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_acc3(size_t elements,const float * input,float * output,float * sum,float max)18 void xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc3(
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 for (; elements >= 192 * sizeof(float); elements -= 192 * sizeof(float)) {
44 // Load 192 (12x16) inputs at a time.
45 const __m512 vi0 = _mm512_loadu_ps(input);
46 const __m512 vi1 = _mm512_loadu_ps(input + 16);
47 const __m512 vi2 = _mm512_loadu_ps(input + 32);
48 const __m512 vi3 = _mm512_loadu_ps(input + 48);
49 const __m512 vi4 = _mm512_loadu_ps(input + 64);
50 const __m512 vi5 = _mm512_loadu_ps(input + 80);
51 const __m512 vi6 = _mm512_loadu_ps(input + 96);
52 const __m512 vi7 = _mm512_loadu_ps(input + 112);
53 const __m512 vi8 = _mm512_loadu_ps(input + 128);
54 const __m512 vi9 = _mm512_loadu_ps(input + 144);
55 const __m512 vi10 = _mm512_loadu_ps(input + 160);
56 const __m512 vi11 = _mm512_loadu_ps(input + 176);
57 input += 192;
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 const __m512 vx11 = _mm512_sub_ps(vi11, vi_max);
72
73 // Compute reduced argument elements := round(x / log(2)).
74 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
75 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
76 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
77 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
78 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
79 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
80 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
81 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
82 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
83 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
84 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
85 const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);
86
87 // Compute reduced argument t := x - elements * log(2).
88 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
89 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
90 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
91 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
92 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
93 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
94 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
95 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
96 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
97 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
98 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
99 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
100 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);
101
102 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
103 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
104 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
105 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
106 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
107 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
108 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
109 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
110 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
111 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
112 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
113 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11);
114
115 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
116 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
117 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
118 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
119 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
120 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
121 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
122 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
123 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
124 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
125 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
126 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
127 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);
128
129 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
130 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
131 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
132 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
133 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
134 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
136 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
137 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
138 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
139 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
140 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3);
141
142 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
143 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
144 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
145 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
146 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
147 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
148 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
149 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
150 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
151 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
152 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
153 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2);
154
155 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
156 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
157 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
158 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
159 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
160 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
161 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
162 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
163 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
164 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
165 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
166 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1);
167
168 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
169 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
170 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
171 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
172 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
173 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
174 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
175 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
176 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
177 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
178 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
179 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0);
180
181 // Reconstruct the final f value:
182 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
183 // = 2**elements * p
184 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
185 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
186 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
187 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
188 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
189 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
190 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
191 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
192 const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
193 const __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
194 const __m512 vf10 = _mm512_scalef_ps(vp10, vn10);
195 const __m512 vf11 = _mm512_scalef_ps(vp11, vn11);
196
197 // Store 192 (12x16) outputs at a time.
198 _mm512_storeu_ps(output, vf0);
199 _mm512_storeu_ps(output + 16, vf1);
200 _mm512_storeu_ps(output + 32, vf2);
201 _mm512_storeu_ps(output + 48, vf3);
202 _mm512_storeu_ps(output + 64, vf4);
203 _mm512_storeu_ps(output + 80, vf5);
204 _mm512_storeu_ps(output + 96, vf6);
205 _mm512_storeu_ps(output + 112, vf7);
206 _mm512_storeu_ps(output + 128, vf8);
207 _mm512_storeu_ps(output + 144, vf9);
208 _mm512_storeu_ps(output + 160, vf10);
209 _mm512_storeu_ps(output + 176, vf11);
210 output += 192;
211
212 // Accumulate computed exponents.
213 vacc0 = _mm512_add_ps(vacc0, vf0);
214 vacc1 = _mm512_add_ps(vacc1, vf1);
215 vacc2 = _mm512_add_ps(vacc2, vf2);
216 vacc0 = _mm512_add_ps(vacc0, vf3);
217 vacc1 = _mm512_add_ps(vacc1, vf4);
218 vacc2 = _mm512_add_ps(vacc2, vf5);
219 vacc0 = _mm512_add_ps(vacc0, vf6);
220 vacc1 = _mm512_add_ps(vacc1, vf7);
221 vacc2 = _mm512_add_ps(vacc2, vf8);
222 vacc0 = _mm512_add_ps(vacc0, vf9);
223 vacc1 = _mm512_add_ps(vacc1, vf10);
224 vacc2 = _mm512_add_ps(vacc2, vf11);
225 }
226 // Add up all accumulators to vacc0
227 vacc0 = _mm512_add_ps(vacc0, vacc1);
228 vacc0 = _mm512_add_ps(vacc0, vacc2);
229
230 __m512 vacc = vacc0;
231 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
232 // Load 16 inputs at a time.
233 const __m512 vi = _mm512_loadu_ps(input);
234 input += 16;
235
236 // Subtract maximum input x := i - i_max.
237 const __m512 vx = _mm512_sub_ps(vi, vi_max);
238
239 // Compute reduced argument elements := round(x / log(2)).
240 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
241
242 // Compute reduced argument t := x - elements * log(2).
243 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
244 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
245 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
246
247 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
248 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
249 vp = _mm512_fmadd_ps(vp, vt, vc3);
250 vp = _mm512_fmadd_ps(vp, vt, vc2);
251 vp = _mm512_fmadd_ps(vp, vt, vc1);
252 vp = _mm512_fmadd_ps(vp, vt, vc0);
253
254 // Reconstruct the final f value:
255 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
256 // = 2**elements * p
257 const __m512 vf = _mm512_scalef_ps(vp, vn);
258
259 // Store 16 outputs at a time.
260 _mm512_storeu_ps(output, vf);
261 output += 16;
262
263 // Accumulate computed exponents.
264 vacc = _mm512_add_ps(vacc, vf);
265 }
266 if (elements != 0) {
267 // Prepare mask for valid 32-bit elements (depends on elements).
268 elements >>= 2 /* log2(sizeof(float)) */;
269 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
270
271 // Load up to 15 inputs at a time.
272 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
273
274 // Subtract maximum input x := i - i_max.
275 const __m512 vx = _mm512_sub_ps(vi, vi_max);
276
277 // Compute reduced argument elements := round(x / log(2)).
278 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
279
280 // Compute reduced argument t := x - elements * log(2).
281 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
282 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
283 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
284
285 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
286 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
287 vp = _mm512_fmadd_ps(vp, vt, vc3);
288 vp = _mm512_fmadd_ps(vp, vt, vc2);
289 vp = _mm512_fmadd_ps(vp, vt, vc1);
290 vp = _mm512_fmadd_ps(vp, vt, vc0);
291
292 // Reconstruct the final f value:
293 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
294 // = 2**elements * p
295 const __m512 vf = _mm512_scalef_ps(vp, vn);
296
297 // Store up to 15 outputs at a time.
298 _mm512_mask_storeu_ps(output, vmask, vf);
299
300 // Accumulate computed exponents.
301 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
302 }
303 *sum = _mm512_reduce_add_ps(vacc);
304 }
305