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
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_velu_ukernel__avx2_rr1_p6_x56(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_velu_ukernel__avx2_rr1_p6_x56(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
29 const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
30 const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
31
32 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
33 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
36 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
37 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
38 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
39 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
41
42 for (; n >= 56 * sizeof(float); n -= 56 * sizeof(float)) {
43 __m256 vx0 = _mm256_loadu_ps(x);
44 __m256 vx1 = _mm256_loadu_ps(x + 8);
45 __m256 vx2 = _mm256_loadu_ps(x + 16);
46 __m256 vx3 = _mm256_loadu_ps(x + 24);
47 __m256 vx4 = _mm256_loadu_ps(x + 32);
48 __m256 vx5 = _mm256_loadu_ps(x + 40);
49 __m256 vx6 = _mm256_loadu_ps(x + 48);
50 x += 56;
51
52 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
53 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
54 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
55 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
56 const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
57 const __m256 vz5 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx5, vprescale));
58 const __m256 vz6 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx6, 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
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
83 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
84 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
85 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
86 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
87 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
88 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5);
89 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6);
90
91 __m256 vp0 = _mm256_fmadd_ps(vc6, vt0, vc5);
92 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
93 __m256 vp2 = _mm256_fmadd_ps(vc6, vt2, vc5);
94 __m256 vp3 = _mm256_fmadd_ps(vc6, vt3, vc5);
95 __m256 vp4 = _mm256_fmadd_ps(vc6, vt4, vc5);
96 __m256 vp5 = _mm256_fmadd_ps(vc6, vt5, vc5);
97 __m256 vp6 = _mm256_fmadd_ps(vc6, vt6, vc5);
98
99 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
100 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
101 vp2 = _mm256_fmadd_ps(vp2, vt2, vc4);
102 vp3 = _mm256_fmadd_ps(vp3, vt3, vc4);
103 vp4 = _mm256_fmadd_ps(vp4, vt4, vc4);
104 vp5 = _mm256_fmadd_ps(vp5, vt5, vc4);
105 vp6 = _mm256_fmadd_ps(vp6, vt6, vc4);
106
107 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
108 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
109 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
110 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
111 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
112 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
113 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
114
115 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
116 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
117 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
118 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
119 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
120 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
121 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
122
123 vp0 = _mm256_mul_ps(vp0, vt0);
124 vt0 = _mm256_mul_ps(vt0, vs0);
125 vp1 = _mm256_mul_ps(vp1, vt1);
126 vt1 = _mm256_mul_ps(vt1, vs1);
127 vp2 = _mm256_mul_ps(vp2, vt2);
128 vt2 = _mm256_mul_ps(vt2, vs2);
129 vp3 = _mm256_mul_ps(vp3, vt3);
130 vt3 = _mm256_mul_ps(vt3, vs3);
131 vp4 = _mm256_mul_ps(vp4, vt4);
132 vt4 = _mm256_mul_ps(vt4, vs4);
133 vp5 = _mm256_mul_ps(vp5, vt5);
134 vt5 = _mm256_mul_ps(vt5, vs5);
135 vp6 = _mm256_mul_ps(vp6, vt6);
136 vt6 = _mm256_mul_ps(vt6, vs6);
137
138 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
139 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
