1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-lut8-p4-perm.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_lut8_p4_perm_x72(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_lut8_p4_perm_x72(
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_lut8_p4.prescale);
27 const __m256 valpha = _mm256_load_ps(params->avx2_rr1_lut8_p4.alpha);
28 const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_lut8_p4.beta);
29 const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_lut8_p4.sat_cutoff);
30 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_lut8_p4.magic_bias);
31 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_lut8_p4.log2e);
32 const __m256i vtable = _mm256_load_si256((const __m256i*) params->avx2_rr1_lut8_p4.table);
33 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_lut8_p4.minus_ln2);
34 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c4);
35 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c3);
36 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_lut8_p4.c2);
37
38 for (; n >= 72 * sizeof(float); n -= 72 * sizeof(float)) {
39 __m256 vx0 = _mm256_loadu_ps(x);
40 __m256 vx1 = _mm256_loadu_ps(x + 8);
41 __m256 vx2 = _mm256_loadu_ps(x + 16);
42 __m256 vx3 = _mm256_loadu_ps(x + 24);
43 __m256 vx4 = _mm256_loadu_ps(x + 32);
44 __m256 vx5 = _mm256_loadu_ps(x + 40);
45 __m256 vx6 = _mm256_loadu_ps(x + 48);
46 __m256 vx7 = _mm256_loadu_ps(x + 56);
47 __m256 vx8 = _mm256_loadu_ps(x + 64);
48 x += 72;
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 const __m256 vz8 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx8, vprescale));
59
60 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
61 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
62 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
63 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
64 __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
65 __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias);
66 __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias);
67 __m256 vn7 = _mm256_fmadd_ps(vz7, vlog2e, vmagic_bias);
68 __m256 vn8 = _mm256_fmadd_ps(vz8, vlog2e, vmagic_bias);
69
70 const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 20);
71 const __m256i vl0 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn0));
72 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
73 const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 20);
74 const __m256i vl1 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn1));
75 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
76 const __m256i ven2 = _mm256_slli_epi32(_mm256_castps_si256(vn2), 20);
77 const __m256i vl2 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn2));
78 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
79 const __m256i ven3 = _mm256_slli_epi32(_mm256_castps_si256(vn3), 20);
80 const __m256i vl3 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn3));
81 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
82 const __m256i ven4 = _mm256_slli_epi32(_mm256_castps_si256(vn4), 20);
83 const __m256i vl4 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn4));
84 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
85 const __m256i ven5 = _mm256_slli_epi32(_mm256_castps_si256(vn5), 20);
86 const __m256i vl5 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn5));
87 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
88 const __m256i ven6 = _mm256_slli_epi32(_mm256_castps_si256(vn6), 20);
89 const __m256i vl6 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn6));
90 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
91 const __m256i ven7 = _mm256_slli_epi32(_mm256_castps_si256(vn7), 20);
92 const __m256i vl7 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn7));
93 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
94 const __m256i ven8 = _mm256_slli_epi32(_mm256_castps_si256(vn8), 20);
