1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddexpminusmax/avx512f-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/raddexpminusmax.h>
16
17
xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3(size_t elements,const float * input,float * sum,float max)18 void xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3(
19 size_t elements,
20 const float* input,
21 float* sum,
22 float max)
23 {
24 assert(elements % sizeof(float) == 0);
25
26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
29
30 const __m512 vc0 = _mm512_set1_ps(1.0f);
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
34 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
36
37 const __m512 vi_max = _mm512_set1_ps(max);
38
39 __m512 vacc0 = _mm512_setzero_ps();
40 __m512 vacc1 = _mm512_setzero_ps();
41 __m512 vacc2 = _mm512_setzero_ps();
42 for (; elements >= 192 * sizeof(float); elements -= 192 * sizeof(float)) {
43 // Load 192 (12x16) inputs at a time.
44 const __m512 vi0 = _mm512_loadu_ps(input);
45 const __m512 vi1 = _mm512_loadu_ps(input + 16);
46 const __m512 vi2 = _mm512_loadu_ps(input + 32);
47 const __m512 vi3 = _mm512_loadu_ps(input + 48);
48 const __m512 vi4 = _mm512_loadu_ps(input + 64);
49 const __m512 vi5 = _mm512_loadu_ps(input + 80);
50 const __m512 vi6 = _mm512_loadu_ps(input + 96);
51 const __m512 vi7 = _mm512_loadu_ps(input + 112);
52 const __m512 vi8 = _mm512_loadu_ps(input + 128);
53 const __m512 vi9 = _mm512_loadu_ps(input + 144);
54 const __m512 vi10 = _mm512_loadu_ps(input + 160);
55 const __m512 vi11 = _mm512_loadu_ps(input + 176);
56 input += 192;
57
58 // Subtract maximum input x := i - i_max.
59 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
60 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
61 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
62 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
63 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
64 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
65 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
66 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
67 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
68 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
69 const __m512 vx10 = _mm512_sub_ps(vi10, vi_max);
70 const __m512 vx11 = _mm512_sub_ps(vi11, vi_max);
71
72 // Compute reduced argument elements := round(x / log(2)).
73 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
74 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
75 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
76 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
77 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
78 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
79 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
80 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
81 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
82 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
83 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
84 const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);
85
86 // Compute reduced argument t := x - elements * log(2).
87 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
88 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
89 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
90 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
91 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
92 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
93 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
94 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
95 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
96 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
97 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
98 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
99 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);
100
101 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
102 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
103 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
104 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
105 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
106 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
107 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
108 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
109 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
110 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
111 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
112 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11);
113
114 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
115 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
116 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
117 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
118 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
119 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
120 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
121 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
122 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
123 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
124 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
125 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
126 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);
127
128 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
129 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
130 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
131 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
132 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
133 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
134 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
135 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
136 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
137 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
138 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
139 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3);
140
141 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
142 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
143 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
144 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
145 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
146 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
147 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
148 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
149 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
150 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
151 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
152 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2);
153
154 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
155 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
156 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
157 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
158 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
159 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
160 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
161 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
162 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
163 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
164 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
165 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1);
166
167 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
168 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
169 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
170 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
171 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
172 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
173 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
174 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
175 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
176 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
177 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
178 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0);
179
180 // Reconstruct the final f value:
181 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
182 // = 2**elements * p
183 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
184 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
185 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
186 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
187 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
188 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
189 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
190 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
191 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
192 const __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
193 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10);
194 const __m512 vf11 = _mm512_scalef_ps(vp11, vn11);
195
196 // Accumulate computed exponents.
197 vacc0 = _mm512_add_ps(vacc0, vf0);
198 vacc1 = _mm512_add_ps(vacc1, vf1);
199 vacc2 = _mm512_add_ps(vacc2, vf2);
200 vacc0 = _mm512_add_ps(vacc0, vf3);
201 vacc1 = _mm512_add_ps(vacc1, vf4);
202 vacc2 = _mm512_add_ps(vacc2, vf5);
203 vacc0 = _mm512_add_ps(vacc0, vf6);
204 vacc1 = _mm512_add_ps(vacc1, vf7);
205 vacc2 = _mm512_add_ps(vacc2, vf8);
206 vacc0 = _mm512_add_ps(vacc0, vf9);
207 vacc1 = _mm512_add_ps(vacc1, vf10);
208 vacc2 = _mm512_add_ps(vacc2, vf11);
209 }
210 // Add up all accumulators to vacc0
211 vacc0 = _mm512_add_ps(vacc0, vacc1);
212 vacc0 = _mm512_add_ps(vacc0, vacc2);
213
214 __m512 vacc = vacc0;
215 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
216 // Load 16 inputs at a time.
217 const __m512 vi = _mm512_loadu_ps(input);
218 input += 16;
219
220 // Subtract maximum input x := i - i_max.
221 const __m512 vx = _mm512_sub_ps(vi, vi_max);
222
223 // Compute reduced argument elements := round(x / log(2)).
224 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
225
226 // Compute reduced argument t := x - elements * log(2).
227 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
228 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
229 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
230
231 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
232 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
233 vp = _mm512_fmadd_ps(vp, vt, vc3);
234 vp = _mm512_fmadd_ps(vp, vt, vc2);
235 vp = _mm512_fmadd_ps(vp, vt, vc1);
236 vp = _mm512_fmadd_ps(vp, vt, vc0);
237
238 // Reconstruct the final f value:
239 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
240 // = 2**elements * p
241 const __m512 vf = _mm512_scalef_ps(vp, vn);
242
243 // Accumulate computed exponents.
244 vacc = _mm512_add_ps(vacc, vf);
245 }
246 if (elements != 0) {
247 // Prepare mask for valid 32-bit elements (depends on elements).
248 elements >>= 2 /* log2(sizeof(float)) */;
249 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
250
251 // Load up to 15 inputs at a time.
252 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
253
254 // Subtract maximum input x := i - i_max.
255 const __m512 vx = _mm512_sub_ps(vi, vi_max);
256
257 // Compute reduced argument elements := round(x / log(2)).
258 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
259
260 // Compute reduced argument t := x - elements * log(2).
261 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
262 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
263 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
264
265 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
266 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
267 vp = _mm512_fmadd_ps(vp, vt, vc3);
268 vp = _mm512_fmadd_ps(vp, vt, vc2);
269 vp = _mm512_fmadd_ps(vp, vt, vc1);
270 vp = _mm512_fmadd_ps(vp, vt, vc0);
271
272 // Reconstruct the final f value:
273 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
274 // = 2**elements * p
275 const __m512 vf = _mm512_scalef_ps(vp, vn);
276
277 // Accumulate computed exponents.
278 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
279 }
280 *sum = _mm512_reduce_add_ps(vacc);
281 }
282