1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/avx2-rr1-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/raddstoreexpminusmax.h>
15
16
xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x80_acc2(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_raddstoreexpminusmax_ukernel__avx2_rr1_p5_x80_acc2(
18 size_t elements,
19 const float* input,
20 const float* max,
21 float* output,
22 float* sum,
23 const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(elements % sizeof(float) == 0);
26
27 const __m256 vi_max = _mm256_broadcast_ss(max);
28 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
29 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
30 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
31 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
32 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
33 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
34 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
35 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
36 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);
37
38 __m256 vacc0 = _mm256_setzero_ps();
39 __m256 vacc1 = _mm256_setzero_ps();
40 for (; elements >= 80 * sizeof(float); elements -= 80 * sizeof(float)) {
41 const __m256 vi0 = _mm256_loadu_ps(input);
42 const __m256 vi1 = _mm256_loadu_ps(input + 8);
43 const __m256 vi2 = _mm256_loadu_ps(input + 16);
44 const __m256 vi3 = _mm256_loadu_ps(input + 24);
45 const __m256 vi4 = _mm256_loadu_ps(input + 32);
46 const __m256 vi5 = _mm256_loadu_ps(input + 40);
47 const __m256 vi6 = _mm256_loadu_ps(input + 48);
48 const __m256 vi7 = _mm256_loadu_ps(input + 56);
49 const __m256 vi8 = _mm256_loadu_ps(input + 64);
50 const __m256 vi9 = _mm256_loadu_ps(input + 72);
51 input += 80;
52
53 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
54 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
55 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
56 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
57 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
58 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
59 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
60 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
61 const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
62 const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
63
64 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
65 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
66 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
67 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
68 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
69 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
70 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
71 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
72 __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
73 __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
74
75 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
76 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
77 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
78 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
79 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
80 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
81 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
82 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
83 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
84 const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
85
86 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
87 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
88 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
89 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
90 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
91 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
92 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
93 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
94 vn8 = _mm256_sub_ps(vn8, vmagic_bias);
95 vn9 = _mm256_sub_ps(vn9, vmagic_bias);
96
97 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vx0);
98 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vx1);
99 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vx2);
100 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vx3);
101 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vx4);
102 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vx5);
103 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vx6);
104 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vx7);
105 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2, vx8);
106 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vx9);
107
108 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
109 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
110 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
111 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
112 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
113 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
114 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
115 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
116 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
117 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
118
119 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
120 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
121 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
122 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
123 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
124 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
125 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
126 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
127 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
128 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
129
130 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
131 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
132 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
133 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
134 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
135 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
136 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
137 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
138 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
139 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
140
141 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
142 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
143 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
144 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
145 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
146 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
147 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
148 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
149 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
150 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
151
152 vt0 = _mm256_mul_ps(vt0, vs0);
153 vt1 = _mm256_mul_ps(vt1, vs1);
154 vt2 = _mm256_mul_ps(vt2, vs2);
155 vt3 = _mm256_mul_ps(vt3, vs3);
156 vt4 = _mm256_mul_ps(vt4, vs4);
157 vt5 = _mm256_mul_ps(vt5, vs5);
158 vt6 = _mm256_mul_ps(vt6, vs6);
159 vt7 = _mm256_mul_ps(vt7, vs7);
160 vt8 = _mm256_mul_ps(vt8, vs8);
161 vt9 = _mm256_mul_ps(vt9, vs9);
162
163 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
164 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
165 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
166 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
167 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
168 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
169 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
170 __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
171 __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
172 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
173
174 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
175 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
176 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
177 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
178 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
179 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
180 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
181 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
182 vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
183 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
184
185 _mm256_storeu_ps(output, vf0);
186 _mm256_storeu_ps(output + 8, vf1);
187 _mm256_storeu_ps(output + 16, vf2);
188 _mm256_storeu_ps(output + 24, vf3);
189 _mm256_storeu_ps(output + 32, vf4);
190 _mm256_storeu_ps(output + 40, vf5);
191 _mm256_storeu_ps(output + 48, vf6);
192 _mm256_storeu_ps(output + 56, vf7);
193 _mm256_storeu_ps(output + 64, vf8);
194 _mm256_storeu_ps(output + 72, vf9);
195 output += 80;
196
197 vacc0 = _mm256_add_ps(vacc0, vf0);
198 vacc1 = _mm256_add_ps(vacc1, vf1);
199 vacc0 = _mm256_add_ps(vacc0, vf2);
200 vacc1 = _mm256_add_ps(vacc1, vf3);
201 vacc0 = _mm256_add_ps(vacc0, vf4);
202 vacc1 = _mm256_add_ps(vacc1, vf5);
203 vacc0 = _mm256_add_ps(vacc0, vf6);
204 vacc1 = _mm256_add_ps(vacc1, vf7);
205 vacc0 = _mm256_add_ps(vacc0, vf8);
206 vacc1 = _mm256_add_ps(vacc1, vf9);
207 }
208 vacc0 = _mm256_add_ps(vacc0, vacc1);
209
210 __m256 vacc = vacc0;
211 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
212 const __m256 vi = _mm256_loadu_ps(input);
213 input += 8;
214
215 const __m256 vx = _mm256_sub_ps(vi, vi_max);
216
217 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
218
219 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
220
221 vn = _mm256_sub_ps(vn, vmagic_bias);
222
223 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vx);
224
225 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
226 vp = _mm256_fmadd_ps(vp, vt, vc3);
227 vp = _mm256_fmadd_ps(vp, vt, vc2);
228 vp = _mm256_fmadd_ps(vp, vt, vc1);
229
230 vt = _mm256_mul_ps(vt, vs);
231 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
232
233 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
234
235 _mm256_storeu_ps(output, vf);
236 output += 8;
237
238 vacc = _mm256_add_ps(vacc, vf);
239 }
240 if (elements != 0) {
241 assert(elements >= 1 * sizeof(float));
242 assert(elements <= 7 * sizeof(float));
243 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - elements));
244
245 const __m256 vi = _mm256_maskload_ps(input, vmask);
246
247 const __m256 vx = _mm256_sub_ps(vi, vi_max);
248
249 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
250
251 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
252
253 vn = _mm256_sub_ps(vn, vmagic_bias);
254
255 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vx);
256
257 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
258 vp = _mm256_fmadd_ps(vp, vt, vc3);
259 vp = _mm256_fmadd_ps(vp, vt, vc2);
260 vp = _mm256_fmadd_ps(vp, vt, vc1);
261
262 vt = _mm256_mul_ps(vt, vs);
263 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
264
265 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
266
267 __m128 vf_lo = _mm256_castps256_ps128(vf);
268 if (elements & (4 * sizeof(float))) {
269 _mm_storeu_ps(output, vf_lo);
270 vf_lo = _mm256_extractf128_ps(vf, 1);
271 output += 4;
272 }
273 if (elements & (2 * sizeof(float))) {
274 _mm_storel_pi((__m64*) output, vf_lo);
275 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
276 output += 2;
277 }
278 if (elements & (1 * sizeof(float))) {
279 _mm_store_ss(output, vf_lo);
280 }
281
282 vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
283 }
284 __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
285 vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
286 vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
287 _mm_store_ss(sum, vacc_lo);
288 _mm256_zeroupper();
289 }
290