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
xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x192(
19 size_t elements,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
24 {
25 assert(elements % sizeof(float) == 0);
26
27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
28 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
30
31 const __m512 vc0 = _mm512_set1_ps(1.0f);
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
34 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
35 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
37
38 const __m512 vscale = _mm512_set1_ps(scale);
39 const __m512 vi_max = _mm512_set1_ps(max);
40
41 for (; elements >= 192 * sizeof(float); elements -= 192 * sizeof(float)) {
42 // Load 192 (12x16) inputs at a time.
43 const __m512 vi0 = _mm512_loadu_ps(input);
44 const __m512 vi1 = _mm512_loadu_ps(input + 16);
45 const __m512 vi2 = _mm512_loadu_ps(input + 32);
46 const __m512 vi3 = _mm512_loadu_ps(input + 48);
47 const __m512 vi4 = _mm512_loadu_ps(input + 64);
48 const __m512 vi5 = _mm512_loadu_ps(input + 80);
49 const __m512 vi6 = _mm512_loadu_ps(input + 96);
50 const __m512 vi7 = _mm512_loadu_ps(input + 112);
51 const __m512 vi8 = _mm512_loadu_ps(input + 128);
52 const __m512 vi9 = _mm512_loadu_ps(input + 144);
53 const __m512 vi10 = _mm512_loadu_ps(input + 160);
54 const __m512 vi11 = _mm512_loadu_ps(input + 176);
55 input += 192;
56
57 // Subtract maximum input x := i - i_max.
58 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
59 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
60 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
61 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
62 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
63 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
64 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
65 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
66 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
67 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
68 const __m512 vx10 = _mm512_sub_ps(vi10, vi_max);
69 const __m512 vx11 = _mm512_sub_ps(vi11, vi_max);
70
71 // Compute reduced argument elements := round(x / log(2)).
72 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
73 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
74 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
75 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
76 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
77 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
78 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
79 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
80 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
81 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
82 __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
83 __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);
84
85 // Compute reduced argument t := x - elements * 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 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);
99
100 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
101 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
102 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
103 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
104 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
105 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
106 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
107 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
108 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
109 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
110 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
111 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11);
112
113 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
114 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
115 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
116 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
117 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
118 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
119 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
120 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
121 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
122 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
123 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
124 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
125 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);
126
127 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
128 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
129 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
130 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
131 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
132 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
133 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
134 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
135 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
136 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3);
139
140 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
141 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
142 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
143 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
144 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
145 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
146 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
147 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
148 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
149 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2);
152
153 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
154 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
155 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
156 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
157 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
158 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
159 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
160 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
161 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
162 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1);
165
166 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
167 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
168 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
169 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
170 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
171 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
172 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
173 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
174 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
175 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
176 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
177 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0);
178
179 // Reconstruct the final f value:
180 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
181 // = 2**elements * p
182 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
183 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
184 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
185 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
186 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
187 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
188 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
189 __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
190 __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
191 __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
192 __m512 vf10 = _mm512_scalef_ps(vp10, vn10);
193 __m512 vf11 = _mm512_scalef_ps(vp11, vn11);
194
195 // Multiply by scale.
196 vf0 = _mm512_mul_ps(vf0, vscale);
197 vf1 = _mm512_mul_ps(vf1, vscale);
198 vf2 = _mm512_mul_ps(vf2, vscale);
199 vf3 = _mm512_mul_ps(vf3, vscale);
200 vf4 = _mm512_mul_ps(vf4, vscale);
201 vf5 = _mm512_mul_ps(vf5, vscale);
202 vf6 = _mm512_mul_ps(vf6, vscale);
203 vf7 = _mm512_mul_ps(vf7, vscale);
204 vf8 = _mm512_mul_ps(vf8, vscale);
205 vf9 = _mm512_mul_ps(vf9, vscale);
206 vf10 = _mm512_mul_ps(vf10, vscale);
207 vf11 = _mm512_mul_ps(vf11, vscale);
208
209 // Store 192 (12x16) outputs at a time.
210 _mm512_storeu_ps(output, vf0);
211 _mm512_storeu_ps(output + 0, vf0);
212 _mm512_storeu_ps(output + 16, vf1);
213 _mm512_storeu_ps(output + 32, vf2);
214 _mm512_storeu_ps(output + 48, vf3);
215 _mm512_storeu_ps(output + 64, vf4);
216 _mm512_storeu_ps(output + 80, vf5);
217 _mm512_storeu_ps(output + 96, vf6);
218 _mm512_storeu_ps(output + 112, vf7);
219 _mm512_storeu_ps(output + 128, vf8);
220 _mm512_storeu_ps(output + 144, vf9);
221 _mm512_storeu_ps(output + 160, vf10);
222 _mm512_storeu_ps(output + 176, vf11);
223 output += 192;
224 }
225 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
226 // Load 16 inputs at a time.
227 const __m512 vi = _mm512_loadu_ps(input);
228 input += 16;
229
230 // Subtract maximum input x := i - i_max.
231 const __m512 vx = _mm512_sub_ps(vi, vi_max);
232
233 // Compute reduced argument elements := round(x / log(2)).
234 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
235
236 // Compute reduced argument t := x - elements * log(2).
237 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
238 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
239 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
240
241 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
242 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
243 vp = _mm512_fmadd_ps(vp, vt, vc3);
244 vp = _mm512_fmadd_ps(vp, vt, vc2);
245 vp = _mm512_fmadd_ps(vp, vt, vc1);
246 vp = _mm512_fmadd_ps(vp, vt, vc0);
247
248 // Reconstruct the final f value:
249 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
250 // = 2**elements * p
251 __m512 vf = _mm512_scalef_ps(vp, vn);
252
253 // Multiply by scale.
254 vf = _mm512_mul_ps(vf, vscale);
255
256 // Store 16 outputs at a time.
257 _mm512_storeu_ps(output, vf);
258 output += 16;
259 }
260 if (elements != 0) {
261 // Prepare mask for valid 32-bit elements (depends on elements).
262 elements >>= 2 /* log2(sizeof(float)) */;
263 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
264
265 // Load up to 15 inputs at a time.
266 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
267
268 // Subtract maximum input x := i - i_max.
269 const __m512 vx = _mm512_sub_ps(vi, vi_max);
270
271 // Compute reduced argument elements := round(x / log(2)).
272 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
273
274 // Compute reduced argument t := x - elements * log(2).
275 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
276 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
277 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
278
279 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
280 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
281 vp = _mm512_fmadd_ps(vp, vt, vc3);
282 vp = _mm512_fmadd_ps(vp, vt, vc2);
283 vp = _mm512_fmadd_ps(vp, vt, vc1);
284 vp = _mm512_fmadd_ps(vp, vt, vc0);
285
286 // Reconstruct the final f value:
287 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
288 // = 2**elements * p
289 __m512 vf = _mm512_scalef_ps(vp, vn);
290
291 // Multiply by scale.
292 vf = _mm512_mul_ps(vf, vscale);
293
294 // Store up to 15 outputs at a time.
295 _mm512_mask_storeu_ps(output, vmask, vf);
296 }
297 }
298