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_x48(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_x48(
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 >= 48 * sizeof(float); n -= 48 * 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     __m256 vx5 = _mm256_loadu_ps(x + 40);
53     x += 48;
54 
55     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
56     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
57     const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
58     const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
59     const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));
60     const __m256 vz5 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx5, vprescale));
61 
62     __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
63     __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
64     __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
65     __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
66     __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);
67     __m256 vn5 = _mm256_add_ps(_mm256_mul_ps(vz5, vlog2e), vmagic_bias);
68 
69     __m256 ven0 = _mm256_andnot_ps(vindex_mask, vn0);
70     const __m256 vl0 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn0));
71     const __m128 ven0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven0)), 21));
72     __m256 ven1 = _mm256_andnot_ps(vindex_mask, vn1);
73     const __m256 vl1 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn1));
74     const __m128 ven1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven1)), 21));
75     __m256 ven2 = _mm256_andnot_ps(vindex_mask, vn2);
76     const __m256 vl2 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn2));
77     const __m128 ven2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven2)), 21));
78     __m256 ven3 = _mm256_andnot_ps(vindex_mask, vn3);
79     const __m256 vl3 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn3));
80     const __m128 ven3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven3)), 21));
81     __m256 ven4 = _mm256_andnot_ps(vindex_mask, vn4);
82     const __m256 vl4 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn4));
83     const __m128 ven4_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven4)), 21));
84     __m256 ven5 = _mm256_andnot_ps(vindex_mask, vn5);
85     const __m256 vl5 = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn5));
86     const __m128 ven5_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven5)), 21));
87 
88     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
89     const __m128 ven0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven0, 1)), 21));
90     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
91     const __m128 ven1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven1, 1)), 21));
92     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
93     const __m128 ven2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven2, 1)), 21));
94     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
95     const __m128 ven3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven3, 1)), 21));
96     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
97     const __m128 ven4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven4, 1)), 21));
98     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
99     const __m128 ven5_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven5, 1)), 21));
100 
101     __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
102     ven0 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven0_lo), ven0_hi, 1);
103     __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
104     ven1 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven1_lo), ven1_hi, 1);
105     __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
106     ven2 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven2_lo), ven2_hi, 1);
107     __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
108     ven3 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven3_lo), ven3_hi, 1);
109     __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_hi), vz4);
110     ven4 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven4_lo), ven4_hi, 1);
111     __m256 vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_hi), vz5);
112     ven5 = _mm256_insertf128_ps(_mm256_castps128_ps256(ven5_lo), ven5_hi, 1);
113 
114     vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
115     __m256 vs0 = _mm256_mul_ps(vl0, ven0);
116     vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
117     __m256 vs1 = _mm256_mul_ps(vl1, ven1);
118     vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
119     __m256 vs2 = _mm256_mul_ps(vl2, ven2);
120     vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
121     __m256 vs3 = _mm256_mul_ps(vl3, ven3);
122     vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_lo), vt4);
123     __m256 vs4 = _mm256_mul_ps(vl4, ven4);
124     vt5 = _mm256_add_ps(_mm256_mul_ps(vn5, vminus_ln2_lo), vt5);
125     __m256 vs5 = _mm256_mul_ps(vl5, ven5);
126 
127     __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc4, vt0), vc3);
128     __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc4, vt1), vc3);
129     __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc4, vt2), vc3);
130     __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc4, vt3), vc3);
131     __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc4, vt4), vc3);
132     __m256 vp5 = _mm256_add_ps(_mm256_mul_ps(vc4, vt5), vc3);
133 
134     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
135     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
136     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
137     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
