Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u176.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleexpminusmax/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/vscaleexpminusmax.h"
16
17
18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_u176(
19 size_t batch,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
24 {
25 assert(batch != 0);
26 assert(batch % sizeof(float) == 0);
27 assert(input != NULL);
28 assert(output != NULL);
29
30 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
31 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
32 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
33
34 const __m512 vc0 = _mm512_set1_ps(1.0f);
35 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
36 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
37 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
38 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
39 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
40
41 const __m512 vscale = _mm512_set1_ps(scale);
42 const __m512 vi_max = _mm512_set1_ps(max);
43
44 for (; batch >= 176 * sizeof(float); batch -= 176 * sizeof(float)) {
45 // Load 176 (11x16) inputs at a time.
46 const __m512 vi0 = _mm512_loadu_ps(input);
47 const __m512 vi1 = _mm512_loadu_ps(input + 16);
48 const __m512 vi2 = _mm512_loadu_ps(input + 32);
49 const __m512 vi3 = _mm512_loadu_ps(input + 48);
50 const __m512 vi4 = _mm512_loadu_ps(input + 64);
51 const __m512 vi5 = _mm512_loadu_ps(input + 80);
52 const __m512 vi6 = _mm512_loadu_ps(input + 96);
53 const __m512 vi7 = _mm512_loadu_ps(input + 112);
54 const __m512 vi8 = _mm512_loadu_ps(input + 128);
55 const __m512 vi9 = _mm512_loadu_ps(input + 144);
56 const __m512 vi10 = _mm512_loadu_ps(input + 160);
57 input += 176;
58
59 // Subtract maximum input x := i - i_max.
60 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
61 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
62 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
63 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
64 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
65 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
66 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
67 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
68 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
69 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
70 const __m512 vx10 = _mm512_sub_ps(vi10, vi_max);
71
72 // Compute reduced argument batch := round(x / log(2)).
73 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
74 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
75 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
76 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
77 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
78 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
79 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
80 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
81 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
82 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
83 __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
84
85 // Compute reduced argument t := x - batch * log(2).
86 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
87 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
88 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
89 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
90 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
91 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
92 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
93 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
94 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
95 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
96 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
97 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
98
99 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
100 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
101 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
102 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
103 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
104 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
105 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
106 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
107 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
108 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
109 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
110
111 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
112 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
113 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
114 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
115 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
116 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
117 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
118 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
119 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
120 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
121 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
122 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
123
124 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
125 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
126 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
127 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
128 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
129 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
130 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
131 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
132 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
133 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
134 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
135
136 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
137 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
138 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
139 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
140 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
141 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
142 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
143 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
144 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
145 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
146 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
147
148 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
149 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
150 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
151 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
152 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
153 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
154 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
155 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
156 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
157 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
158 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
159
160 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
161 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
162 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
163 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
164 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
165 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
166 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
167 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
168 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
169 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
170 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
171
172 // Reconstruct the final f value:
173 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
174 // = 2**batch * p
175 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
176 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
177 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
178 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
179 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
180 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
181 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
182 __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
183 __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
184 __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
185 __m512 vf10 = _mm512_scalef_ps(vp10, vn10);
186
187 // Multiply by scale.
188 vf0 = _mm512_mul_ps(vf0, vscale);
189 vf1 = _mm512_mul_ps(vf1, vscale);
190 vf2 = _mm512_mul_ps(vf2, vscale);
191 vf3 = _mm512_mul_ps(vf3, vscale);
192 vf4 = _mm512_mul_ps(vf4, vscale);
193 vf5 = _mm512_mul_ps(vf5, vscale);
194 vf6 = _mm512_mul_ps(vf6, vscale);
195 vf7 = _mm512_mul_ps(vf7, vscale);
196 vf8 = _mm512_mul_ps(vf8, vscale);
197 vf9 = _mm512_mul_ps(vf9, vscale);
198 vf10 = _mm512_mul_ps(vf10, vscale);
199
200 // Store 176 (11x16) outputs at a time.
201 _mm512_storeu_ps(output, vf0);
202 _mm512_storeu_ps(output + 0, vf0);
203 _mm512_storeu_ps(output + 16, vf1);
204 _mm512_storeu_ps(output + 32, vf2);
205 _mm512_storeu_ps(output + 48, vf3);
206 _mm512_storeu_ps(output + 64, vf4);
207 _mm512_storeu_ps(output + 80, vf5);
208 _mm512_storeu_ps(output + 96, vf6);
209 _mm512_storeu_ps(output + 112, vf7);
210 _mm512_storeu_ps(output + 128, vf8);
211 _mm512_storeu_ps(output + 144, vf9);
212 _mm512_storeu_ps(output + 160, vf10);
213 output += 176;
214 }
215 for (; batch >= 16 * sizeof(float); batch -= 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 batch := round(x / log(2)).
224 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
225
226 // Compute reduced argument t := x - batch * 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 approximation 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**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
240 // = 2**batch * p
241 __m512 vf = _mm512_scalef_ps(vp, vn);
242
243 // Multiply by scale.
244 vf = _mm512_mul_ps(vf, vscale);
245
246 // Store 16 outputs at a time.
247 _mm512_storeu_ps(output, vf);
248 output += 16;
249 }
250 if (batch != 0) {
251 // Prepare mask for valid 32-bit batch (depends on batch).
252 batch >>= XNN_LOG2_SIZEOF_FLOAT;
253 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
254
255 // Load up to 15 inputs at a time.
256 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
257
258 // Subtract maximum input x := i - i_max.
259 const __m512 vx = _mm512_sub_ps(vi, vi_max);
260
261 // Compute reduced argument batch := round(x / log(2)).
262 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
263
264 // Compute reduced argument t := x - batch * log(2).
265 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
266 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
267 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
268
269 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
270 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
271 vp = _mm512_fmadd_ps(vp, vt, vc3);
272 vp = _mm512_fmadd_ps(vp, vt, vc2);
273 vp = _mm512_fmadd_ps(vp, vt, vc1);
274 vp = _mm512_fmadd_ps(vp, vt, vc0);
275
276 // Reconstruct the final f value:
277 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
278 // = 2**batch * p
279 __m512 vf = _mm512_scalef_ps(vp, vn);
280
281 // Multiply by scale.
282 vf = _mm512_mul_ps(vf, vscale);
283
284 // Store up to 15 outputs at a time.
285 _mm512_mask_storeu_ps(output, vmask, vf);
286 }
287 }