1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-p6.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_velu_ukernel__avx2_rr1_p6_x64(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__avx2_rr1_p6_x64(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vprescale = _mm256_load_ps(params->avx2_rr1_p6.prescale);
27 const __m256 valpha = _mm256_load_ps(params->avx2_rr1_p6.alpha);
28 const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_p6.beta);
29 const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_p6.sat_cutoff);
30 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p6.magic_bias);
31 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p6.log2e);
32 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p6.minus_ln2);
33 const __m256 vc6 = _mm256_load_ps(params->avx2_rr1_p6.c6);
34 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p6.c5);
35 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p6.c4);
36 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p6.c3);
37 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p6.c2);
38
39 for (; n >= 64 * sizeof(float); n -= 64 * sizeof(float)) {
40 __m256 vx0 = _mm256_loadu_ps(x);
41 __m256 vx1 = _mm256_loadu_ps(x + 8);
42 __m256 vx2 = _mm256_loadu_ps(x + 16);
43 __m256 vx3 = _mm256_loadu_ps(x + 24);
44 __m256 vx4 = _mm256_loadu_ps(x + 32);
45 __m256 vx5 = _mm256_loadu_ps(x + 40);
46 __m256 vx6 = _mm256_loadu_ps(x + 48);
47 __m256 vx7 = _mm256_loadu_ps(x + 56);
48 x += 64;
49
50 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
51 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
52 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
53 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
54 const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
55 const __m256 vz5 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx5, vprescale));
56 const __m256 vz6 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx6, vprescale));
57 const __m256 vz7 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx7, vprescale));
58
59 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
60 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
61 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
62 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
63 __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
64 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias);
65 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias);
66 __m256 vn7 = _mm256_fmadd_ps(vz7, vlog2e, vmagic_bias);
67
68 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
69 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
70 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
71 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
72 __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
73 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
74 __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
75 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
76 __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
77 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
78 __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
79 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
80 __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
81 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
82 __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
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(vc6, vt0, vc5);
95 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
96 __m256 vp2 = _mm256_fmadd_ps(vc6, vt2, vc5);
97 __m256 vp3 = _mm256_fmadd_ps(vc6, vt3, vc5);
98 __m256 vp4 = _mm256_fmadd_ps(vc6, vt4, vc5);
99 __m256 vp5 = _mm256_fmadd_ps(vc6, vt5, vc5);
100 __m256 vp6 = _mm256_fmadd_ps(vc6, vt6, vc5);
101 __m256 vp7 = _mm256_fmadd_ps(vc6, vt7, vc5);
102
103 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
104 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
105 vp2 = _mm256_fmadd_ps(vp2, vt2, vc4);
106 vp3 = _mm256_fmadd_ps(vp3, vt3, vc4);
107 vp4 = _mm256_fmadd_ps(vp4, vt4, vc4);
108 vp5 = _mm256_fmadd_ps(vp5, vt5, vc4);
109 vp6 = _mm256_fmadd_ps(vp6, vt6, vc4);
110 vp7 = _mm256_fmadd_ps(vp7, vt7, vc4);
111
112 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
113 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
114 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
115 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
116 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
117 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
118 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
119 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
120
121 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
122 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
123 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
124 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
125 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
126 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
127 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
128 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
129
130 vp0 = _mm256_mul_ps(vp0, vt0);
131 vt0 = _mm256_mul_ps(vt0, vs0);
