1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-rr2-lut4-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
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__avx_rr2_lut4_p4_perm_x40(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__avx_rr2_lut4_p4_perm_x40(
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(0x1.8003F8p21f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x3));
36 const __m256 vtable = _mm256_set_ps(
37 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f,
38 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f);
39 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
40 const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
41 const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
42 const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
43 const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
44 const __m256 vone = _mm256_set1_ps(1.0f);
45
46 for (; n >= 40 * sizeof(float); n -= 40 * sizeof(float)) {
47 __m256 vx0 = _mm256_loadu_ps(x);
48 __m256 vx1 = _mm256_loadu_ps(x + 8);
49 __m256 vx2 = _mm256_loadu_ps(x + 16);
50 __m256 vx3 = _mm256_loadu_ps(x + 24);
51 __m256 vx4 = _mm256_loadu_ps(x + 32);
52 x += 40;
53
54 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
55 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
56 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
57 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
58 const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
59
60 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
61 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
62 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
63 __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
64 __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);
65
66 __m256 ven0 = _mm256_andnot_ps(vindex_mask, vn0);
67 const __m256 vl0 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0));
68 const __m128 ven0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven0)), 21));
69 __m256 ven1 = _mm256_andnot_ps(vindex_mask, vn1);
70 const __m256 vl1 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1));
71 const __m128 ven1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven1)), 21));
72 __m256 ven2 = _mm256_andnot_ps(vindex_mask, vn2);
73 const __m256 vl2 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn2));
74 const __m128 ven2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven2)), 21));
75 __m256 ven3 = _mm256_andnot_ps(vindex_mask, vn3);
76 const __m256 vl3 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn3));
77 const __m128 ven3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven3)), 21));
78 __m256 ven4 = _mm256_andnot_ps(vindex_mask, vn4);
79 const __m256 vl4 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn4));
80 const __m128 ven4_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven4)), 21));
81
82 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
83 const __m128 ven0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven0, 1)), 21));
84 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
85 const __m128 ven1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven1, 1)), 21));
86 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
87 const __m128 ven2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven2, 1)), 21));
88 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
89 const __m128 ven3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven3, 1)), 21));
90 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
91 const __m128 ven4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven4, 1)), 21));
92
93 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
94 ven0 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven0_lo), ven0_hi, 1);
95 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
96 ven1 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven1_lo), ven1_hi, 1);
97 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
98 ven2 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven2_lo), ven2_hi, 1);
99 __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
100 ven3 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven3_lo), ven3_hi, 1);
101 __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_hi), vz4);
102 ven4 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven4_lo), ven4_hi, 1);
103
104 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
105 __m256 vs0 = _mm256_mul_ps(vl0, ven0);
106 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
107 __m256 vs1 = _mm256_mul_ps(vl1, ven1);
108 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
109 __m256 vs2 = _mm256_mul_ps(vl2, ven2);
110 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
111 __m256 vs3 = _mm256_mul_ps(vl3, ven3);
112 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_lo), vt4);
113 __m256 vs4 = _mm256_mul_ps(vl4, ven4);
114
115 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc4, vt0), vc3);
116 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc4, vt1), vc3);
117 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc4, vt2), vc3);
118 __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc4, vt3), vc3);
119 __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc4, vt4), vc3);
120
121 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
122 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
123 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
124 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
125 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc2);
126
127 vp0 = _mm256_mul_ps(vp0, vt0);
128 vp1 = _mm256_mul_ps(vp1, vt1);
129 vp2 = _mm256_mul_ps(vp2, vt2);
130 vp3 = _mm256_mul_ps(vp3, vt3);
131 vp4 = _mm256_mul_ps(vp4, vt4);
132
133 vt0 = _mm256_mul_ps(vt0, vs0);
134 vs0 = _mm256_sub_ps(vs0, vone);
135 vt1 = _mm256_mul_ps(vt1, vs1);
136 vs1 = _mm256_sub_ps(vs1, vone);
137 vt2 = _mm256_mul_ps(vt2, vs2);
138 vs2 = _mm256_sub_ps(vs2, vone);
139 vt3 = _mm256_mul_ps(vt3, vs3);
140 vs3 = _mm256_sub_ps(vs3, vone);
141 vt4 = _mm256_mul_ps(vt4, vs4);
142 vs4 = _mm256_sub_ps(vs4, vone);
143
144 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
145 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
146 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
147 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
148 vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vt4);
149
150 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
151 vx0 = _mm256_mul_ps(vx0, vbeta);
152 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
153 vx1 = _mm256_mul_ps(vx1, vbeta);
154 const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
155 vx2 = _mm256_mul_ps(vx2, vbeta);
156 const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
157 vx3 = _mm256_mul_ps(vx3, vbeta);
158 const __m256 ve4 = _mm256_mul_ps(_mm256_add_ps(vp4, vs4), valpha);
159 vx4 = _mm256_mul_ps(vx4, vbeta);
160
161 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
162 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
163 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
164 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
165 const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
166
167 _mm256_storeu_ps(y, vy0);
168 _mm256_storeu_ps(y + 8, vy1);
169 _mm256_storeu_ps(y + 16, vy2);
170 _mm256_storeu_ps(y + 24, vy3);
171 _mm256_storeu_ps(y + 32, vy4);
172 y += 40;
173 }
174 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
175 __m256 vx = _mm256_loadu_ps(x);
176 x += 8;
177
178 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
179
180 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
181 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
182 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
183 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
184 vn = _mm256_sub_ps(vn, vmagic_bias);
185 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
186
187 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
188 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
189 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
190 __m256 vs = _mm256_mul_ps(vl, ven);
191
192 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
193 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
194 vp = _mm256_mul_ps(vp, vt);
195
196 vt = _mm256_mul_ps(vt, vs);
197 vs = _mm256_sub_ps(vs, vone);
198 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
199
200 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
201 vx = _mm256_mul_ps(vx, vbeta);
202 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
203
204 _mm256_storeu_ps(y, vy);
205 y += 8;
206 }
207 if XNN_UNLIKELY(n != 0) {
208 assert(n >= 1 * sizeof(float));
209 assert(n <= 7 * sizeof(float));
210 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
211
212 __m256 vx = _mm256_maskload_ps(x, vmask);
213
214 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
215
216 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
217 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
218 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
219 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
220 vn = _mm256_sub_ps(vn, vmagic_bias);
221 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
222
223 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
224 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
225 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
226 __m256 vs = _mm256_mul_ps(vl, ven);
227
228 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
229 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
230 vp = _mm256_mul_ps(vp, vt);
231
232 vt = _mm256_mul_ps(vt, vs);
233 vs = _mm256_sub_ps(vs, vone);
234 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
235
236 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
237 vx = _mm256_mul_ps(vx, vbeta);
238 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
239
240 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
241 __m128 vy_lo = _mm256_castps256_ps128(vy);
242 if (n & (4 * sizeof(float))) {
243 _mm_storeu_ps(y, vy_lo);
244 vy_lo = _mm256_extractf128_ps(vy, 1);
245 y += 4;
246 }
247 if (n & (2 * sizeof(float))) {
248 _mm_storel_pi((__m64*) y, vy_lo);
249 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
250 y += 2;
251 }
252 if (n & (1 * sizeof(float))) {
253 _mm_store_ss(y, vy_lo);
254 }
255 }
256 }
257