140 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
141 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
142 vs2 = _mm256_fmsub_ps(vs2, valpha, valpha);
143 vp2 = _mm256_fmadd_ps(vp2, vt2, vt2);
144 vs3 = _mm256_fmsub_ps(vs3, valpha, valpha);
145 vp3 = _mm256_fmadd_ps(vp3, vt3, vt3);
146 vs4 = _mm256_fmsub_ps(vs4, valpha, valpha);
147 vp4 = _mm256_fmadd_ps(vp4, vt4, vt4);
148 vs5 = _mm256_fmsub_ps(vs5, valpha, valpha);
149 vp5 = _mm256_fmadd_ps(vp5, vt5, vt5);
150 vs6 = _mm256_fmsub_ps(vs6, valpha, valpha);
151 vp6 = _mm256_fmadd_ps(vp6, vt6, vt6);
152
153 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
154 vx0 = _mm256_mul_ps(vx0, vbeta);
155 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
156 vx1 = _mm256_mul_ps(vx1, vbeta);
157 const __m256 ve2 = _mm256_fmadd_ps(vp2, valpha, vs2);
158 vx2 = _mm256_mul_ps(vx2, vbeta);
159 const __m256 ve3 = _mm256_fmadd_ps(vp3, valpha, vs3);
160 vx3 = _mm256_mul_ps(vx3, vbeta);
161 const __m256 ve4 = _mm256_fmadd_ps(vp4, valpha, vs4);
162 vx4 = _mm256_mul_ps(vx4, vbeta);
163 const __m256 ve5 = _mm256_fmadd_ps(vp5, valpha, vs5);
164 vx5 = _mm256_mul_ps(vx5, vbeta);
165 const __m256 ve6 = _mm256_fmadd_ps(vp6, valpha, vs6);
166 vx6 = _mm256_mul_ps(vx6, vbeta);
167
168 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
169 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
170 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
171 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
172 const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
173 const __m256 vy5 = _mm256_blendv_ps(vx5, ve5, vx5);
174 const __m256 vy6 = _mm256_blendv_ps(vx6, ve6, vx6);
175
176 _mm256_storeu_ps(y, vy0);
177 _mm256_storeu_ps(y + 8, vy1);
178 _mm256_storeu_ps(y + 16, vy2);
179 _mm256_storeu_ps(y + 24, vy3);
180 _mm256_storeu_ps(y + 32, vy4);
181 _mm256_storeu_ps(y + 40, vy5);
182 _mm256_storeu_ps(y + 48, vy6);
183 y += 56;
184 }
185 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
186 __m256 vx = _mm256_loadu_ps(x);
187 x += 8;
188
189 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
190
191 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
192 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
193 vn = _mm256_sub_ps(vn, vmagic_bias);
194
195 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
196
197 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
198 vp = _mm256_fmadd_ps(vp, vt, vc4);
199 vp = _mm256_fmadd_ps(vp, vt, vc3);
200 vp = _mm256_fmadd_ps(vp, vt, vc2);
201 vp = _mm256_mul_ps(vp, vt);
202
203 vt = _mm256_mul_ps(vt, vs);
204 vs = _mm256_fmsub_ps(vs, valpha, valpha);
205 vp = _mm256_fmadd_ps(vp, vt, vt);
206 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
207
208 vx = _mm256_mul_ps(vx, vbeta);
209 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
210
211 _mm256_storeu_ps(y, vy);
212 y += 8;
213 }
214 if XNN_UNLIKELY(n != 0) {
215 assert(n >= 1 * sizeof(float));
216 assert(n <= 7 * sizeof(float));
217 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
218
219 __m256 vx = _mm256_maskload_ps(x, vmask);
220
221 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
222
223 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
224 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
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(vc6, vt, vc5);
230 vp = _mm256_fmadd_ps(vp, vt, vc4);
231 vp = _mm256_fmadd_ps(vp, vt, vc3);
232 vp = _mm256_fmadd_ps(vp, vt, vc2);
233 vp = _mm256_mul_ps(vp, vt);
234
235 vt = _mm256_mul_ps(vt, vs);
236 vs = _mm256_fmsub_ps(vs, valpha, valpha);
237 vp = _mm256_fmadd_ps(vp, vt, vt);
238 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
239
240 vx = _mm256_mul_ps(vx, vbeta);
241 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
242
243 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
244 __m128 vy_lo = _mm256_castps256_ps128(vy);
245 if (n & (4 * sizeof(float))) {
246 _mm_storeu_ps(y, vy_lo);
247 vy_lo = _mm256_extractf128_ps(vy, 1);
248 y += 4;
249 }
250 if (n & (2 * sizeof(float))) {
251 _mm_storel_pi((__m64*) y, vy_lo);
252 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
253 y += 2;
254 }
255 if (n & (1 * sizeof(float))) {
256 _mm_store_ss(y, vy_lo);
257 }
258 }
259 }
260