1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-rr2-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__avx_rr2_p6_x40(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__avx_rr2_p6_x40(
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->avx_rr2_p6.prescale);
27 const __m256 valpha = _mm256_load_ps(params->avx_rr2_p6.alpha);
28 const __m256 vbeta = _mm256_load_ps(params->avx_rr2_p6.beta);
29 const __m256 vsat_cutoff = _mm256_load_ps(params->avx_rr2_p6.sat_cutoff);
30 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p6.magic_bias);
31 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p6.log2e);
32 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p6.minus_ln2_hi);
33 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p6.minus_ln2_lo);
34 const __m256 vc6 = _mm256_load_ps(params->avx_rr2_p6.c6);
35 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p6.c5);
36 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p6.c4);
37 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p6.c3);
38 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p6.c2);
39 const __m256 vone = _mm256_load_ps(params->avx_rr2_p6.one);
40
41 for (; n >= 40 * sizeof(float); n -= 40 * sizeof(float)) {
42 __m256 vx0 = _mm256_loadu_ps(x);
43 __m256 vx1 = _mm256_loadu_ps(x + 8);
44 __m256 vx2 = _mm256_loadu_ps(x + 16);
45 __m256 vx3 = _mm256_loadu_ps(x + 24);
46 __m256 vx4 = _mm256_loadu_ps(x + 32);
47 x += 40;
48
49 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
50 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
51 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
52 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
53 const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
54
55 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
56 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
57 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
58 __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
59 __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);
60
61 const __m128 vs0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
62 const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
63 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
64 const __m128 vs1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
65 const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
66 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
67 const __m128 vs2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
68 const __m128 vs2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
69 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
70 const __m128 vs3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)), 23));
71 const __m128 vs3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, 1)), 23));
72 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
73 const __m128 vs4_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)), 23));
74 const __m128 vs4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, 1)), 23));
75 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
76
77 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
78 __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
79 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
80 __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
81 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
82 __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1);
83 __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
84 __m256 vs3 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs3_lo), vs3_hi, 1);
85 __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_hi), vz4);
86 __m256 vs4 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs4_lo), vs4_hi, 1);
87
88 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
89 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
90 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
91 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
92 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_lo), vt4);
93
94 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc6, vt0), vc5);
95 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc6, vt1), vc5);
96 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc6, vt2), vc5);
97 __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc6, vt3), vc5);
98 __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc6, vt4), vc5);
99
100 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc4);
101 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc4);
102 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc4);
103 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc4);
104 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc4);
105
106 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
107 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
108 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
109 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc3);
110 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc3);
111
112 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
113 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
114 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
115 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
116 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc2);
117
118 vp0 = _mm256_mul_ps(vp0, vt0);
119 vp1 = _mm256_mul_ps(vp1, vt1);
120 vp2 = _mm256_mul_ps(vp2, vt2);
121 vp3 = _mm256_mul_ps(vp3, vt3);
122 vp4 = _mm256_mul_ps(vp4, vt4);
123
124 vt0 = _mm256_mul_ps(vt0, vs0);
125 vs0 = _mm256_sub_ps(vs0, vone);
126 vt1 = _mm256_mul_ps(vt1, vs1);
127 vs1 = _mm256_sub_ps(vs1, vone);
128 vt2 = _mm256_mul_ps(vt2, vs2);
129 vs2 = _mm256_sub_ps(vs2, vone);
130 vt3 = _mm256_mul_ps(vt3, vs3);
131 vs3 = _mm256_sub_ps(vs3, vone);
132 vt4 = _mm256_mul_ps(vt4, vs4);
133 vs4 = _mm256_sub_ps(vs4, vone);
134
135 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
136 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
137 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
138 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
139 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vt4);
140
141 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
142 vx0 = _mm256_mul_ps(vx0, vbeta);
143 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
144 vx1 = _mm256_mul_ps(vx1, vbeta);
145 const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
146 vx2 = _mm256_mul_ps(vx2, vbeta);
147 const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
148 vx3 = _mm256_mul_ps(vx3, vbeta);
149 const __m256 ve4 = _mm256_mul_ps(_mm256_add_ps(vp4, vs4), valpha);
150 vx4 = _mm256_mul_ps(vx4, vbeta);
151
152 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
153 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
154 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
155 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
156 const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
157
158 _mm256_storeu_ps(y, vy0);
159 _mm256_storeu_ps(y + 8, vy1);
160 _mm256_storeu_ps(y + 16, vy2);
161 _mm256_storeu_ps(y + 24, vy3);
162 _mm256_storeu_ps(y + 32, vy4);
163 y += 40;
164 }
165 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
166 __m256 vx = _mm256_loadu_ps(x);
167 x += 8;
168
169 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
170
171 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
172 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
173 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
174 vn = _mm256_sub_ps(vn, vmagic_bias);
175
176 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
177 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
178 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
179
180 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
181 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
182 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
183 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
184 vp = _mm256_mul_ps(vp, vt);
185
186 vt = _mm256_mul_ps(vt, vs);
187 vs = _mm256_sub_ps(vs, vone);
188 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
189
190 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
191 vx = _mm256_mul_ps(vx, vbeta);
192 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
193
194 _mm256_storeu_ps(y, vy);
195 y += 8;
196 }
197 if XNN_UNLIKELY(n != 0) {
198 assert(n >= 1 * sizeof(float));
199 assert(n <= 7 * sizeof(float));
200 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p6.mask_table[7] - n));
201
202 __m256 vx = _mm256_maskload_ps(x, vmask);
203
204 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
205
206 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
207 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
208 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
209 vn = _mm256_sub_ps(vn, vmagic_bias);
210
211 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
212 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
213 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
214
215 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
216 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
217 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
218 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
219 vp = _mm256_mul_ps(vp, vt);
220
221 vt = _mm256_mul_ps(vt, vs);
222 vs = _mm256_sub_ps(vs, vone);
223 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
224
225 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
226 vx = _mm256_mul_ps(vx, vbeta);
227 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
228
229 __m128 vy_lo = _mm256_castps256_ps128(vy);
230 if (n & (4 * sizeof(float))) {
231 _mm_storeu_ps(y, vy_lo);
232 vy_lo = _mm256_extractf128_ps(vy, 1);
233 y += 4;
234 }
235 if (n & (2 * sizeof(float))) {
236 _mm_storel_pi((__m64*) y, vy_lo);
237 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
238 y += 2;
239 }
240 if (n & (1 * sizeof(float))) {
241 _mm_store_ss(y, vy_lo);
242 }
243 }
244 }
245