95 const __m256i vl8 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn8));
96 vn8 = _mm256_sub_ps(vn8, vmagic_bias);
97
98 __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
99 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
100 __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
101 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
102 __m256 vs2 = _mm256_castsi256_ps(_mm256_add_epi32(vl2, ven2));
103 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
104 __m256 vs3 = _mm256_castsi256_ps(_mm256_add_epi32(vl3, ven3));
105 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
106 __m256 vs4 = _mm256_castsi256_ps(_mm256_add_epi32(vl4, ven4));
107 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
108 __m256 vs5 = _mm256_castsi256_ps(_mm256_add_epi32(vl5, ven5));
109 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5);
110 __m256 vs6 = _mm256_castsi256_ps(_mm256_add_epi32(vl6, ven6));
111 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6);
112 __m256 vs7 = _mm256_castsi256_ps(_mm256_add_epi32(vl7, ven7));
113 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vz7);
114 __m256 vs8 = _mm256_castsi256_ps(_mm256_add_epi32(vl8, ven8));
115 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2, vz8);
116
117 __m256 vp0 = _mm256_fmadd_ps(vc4, vt0, vc3);
118 __m256 vp1 = _mm256_fmadd_ps(vc4, vt1, vc3);
119 __m256 vp2 = _mm256_fmadd_ps(vc4, vt2, vc3);
120 __m256 vp3 = _mm256_fmadd_ps(vc4, vt3, vc3);
121 __m256 vp4 = _mm256_fmadd_ps(vc4, vt4, vc3);
122 __m256 vp5 = _mm256_fmadd_ps(vc4, vt5, vc3);
123 __m256 vp6 = _mm256_fmadd_ps(vc4, vt6, vc3);
124 __m256 vp7 = _mm256_fmadd_ps(vc4, vt7, vc3);
125 __m256 vp8 = _mm256_fmadd_ps(vc4, vt8, vc3);
126
127 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
128 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
129 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
130 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
131 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
132 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
133 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
134 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
135 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
136
137 vp0 = _mm256_mul_ps(vp0, vt0);
138 vt0 = _mm256_mul_ps(vt0, vs0);
139 vp1 = _mm256_mul_ps(vp1, vt1);
140 vt1 = _mm256_mul_ps(vt1, vs1);
141 vp2 = _mm256_mul_ps(vp2, vt2);
142 vt2 = _mm256_mul_ps(vt2, vs2);
143 vp3 = _mm256_mul_ps(vp3, vt3);
144 vt3 = _mm256_mul_ps(vt3, vs3);
145 vp4 = _mm256_mul_ps(vp4, vt4);
146 vt4 = _mm256_mul_ps(vt4, vs4);
147 vp5 = _mm256_mul_ps(vp5, vt5);
148 vt5 = _mm256_mul_ps(vt5, vs5);
149 vp6 = _mm256_mul_ps(vp6, vt6);
150 vt6 = _mm256_mul_ps(vt6, vs6);
151 vp7 = _mm256_mul_ps(vp7, vt7);
152 vt7 = _mm256_mul_ps(vt7, vs7);
153 vp8 = _mm256_mul_ps(vp8, vt8);
154 vt8 = _mm256_mul_ps(vt8, vs8);
155
156 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
157 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
158 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
159 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
160 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
161 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
162 vs3 = _mm256_fmsub_ps(vs3, valpha, valpha);
163 vp3 = _mm256_fmadd_ps(vp3, vt3, vt3);
164 vs4 = _mm256_fmsub_ps(vs4, valpha, valpha);
165 vp4 = _mm256_fmadd_ps(vp4, vt4, vt4);
166 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha);
167 vp5 = _mm256_fmadd_ps(vp5, vt5, vt5);
168 vs6 = _mm256_fmsub_ps(vs6, valpha, valpha);
169 vp6 = _mm256_fmadd_ps(vp6, vt6, vt6);
170 vs7 = _mm256_fmsub_ps(vs7, valpha, valpha);
171 vp7 = _mm256_fmadd_ps(vp7, vt7, vt7);
172 vs8 = _mm256_fmsub_ps(vs8, valpha, valpha);
173 vp8 = _mm256_fmadd_ps(vp8, vt8, vt8);
174
175 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
176 vx0 = _mm256_mul_ps(vx0, vbeta);
177 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
178 vx1 = _mm256_mul_ps(vx1, vbeta);
