1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/avx512f-rr1-p5-scalef.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/intrinsics-polyfill.h>
15 #include <xnnpack/raddstoreexpminusmax.h>
16
17
xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x144_acc3(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)])18 void xnn_f32_raddstoreexpminusmax_ukernel__avx512f_rr1_p5_scalef_x144_acc3(
19 size_t elements,
20 const float* input,
21 const float* max,
22 float* output,
23 float* sum,
24 const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(elements % sizeof(float) == 0);
27
28 const __m512 vi_max = _mm512_set1_ps(*max);
29 const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_p5.log2e);
30 const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_p5.minus_ln2);
31 const __m512 vc5 = _mm512_set1_ps(params->avx512_rr1_p5.c5);
32 const __m512 vc4 = _mm512_set1_ps(params->avx512_rr1_p5.c4);
33 const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_p5.c3);
34 const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_p5.c2);
35 const __m512 vc1 = _mm512_set1_ps(params->avx512_rr1_p5.c1);
36 const __m512 vc0 = _mm512_set1_ps(params->avx512_rr1_p5.c0);
37
38 __m512 vacc0 = _mm512_setzero_ps();
39 __m512 vacc1 = _mm512_setzero_ps();
40 __m512 vacc2 = _mm512_setzero_ps();
41 for (; elements >= 144 * sizeof(float); elements -= 144 * sizeof(float)) {
42 const __m512 vi0 = _mm512_loadu_ps(input);
43 const __m512 vi1 = _mm512_loadu_ps(input + 16);
44 const __m512 vi2 = _mm512_loadu_ps(input + 32);
45 const __m512 vi3 = _mm512_loadu_ps(input + 48);
46 const __m512 vi4 = _mm512_loadu_ps(input + 64);
47 const __m512 vi5 = _mm512_loadu_ps(input + 80);
48 const __m512 vi6 = _mm512_loadu_ps(input + 96);
49 const __m512 vi7 = _mm512_loadu_ps(input + 112);
50 const __m512 vi8 = _mm512_loadu_ps(input + 128);
51 input += 144;
52
53 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
54 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
55 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
56 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
57 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
58 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
59 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
60 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
61 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
62
63 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
64 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
65 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
66 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
67 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
68 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
69 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
70 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
71 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
72
73 const __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vx0);
74 const __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vx1);
75 const __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vx2);
76 const __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vx3);
77 const __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vx4);
78 const __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vx5);
79 const __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vx6);
80 const __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2, vx7);
81 const __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2, vx8);
82
83 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
84 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
85 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
86 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
87 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
88 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
89 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
90 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
91 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
92
93 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
94 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
95 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
96 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
97 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
98 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
99 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
100 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
101 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
102
103 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
104 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
105 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
106 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
107 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
108 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
110 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
111 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
112
113 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
114 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
115 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
116 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
117 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
118 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
119 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
120 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
121 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
122
123 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
124 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
125 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
126 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
127 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
128 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
129 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
130 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
131 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
132
133 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
134 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
135 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
136 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
137 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
138 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
139 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
140 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
141 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
142
143 _mm512_storeu_ps(output, vf0);
144 _mm512_storeu_ps(output + 16, vf1);
145 _mm512_storeu_ps(output + 32, vf2);
146 _mm512_storeu_ps(output + 48, vf3);
147 _mm512_storeu_ps(output + 64, vf4);
148 _mm512_storeu_ps(output + 80, vf5);
149 _mm512_storeu_ps(output + 96, vf6);
150 _mm512_storeu_ps(output + 112, vf7);
151 _mm512_storeu_ps(output + 128, vf8);
152 output += 144;
153
154 vacc0 = _mm512_add_ps(vacc0, vf0);
155 vacc1 = _mm512_add_ps(vacc1, vf1);
156 vacc2 = _mm512_add_ps(vacc2, vf2);
157 vacc0 = _mm512_add_ps(vacc0, vf3);
158 vacc1 = _mm512_add_ps(vacc1, vf4);
159 vacc2 = _mm512_add_ps(vacc2, vf5);
160 vacc0 = _mm512_add_ps(vacc0, vf6);
161 vacc1 = _mm512_add_ps(vacc1, vf7);
162 vacc2 = _mm512_add_ps(vacc2, vf8);
163 }
164 vacc0 = _mm512_add_ps(vacc0, vacc1);
165 vacc0 = _mm512_add_ps(vacc0, vacc2);
166
167 __m512 vacc = vacc0;
168 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
169 const __m512 vi = _mm512_loadu_ps(input);
170 input += 16;
171
172 const __m512 vx = _mm512_sub_ps(vi, vi_max);
173
174 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
175
176 const __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vx);
177
178 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
179 vp = _mm512_fmadd_ps(vp, vt, vc3);
180 vp = _mm512_fmadd_ps(vp, vt, vc2);
181 vp = _mm512_fmadd_ps(vp, vt, vc1);
182 vp = _mm512_fmadd_ps(vp, vt, vc0);
183
184 const __m512 vf = _mm512_scalef_ps(vp, vn);
185
186 _mm512_storeu_ps(output, vf);
187 output += 16;
188
189 vacc = _mm512_add_ps(vacc, vf);
190 }
191 if (elements != 0) {
192 // Prepare mask for valid 32-bit elements (depends on elements).
193 elements >>= 2 /* log2(sizeof(float)) */;
194 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
195
196 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
197
198 const __m512 vx = _mm512_sub_ps(vi, vi_max);
199
200 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
201
202 const __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vx);
203
204 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
205 vp = _mm512_fmadd_ps(vp, vt, vc3);
206 vp = _mm512_fmadd_ps(vp, vt, vc2);
207 vp = _mm512_fmadd_ps(vp, vt, vc1);
208 vp = _mm512_fmadd_ps(vp, vt, vc0);
209
210 const __m512 vf = _mm512_scalef_ps(vp, vn);
211
212 _mm512_mask_storeu_ps(output, vmask, vf);
213
214 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
215 }
216 *sum = _mm512_reduce_add_ps(vacc);
217 }
218