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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * 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 x += 32;
52
53 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
54 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
55 const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
56 const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
57
58 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
59 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
60 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
61 __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
62
63 __m256 ven0 = _mm256_andnot_ps(vindex_mask, vn0);
64 const __m256 vl0 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0));
65 const __m128 ven0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven0)), 21));
66 __m256 ven1 = _mm256_andnot_ps(vindex_mask, vn1);
67 const __m256 vl1 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1));
68 const __m128 ven1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven1)), 21));
69 __m256 ven2 = _mm256_andnot_ps(vindex_mask, vn2);
70 const __m256 vl2 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn2));
71 const __m128 ven2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven2)), 21));
72 __m256 ven3 = _mm256_andnot_ps(vindex_mask, vn3);
73 const __m256 vl3 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn3));
74 const __m128 ven3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven3)), 21));
75
76 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
77 const __m128 ven0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven0, 1)), 21));
78 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
79 const __m128 ven1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven1, 1)), 21));
80 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
81 const __m128 ven2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven2, 1)), 21));
82 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
83 const __m128 ven3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven3, 1)), 21));
84
85 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
86 ven0 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven0_lo), ven0_hi, 1);
87 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
88 ven1 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven1_lo), ven1_hi, 1);
89 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
90 ven2 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven2_lo), ven2_hi, 1);
91 __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
92 ven3 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven3_lo), ven3_hi, 1);
93
94 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
95 __m256 vs0 = _mm256_mul_ps(vl0, ven0);
96 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
97 __m256 vs1 = _mm256_mul_ps(vl1, ven1);
98 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
99 __m256 vs2 = _mm256_mul_ps(vl2, ven2);
100 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
101 __m256 vs3 = _mm256_mul_ps(vl3, ven3);
102
103 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc4, vt0), vc3);
104 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc4, vt1), vc3);
105 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc4, vt2), vc3);
106 __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc4, vt3), vc3);
107
108 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
109 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
110 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
111 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
112
113 vp0 = _mm256_mul_ps(vp0, vt0);
114 vp1 = _mm256_mul_ps(vp1, vt1);
115 vp2 = _mm256_mul_ps(vp2, vt2);
116 vp3 = _mm256_mul_ps(vp3, vt3);
117
118 vt0 = _mm256_mul_ps(vt0, vs0);
119 vs0 = _mm256_sub_ps(vs0, vone);
120 vt1 = _mm256_mul_ps(vt1, vs1);
121 vs1 = _mm256_sub_ps(vs1, vone);
122 vt2 = _mm256_mul_ps(vt2, vs2);
123 vs2 = _mm256_sub_ps(vs2, vone);
124 vt3 = _mm256_mul_ps(vt3, vs3);
125 vs3 = _mm256_sub_ps(vs3, vone);
126
127 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
128 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
129 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
130 vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
131
132 const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
133 vx0 = _mm256_mul_ps(vx0, vbeta);
134 const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
135 vx1 = _mm256_mul_ps(vx1, vbeta);
136 const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
137 vx2 = _mm256_mul_ps(vx2, vbeta);
138 const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
139 vx3 = _mm256_mul_ps(vx3, vbeta);
140
141 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
142 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
143 const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
144 const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
145
146 _mm256_storeu_ps(y, vy0);
147 _mm256_storeu_ps(y + 8, vy1);
148 _mm256_storeu_ps(y + 16, vy2);
149 _mm256_storeu_ps(y + 24, vy3);
150 y += 32;
151 }
152 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
153 __m256 vx = _mm256_loadu_ps(x);
154 x += 8;
155
156 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
157
158 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
159 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
160 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
161 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
162 vn = _mm256_sub_ps(vn, vmagic_bias);
163 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
164
165 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
166 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
167 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
168 __m256 vs = _mm256_mul_ps(vl, ven);
169
170 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
171 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
172 vp = _mm256_mul_ps(vp, vt);
173
174 vt = _mm256_mul_ps(vt, vs);
175 vs = _mm256_sub_ps(vs, vone);
176 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
177
178 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
179 vx = _mm256_mul_ps(vx, vbeta);
180 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
181
182 _mm256_storeu_ps(y, vy);
183 y += 8;
184 }
185 if XNN_UNLIKELY(n != 0) {
186 assert(n >= 1 * sizeof(float));
187 assert(n <= 7 * sizeof(float));
188 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
189
190 __m256 vx = _mm256_maskload_ps(x, vmask);
191
192 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
193
194 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
195 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
196 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
197 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
198 vn = _mm256_sub_ps(vn, vmagic_bias);
199 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
200
201 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
202 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
203 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
204 __m256 vs = _mm256_mul_ps(vl, ven);
205
206 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
207 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
208 vp = _mm256_mul_ps(vp, vt);
209
210 vt = _mm256_mul_ps(vt, vs);
211 vs = _mm256_sub_ps(vs, vone);
212 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
213
214 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
215 vx = _mm256_mul_ps(vx, vbeta);
216 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
217
218 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
219 __m128 vy_lo = _mm256_castps256_ps128(vy);
220 if (n & (4 * sizeof(float))) {
221 _mm_storeu_ps(y, vy_lo);
222 vy_lo = _mm256_extractf128_ps(vy, 1);
223 y += 4;
224 }
225 if (n & (2 * sizeof(float))) {
226 _mm_storel_pi((__m64*) y, vy_lo);
227 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
228 y += 2;
229 }
230 if (n & (1 * sizeof(float))) {
231 _mm_store_ss(y, vy_lo);
232 }
233 }
234 }
235