1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/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/raddstoreexpminusmax.h>
16
17
xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6(size_t elements,const float * input,float * output,float * sum,float max)18 void xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc6(
19 size_t elements,
20 const float* input,
21 float* output,
22 float* sum,
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 vi_max = _mm512_set1_ps(max);
39
40 __m512 vacc0 = _mm512_setzero_ps();
41 __m512 vacc1 = _mm512_setzero_ps();
42 __m512 vacc2 = _mm512_setzero_ps();
43 __m512 vacc3 = _mm512_setzero_ps();
44 __m512 vacc4 = _mm512_setzero_ps();
45 __m512 vacc5 = _mm512_setzero_ps();
46 for (; elements >= 192 * sizeof(float); elements -= 192 * sizeof(float)) {
47 // Load 192 (12x16) inputs at a time.
48 const __m512 vi0 = _mm512_loadu_ps(input);
49 const __m512 vi1 = _mm512_loadu_ps(input + 16);
50 const __m512 vi2 = _mm512_loadu_ps(input + 32);
51 const __m512 vi3 = _mm512_loadu_ps(input + 48);
52 const __m512 vi4 = _mm512_loadu_ps(input + 64);
53 const __m512 vi5 = _mm512_loadu_ps(input + 80);
54 const __m512 vi6 = _mm512_loadu_ps(input + 96);
55 const __m512 vi7 = _mm512_loadu_ps(input + 112);
56 const __m512 vi8 = _mm512_loadu_ps(input + 128);
57 const __m512 vi9 = _mm512_loadu_ps(input + 144);
58 const __m512 vi10 = _mm512_loadu_ps(input + 160);
59 const __m512 vi11 = _mm512_loadu_ps(input + 176);
60 input += 192;
61
62 // Subtract maximum input x := i - i_max.
63 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
64 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
65 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
66 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
67 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
68 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
69 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
70 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
71 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
72 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
73 const __m512 vx10 = _mm512_sub_ps(vi10, vi_max);
74 const __m512 vx11 = _mm512_sub_ps(vi11, vi_max);
75
76 // Compute reduced argument elements := round(x / log(2)).
77 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
78 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
79 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
80 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
81 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
82 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
83 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
84 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
85 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
86 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
87 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
88 const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);
89
90 // Compute reduced argument t := x - elements * log(2).
91 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
92 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
93 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
94 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
95 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
96 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
97 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
98 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
99 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
100 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
101 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
102 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
103 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);
104
105 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
106 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
107 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
108 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
109 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
110 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
111 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
112 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
113 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
114 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
115 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
116 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11);
117
118 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
119 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
120 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
121 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
122 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
123 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
124 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
125 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
126 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
127 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
128 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
129 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
130 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);
131
132 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
133 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
134 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
135 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
136 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
137 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
138 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
139 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
140 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
141 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
142 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
143 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3);
144
145 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
146 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
147 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
148 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
149 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
150 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
151 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
152 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
153 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
154 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
155 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
156 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2);
157
158 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
159 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
160 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
161 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
162 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
163 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
164 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
165 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
166 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
167 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
168 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
169 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1);
170
171 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
172 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
173 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
174 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
175 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
176 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
177 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
178 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
179 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
180 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
181 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
182 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0);
183
184 // Reconstruct the final f value:
185 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
186 // = 2**elements * p
187 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
188 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
189 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
190 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
191 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
192 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
193 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
194 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
195 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
196 const __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
197 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10);
198 const __m512 vf11 = _mm512_scalef_ps(vp11, vn11);
199
200 // Store 192 (12x16) outputs at a time.
201 _mm512_storeu_ps(output, vf0);
202 _mm512_storeu_ps(output + 16, vf1);
203 _mm512_storeu_ps(output + 32, vf2);
204 _mm512_storeu_ps(output + 48, vf3);
205 _mm512_storeu_ps(output + 64, vf4);
206 _mm512_storeu_ps(output + 80, vf5);
207 _mm512_storeu_ps(output + 96, vf6);
208 _mm512_storeu_ps(output + 112, vf7);
209 _mm512_storeu_ps(output + 128, vf8);
210 _mm512_storeu_ps(output + 144, vf9);
211 _mm512_storeu_ps(output + 160, vf10);
212 _mm512_storeu_ps(output + 176, vf11);
213 output += 192;
214
215 // Accumulate computed exponents.