138     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc2);
139     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vc2);
140 
141     vp0 = _mm256_mul_ps(vp0, vt0);
142     vp1 = _mm256_mul_ps(vp1, vt1);
143     vp2 = _mm256_mul_ps(vp2, vt2);
144     vp3 = _mm256_mul_ps(vp3, vt3);
145     vp4 = _mm256_mul_ps(vp4, vt4);
146     vp5 = _mm256_mul_ps(vp5, vt5);
147 
148     vt0 = _mm256_mul_ps(vt0, vs0);
149     vs0 = _mm256_sub_ps(vs0, vone);
150     vt1 = _mm256_mul_ps(vt1, vs1);
151     vs1 = _mm256_sub_ps(vs1, vone);
152     vt2 = _mm256_mul_ps(vt2, vs2);
153     vs2 = _mm256_sub_ps(vs2, vone);
154     vt3 = _mm256_mul_ps(vt3, vs3);
155     vs3 = _mm256_sub_ps(vs3, vone);
156     vt4 = _mm256_mul_ps(vt4, vs4);
157     vs4 = _mm256_sub_ps(vs4, vone);
158     vt5 = _mm256_mul_ps(vt5, vs5);
159     vs5 = _mm256_sub_ps(vs5, vone);
160 
161     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
162     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
163     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
164     vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
165     vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vt4);
166     vp5 = _mm256_add_ps(_mm256_mul_ps(vp5, vt5), vt5);
167 
168     const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
169     vx0 = _mm256_mul_ps(vx0, vbeta);
170     const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
171     vx1 = _mm256_mul_ps(vx1, vbeta);
172     const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
173     vx2 = _mm256_mul_ps(vx2, vbeta);
174     const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
175     vx3 = _mm256_mul_ps(vx3, vbeta);
176     const __m256 ve4 = _mm256_mul_ps(_mm256_add_ps(vp4, vs4), valpha);
177     vx4 = _mm256_mul_ps(vx4, vbeta);
178     const __m256 ve5 = _mm256_mul_ps(_mm256_add_ps(vp5, vs5), valpha);
179     vx5 = _mm256_mul_ps(vx5, vbeta);
180 
181     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
182     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
183     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
184     const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
185     const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);
186     const __m256 vy5 = _mm256_blendv_ps(vx5, ve5, vx5);
187 
188     _mm256_storeu_ps(y, vy0);
189     _mm256_storeu_ps(y + 8, vy1);
190     _mm256_storeu_ps(y + 16, vy2);
191     _mm256_storeu_ps(y + 24, vy3);
192     _mm256_storeu_ps(y + 32, vy4);
193     _mm256_storeu_ps(y + 40, vy5);
194     y += 48;
195   }
196   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
197     __m256 vx = _mm256_loadu_ps(x);
198     x += 8;
199 
200     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
201 
202     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
203     __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
204     const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
205     const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
206     vn = _mm256_sub_ps(vn, vmagic_bias);
207     const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
208 
209     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
210     ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
211     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
212     __m256 vs = _mm256_mul_ps(vl, ven);
213 
214     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
215     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
216     vp = _mm256_mul_ps(vp, vt);
217 
218     vt = _mm256_mul_ps(vt, vs);
219     vs = _mm256_sub_ps(vs, vone);
220     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
221 
222     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
223     vx = _mm256_mul_ps(vx, vbeta);
224     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
225 
226     _mm256_storeu_ps(y, vy);
227     y += 8;
228   }
229   if XNN_UNLIKELY(n != 0) {
230     assert(n >= 1 * sizeof(float));
231     assert(n <= 7 * sizeof(float));
232     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
233 
234     __m256 vx = _mm256_maskload_ps(x, vmask);
235 
236     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
237 
238     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
239     __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
240     const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
241     const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
242     vn = _mm256_sub_ps(vn, vmagic_bias);
243     const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
244 
245     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
246     ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
247     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
248     __m256 vs = _mm256_mul_ps(vl, ven);
249 
250     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
251     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
252     vp = _mm256_mul_ps(vp, vt);
253 
254     vt = _mm256_mul_ps(vt, vs);
255     vs = _mm256_sub_ps(vs, vone);
256     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
257 
258     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
259     vx = _mm256_mul_ps(vx, vbeta);
260     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
261 
262     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
263     __m128 vy_lo = _mm256_castps256_ps128(vy);
264     if (n & (4 * sizeof(float))) {
265       _mm_storeu_ps(y, vy_lo);
266       vy_lo = _mm256_extractf128_ps(vy, 1);
267       y += 4;
268     }
269     if (n & (2 * sizeof(float))) {
270       _mm_storel_pi((__m64*) y, vy_lo);
271       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
272       y += 2;
273     }
274     if (n & (1 * sizeof(float))) {
275       _mm_store_ss(y, vy_lo);
276     }
277   }
278 }
279