179 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
180 vx2 = _mm256_mul_ps(vx2, vbeta);
181 const __m256 ve3 = _mm256_fmadd_ps(vp3, valpha, vs3);
182 vx3 = _mm256_mul_ps(vx3, vbeta);
183 const __m256 ve4 = _mm256_fmadd_ps(vp4, valpha, vs4);
184 vx4 = _mm256_mul_ps(vx4, vbeta);
185 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5);
186 vx5 = _mm256_mul_ps(vx5, vbeta);
187 const __m256 ve6 = _mm256_fmadd_ps(vp6, valpha, vs6);
188 vx6 = _mm256_mul_ps(vx6, vbeta);
189 const __m256 ve7 = _mm256_fmadd_ps(vp7, valpha, vs7);
190 vx7 = _mm256_mul_ps(vx7, vbeta);
191 const __m256 ve8 = _mm256_fmadd_ps(vp8, valpha, vs8);
192 vx8 = _mm256_mul_ps(vx8, vbeta);
193
194 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
195 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
196 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
197 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
198 const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
199 const __m256 vy5 = _mm256_blendv_ps(vx5, ve5, vx5);
200 const __m256 vy6 = _mm256_blendv_ps(vx6, ve6, vx6);
201 const __m256 vy7 = _mm256_blendv_ps(vx7, ve7, vx7);
202 const __m256 vy8 = _mm256_blendv_ps(vx8, ve8, vx8);
203
204 _mm256_storeu_ps(y, vy0);
205 _mm256_storeu_ps(y + 8, vy1);
206 _mm256_storeu_ps(y + 16, vy2);
207 _mm256_storeu_ps(y + 24, vy3);
208 _mm256_storeu_ps(y + 32, vy4);
209 _mm256_storeu_ps(y + 40, vy5);
210 _mm256_storeu_ps(y + 48, vy6);
211 _mm256_storeu_ps(y + 56, vy7);
212 _mm256_storeu_ps(y + 64, vy8);
213 y += 72;
214 }
215 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
216 __m256 vx = _mm256_loadu_ps(x);
217 x += 8;
218
219 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
220
221 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
222 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
223 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
224 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
225 vn = _mm256_sub_ps(vn, vmagic_bias);
226
227 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
228
229 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
230 vp = _mm256_fmadd_ps(vp, vt, vc2);
231 vp = _mm256_mul_ps(vp, vt);
232
233 vt = _mm256_mul_ps(vt, vs);
234 vs = _mm256_fmsub_ps(vs, valpha, valpha);
235 vp = _mm256_fmadd_ps(vp, vt, vt);
236 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
237
238 vx = _mm256_mul_ps(vx, vbeta);
239 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
240
241 _mm256_storeu_ps(y, vy);
242 y += 8;
243 }
244 if XNN_UNLIKELY(n != 0) {
245 assert(n >= 1 * sizeof(float));
246 assert(n <= 7 * sizeof(float));
247 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_lut8_p4.mask_table[7] - n));
248
249 __m256 vx = _mm256_maskload_ps(x, vmask);
250
251 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
252
253 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
254 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
255 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
256 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
257 vn = _mm256_sub_ps(vn, vmagic_bias);
258
259 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
260
261 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
262 vp = _mm256_fmadd_ps(vp, vt, vc2);
263 vp = _mm256_mul_ps(vp, vt);
264
265 vt = _mm256_mul_ps(vt, vs);
266 vs = _mm256_fmsub_ps(vs, valpha, valpha);
267 vp = _mm256_fmadd_ps(vp, vt, vt);
268 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
269
270 vx = _mm256_mul_ps(vx, vbeta);
271 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
272
273 __m128 vy_lo = _mm256_castps256_ps128(vy);
274 if (n & (4 * sizeof(float))) {
275 _mm_storeu_ps(y, vy_lo);
276 vy_lo = _mm256_extractf128_ps(vy, 1);
277 y += 4;
278 }
279 if (n & (2 * sizeof(float))) {
280 _mm_storel_pi((__m64*) y, vy_lo);
281 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
282 y += 2;
283 }
284 if (n & (1 * sizeof(float))) {
285 _mm_store_ss(y, vy_lo);
286 }
287 }
288 }
289