132 vp1 = _mm256_mul_ps(vp1, vt1);
133 vt1 = _mm256_mul_ps(vt1, vs1);
134 vp2 = _mm256_mul_ps(vp2, vt2);
135 vt2 = _mm256_mul_ps(vt2, vs2);
136 vp3 = _mm256_mul_ps(vp3, vt3);
137 vt3 = _mm256_mul_ps(vt3, vs3);
138 vp4 = _mm256_mul_ps(vp4, vt4);
139 vt4 = _mm256_mul_ps(vt4, vs4);
140 vp5 = _mm256_mul_ps(vp5, vt5);
141 vt5 = _mm256_mul_ps(vt5, vs5);
142 vp6 = _mm256_mul_ps(vp6, vt6);
143 vt6 = _mm256_mul_ps(vt6, vs6);
144 vp7 = _mm256_mul_ps(vp7, vt7);
145 vt7 = _mm256_mul_ps(vt7, vs7);
146
147 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
148 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
149 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
150 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
151 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
152 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
153 vs3 = _mm256_fmsub_ps(vs3, valpha, valpha);
154 vp3 = _mm256_fmadd_ps(vp3, vt3, vt3);
155 vs4 = _mm256_fmsub_ps(vs4, valpha, valpha);
156 vp4 = _mm256_fmadd_ps(vp4, vt4, vt4);
157 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha);
158 vp5 = _mm256_fmadd_ps(vp5, vt5, vt5);
159 vs6 = _mm256_fmsub_ps(vs6, valpha, valpha);
160 vp6 = _mm256_fmadd_ps(vp6, vt6, vt6);
161 vs7 = _mm256_fmsub_ps(vs7, valpha, valpha);
162 vp7 = _mm256_fmadd_ps(vp7, vt7, vt7);
163
164 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
165 vx0 = _mm256_mul_ps(vx0, vbeta);
166 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
167 vx1 = _mm256_mul_ps(vx1, vbeta);
168 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
169 vx2 = _mm256_mul_ps(vx2, vbeta);
170 const __m256 ve3 = _mm256_fmadd_ps(vp3, valpha, vs3);
171 vx3 = _mm256_mul_ps(vx3, vbeta);
172 const __m256 ve4 = _mm256_fmadd_ps(vp4, valpha, vs4);
173 vx4 = _mm256_mul_ps(vx4, vbeta);
174 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5);
175 vx5 = _mm256_mul_ps(vx5, vbeta);
176 const __m256 ve6 = _mm256_fmadd_ps(vp6, valpha, vs6);
177 vx6 = _mm256_mul_ps(vx6, vbeta);
178 const __m256 ve7 = _mm256_fmadd_ps(vp7, valpha, vs7);
179 vx7 = _mm256_mul_ps(vx7, vbeta);
180
181 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
182 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
183 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
184 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
185 const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
186 const __m256 vy5 = _mm256_blendv_ps(vx5, ve5, vx5);
187 const __m256 vy6 = _mm256_blendv_ps(vx6, ve6, vx6);
188 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7);
189
190 _mm256_storeu_ps(y, vy0);
191 _mm256_storeu_ps(y + 8, vy1);
192 _mm256_storeu_ps(y + 16, vy2);
193 _mm256_storeu_ps(y + 24, vy3);
194 _mm256_storeu_ps(y + 32, vy4);
195 _mm256_storeu_ps(y + 40, vy5);
196 _mm256_storeu_ps(y + 48, vy6);
197 _mm256_storeu_ps(y + 56, vy7);
198 y += 64;
199 }
200 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
201 __m256 vx = _mm256_loadu_ps(x);
202 x += 8;
203
204 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
205
206 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
207 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
208 vn = _mm256_sub_ps(vn, vmagic_bias);
209
210 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
211
212 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
213 vp = _mm256_fmadd_ps(vp, vt, vc4);
214 vp = _mm256_fmadd_ps(vp, vt, vc3);
215 vp = _mm256_fmadd_ps(vp, vt, vc2);
216 vp = _mm256_mul_ps(vp, vt);
217
218 vt = _mm256_mul_ps(vt, vs);
219 vs = _mm256_fmsub_ps(vs, valpha, valpha);
220 vp = _mm256_fmadd_ps(vp, vt, vt);
221 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
222
223 vx = _mm256_mul_ps(vx, vbeta);
224 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
225
226 _mm256_storeu_ps(y, vy);
227 y += 8;
228 }
229 if XNN_UNLIKELY(n != 0) {
230 assert(n >= 1 * sizeof(float));
231 assert(n <= 7 * sizeof(float));
232 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p6.mask_table[7] - n));
233
234 __m256 vx = _mm256_maskload_ps(x, vmask);
235
236 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
237
238 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
239 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
240 vn = _mm256_sub_ps(vn, vmagic_bias);
241
242 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
243
244 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
245 vp = _mm256_fmadd_ps(vp, vt, vc4);
246 vp = _mm256_fmadd_ps(vp, vt, vc3);
247 vp = _mm256_fmadd_ps(vp, vt, vc2);
248 vp = _mm256_mul_ps(vp, vt);
249
250 vt = _mm256_mul_ps(vt, vs);
251 vs = _mm256_fmsub_ps(vs, valpha, valpha);
252 vp = _mm256_fmadd_ps(vp, vt, vt);
253 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
254
255 vx = _mm256_mul_ps(vx, vbeta);
256 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
257
258 __m128 vy_lo = _mm256_castps256_ps128(vy);
259 if (n & (4 * sizeof(float))) {
260 _mm_storeu_ps(y, vy_lo);
261 vy_lo = _mm256_extractf128_ps(vy, 1);
262 y += 4;
263 }
264 if (n & (2 * sizeof(float))) {
265 _mm_storel_pi((__m64*) y, vy_lo);
266 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
267 y += 2;
268 }
269 if (n & (1 * sizeof(float))) {
270 _mm_store_ss(y, vy_lo);
271 }
272 }
273 }
274