216 vacc0 = _mm512_add_ps(vacc0, vf0);
217 vacc1 = _mm512_add_ps(vacc1, vf1);
218 vacc2 = _mm512_add_ps(vacc2, vf2);
219 vacc3 = _mm512_add_ps(vacc3, vf3);
220 vacc4 = _mm512_add_ps(vacc4, vf4);
221 vacc5 = _mm512_add_ps(vacc5, vf5);
222 vacc0 = _mm512_add_ps(vacc0, vf6);
223 vacc1 = _mm512_add_ps(vacc1, vf7);
224 vacc2 = _mm512_add_ps(vacc2, vf8);
225 vacc3 = _mm512_add_ps(vacc3, vf9);
226 vacc4 = _mm512_add_ps(vacc4, vf10);
227 vacc5 = _mm512_add_ps(vacc5, vf11);
228 }
229 // Add up all accumulators to vacc0
230 vacc0 = _mm512_add_ps(vacc0, vacc1);
231 vacc2 = _mm512_add_ps(vacc2, vacc3);
232 vacc4 = _mm512_add_ps(vacc4, vacc5);
233 vacc0 = _mm512_add_ps(vacc0, vacc2);
234 vacc0 = _mm512_add_ps(vacc0, vacc4);
235
236 __m512 vacc = vacc0;
237 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
238 // Load 16 inputs at a time.
239 const __m512 vi = _mm512_loadu_ps(input);
240 input += 16;
241
242 // Subtract maximum input x := i - i_max.
243 const __m512 vx = _mm512_sub_ps(vi, vi_max);
244
245 // Compute reduced argument elements := round(x / log(2)).
246 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
247
248 // Compute reduced argument t := x - elements * log(2).
249 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
250 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
251 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
252
253 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
254 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
255 vp = _mm512_fmadd_ps(vp, vt, vc3);
256 vp = _mm512_fmadd_ps(vp, vt, vc2);
257 vp = _mm512_fmadd_ps(vp, vt, vc1);
258 vp = _mm512_fmadd_ps(vp, vt, vc0);
259
260 // Reconstruct the final f value:
261 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
262 // = 2**elements * p
263 const __m512 vf = _mm512_scalef_ps(vp, vn);
264
265 // Store 16 outputs at a time.
266 _mm512_storeu_ps(output, vf);
267 output += 16;
268
269 // Accumulate computed exponents.
270 vacc = _mm512_add_ps(vacc, vf);
271 }
272 if (elements != 0) {
273 // Prepare mask for valid 32-bit elements (depends on elements).
274 elements >>= 2 /* log2(sizeof(float)) */;
275 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
276
277 // Load up to 15 inputs at a time.
278 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
279
280 // Subtract maximum input x := i - i_max.
281 const __m512 vx = _mm512_sub_ps(vi, vi_max);
282
283 // Compute reduced argument elements := round(x / log(2)).
284 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
285
286 // Compute reduced argument t := x - elements * log(2).
287 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
288 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
289 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
290
291 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
292 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
293 vp = _mm512_fmadd_ps(vp, vt, vc3);
294 vp = _mm512_fmadd_ps(vp, vt, vc2);
295 vp = _mm512_fmadd_ps(vp, vt, vc1);
296 vp = _mm512_fmadd_ps(vp, vt, vc0);
297
298 // Reconstruct the final f value:
299 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
300 // = 2**elements * p
301 const __m512 vf = _mm512_scalef_ps(vp, vn);
302
303 // Store up to 15 outputs at a time.
304 _mm512_mask_storeu_ps(output, vmask, vf);
305
306 // Accumulate computed exponents.
307 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
308 }
309 *sum = _mm512_reduce_add_ps(vacc);
310 }
311