(x);
40 const __m256 vx1 = _mm256_loadu_ps(x + 8);
41 const __m256 vx2 = _mm256_loadu_ps(x + 16);
42 const __m256 vx3 = _mm256_loadu_ps(x + 24);
43 const __m256 vx4 = _mm256_loadu_ps(x + 32);
44 const __m256 vx5 = _mm256_loadu_ps(x + 40);
45 const __m256 vx6 = _mm256_loadu_ps(x + 48);
46 const __m256 vx7 = _mm256_loadu_ps(x + 56);
47 x += 64;
48
49 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
50 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
51 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
52 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
53 const __m256 vz4 = _mm256_or_ps(vx4, vsign_mask);
54 const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask);
55 const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask);
56 const __m256 vz7 = _mm256_or_ps(vx7, vsign_mask);
57
58 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
59 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
60 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
61 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
62 __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
63 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias);
64 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias);
65 __m256 vn7 = _mm256_fmadd_ps(vz7, vlog2e, vmagic_bias);
66
67 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
68 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
69 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
70 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
71 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
72 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
73 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
74 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
75
76 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
77 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
78 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
79 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
80 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
81 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
82 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
83 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
84
85 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
86 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
87 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
88 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
89 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
90 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5);
91 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6);
92 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vz7);
93
94 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
95 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
96 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
97 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
98 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
99 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
100 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
101 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
102
103 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
104 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
105 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
106 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
107 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
108 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
109 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
110 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
111
112 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
113 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
114 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
115 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
116 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
117 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
118 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
119 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
120
121 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
122 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
123 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
124 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
125 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
126 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
127 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
128 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
129
130 vt0 = _mm256_mul_ps(vt0, vs0);
131 vt1 = _mm256_mul_ps(vt1, vs1);
132 vt2 = _mm256_mul_ps(vt2, vs2);
133 vt3 = _mm256_mul_ps(vt3, vs3);
134 vt4 = _mm256_mul_ps(vt4, vs4);
135 vt5 = _mm256_mul_ps(vt5, vs5);
136 vt6 = _mm256_mul_ps(vt6, vs6);
137 vt7 = _mm256_mul_ps(vt7, vs7);
138
139 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
140 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
141 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
142 const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
143 const __m256 ve4 = _mm256_fmadd_ps(vt4, vp4, vs4);
144 const __m256 ve5 = _mm256_fmadd_ps(vt5, vp5, vs5);
145 const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6);
146 const __m256 ve7 = _mm256_fmadd_ps(vt7, vp7, vs7);
147
148 const __m256 vd0 = _mm256_add_ps(ve0, vone);
149 const __m256 vd1 = _mm256_add_ps(ve1, vone);
150 const __m256 vd2 = _mm256_add_ps(ve2, vone);
151 const __m256 vd3 = _mm256_add_ps(ve3, vone);
152 const __m256 vd4 = _mm256_add_ps(ve4, vone);
153 const __m256 vd5 = _mm256_add_ps(ve5, vone);
154 const __m256 vd6 = _mm256_add_ps(ve6, vone);
155 const __m256 vd7 = _mm256_add_ps(ve7, vone);
156
157 __m256 vf0 = _mm256_div_ps(ve0, vd0);
158 __m256 vf1 = _mm256_div_ps(ve1, vd1);
159 __m256 vf2 = _mm256_div_ps(ve2, vd2);
160 __m256 vf3 = _mm256_div_ps(ve3, vd3);
161 __m256 vf4 = _mm256_div_ps(ve4, vd4);
162 __m256 vf5 = _mm256_div_ps(ve5, vd5);
163 __m256 vf6 = _mm256_div_ps(ve6, vd6);
164 __m256 vf7 = _mm256_div_ps(ve7, vd7);
165
166 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
167 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
168 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
169 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
170 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vz4, vdenorm_cutoff, _CMP_LT_OS), vf4);
171 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5);
172 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6);
173 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vz7, vdenorm_cutoff, _CMP_LT_OS), vf7);
174
175 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
176 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
177 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
178 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
179 vf4 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf4), vf4, vx4);
180 vf5 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf5), vf5, vx5);
181 vf6 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf6), vf6, vx6);
182 vf7 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf7), vf7, vx7);
183
184 _mm256_storeu_ps(y, vf0);
185 _mm256_storeu_ps(y + 8, vf1);
186 _mm256_storeu_ps(y + 16, vf2);
187 _mm256_storeu_ps(y + 24, vf3);
188 _mm256_storeu_ps(y + 32, vf4);
189 _mm256_storeu_ps(y + 40, vf5);
190 _mm256_storeu_ps(y + 48, vf6);
191 _mm256_storeu_ps(y + 56, vf7);
192 y += 64;
193 }
194 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
195 const __m256 vx = _mm256_loadu_ps(x);
196 x += 8;
197
198 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
199
200 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
201 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
202 vn = _mm256_sub_ps(vn, vmagic_bias);
203
204 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
205
206 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
207 vp = _mm256_fmadd_ps(vp, vt, vc3);
208 vp = _mm256_fmadd_ps(vp, vt, vc2);
209 vp = _mm256_fmadd_ps(vp, vt, vc1);
210
211 vt = _mm256_mul_ps(vt, vs);
212 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
213
214 const __m256 vd = _mm256_add_ps(ve, vone);
215 __m256 vf = _mm256_div_ps(ve, vd);
216
217 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
218 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
219
220 _mm256_storeu_ps(y, vf);
221 y += 8;
222 }
223 if XNN_UNLIKELY(n != 0) {
224 assert(n >= 1 * sizeof(float));
225 assert(n <= 7 * sizeof(float));
226 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - n));
227
228 const __m256 vx = _mm256_maskload_ps(x, vmask);
229
230 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
231
232 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
233 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
234 vn = _mm256_sub_ps(vn, vmagic_bias);
235
236 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
237
238 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
239 vp = _mm256_fmadd_ps(vp, vt, vc3);
240 vp = _mm256_fmadd_ps(vp, vt, vc2);
241 vp = _mm256_fmadd_ps(vp, vt, vc1);
242
243 vt = _mm256_mul_ps(vt, vs);
244 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
245
246 const __m256 vd = _mm256_add_ps(ve, vone);
247 __m256 vf = _mm256_div_ps(ve, vd);
248
249 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
250 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
251
252 __m128 vf_lo = _mm256_castps256_ps128(vf);
253 if (n & (4 * sizeof(float))) {
254 _mm_storeu_ps(y, vf_lo);
255 vf_lo = _mm256_extractf128_ps(vf, 1);
256 y += 4;
257 }
258 if (n & (2 * sizeof(float))) {
259 _mm_storel_pi((__m64*) y, vf_lo);
260 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
261 y += 2;
262 }
263 if (n & (1 * sizeof(float))) {
264 _mm_store_ss(y, vf_lo);
265 }
266 }
267 }
268