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_x160(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x160(
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 >= 160 * sizeof(float); elements -= 160 * sizeof(float)) {
42 // Load 160 (10x16) 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 input += 160;
54
55 // Subtract maximum input x := i - i_max.
56 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
57 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
58 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
59 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
60 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
61 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
62 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
63 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
64 const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
65 const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
66
67 // Compute reduced argument elements := round(x / log(2)).
68 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
69 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
70 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
71 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
72 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
73 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
74 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
75 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
76 __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
77 __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
78
79 // Compute reduced argument t := x - elements * log(2).
80 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
81 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
82 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
83 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
84 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
85 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
86 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
87 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
88 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
89 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
90 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
91
92 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
93 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
94 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
95 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
96 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
97 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
98 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
99 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
100 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
101 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
102
103 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
104 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
105 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
106 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
107 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
108 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
109 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
110 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
111 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
112 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
113 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
114
115 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
116 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
117 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
118 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
119 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
120 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
121 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
122 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
123 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
125
126 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
127 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
128 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
129 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
130 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
131 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
132 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
133 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
134 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
136
137 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
138 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
139 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
140 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
141 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
142 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
143 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
144 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
145 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
147
148 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
149 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
150 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
151 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
152 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
153 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
154 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
155 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
156 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
157 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
158
159 // Reconstruct the final f value:
160 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
161 // = 2**elements * p
162 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
163 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
164 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
165 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
166 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
167 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
168 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
169 __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
170 __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
171 __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
172
173 // Multiply by scale.
174 vf0 = _mm512_mul_ps(vf0, vscale);
175 vf1 = _mm512_mul_ps(vf1, vscale);
176 vf2 = _mm512_mul_ps(vf2, vscale);
177 vf3 = _mm512_mul_ps(vf3, vscale);
178 vf4 = _mm512_mul_ps(vf4, vscale);
179 vf5 = _mm512_mul_ps(vf5, vscale);
180 vf6 = _mm512_mul_ps(vf6, vscale);
181 vf7 = _mm512_mul_ps(vf7, vscale);
182 vf8 = _mm512_mul_ps(vf8, vscale);
183 vf9 = _mm512_mul_ps(vf9, vscale);
184
185 // Store 160 (10x16) outputs at a time.
186 _mm512_storeu_ps(output, vf0);
187 _mm512_storeu_ps(output + 0, vf0);
188 _mm512_storeu_ps(output + 16, vf1);
189 _mm512_storeu_ps(output + 32, vf2);
190 _mm512_storeu_ps(output + 48, vf3);
191 _mm512_storeu_ps(output + 64, vf4);
192 _mm512_storeu_ps(output + 80, vf5);
193 _mm512_storeu_ps(output + 96, vf6);
194 _mm512_storeu_ps(output + 112, vf7);
195 _mm512_storeu_ps(output + 128, vf8);
196 _mm512_storeu_ps(output + 144, vf9);
197 output += 160;
198 }
199 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
200 // Load 16 inputs at a time.
201 const __m512 vi = _mm512_loadu_ps(input);
202 input += 16;
203
204 // Subtract maximum input x := i - i_max.
205 const __m512 vx = _mm512_sub_ps(vi, vi_max);
206
207 // Compute reduced argument elements := round(x / log(2)).
208 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
209
210 // Compute reduced argument t := x - elements * log(2).
211 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
212 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
213 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
214
215 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
216 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
217 vp = _mm512_fmadd_ps(vp, vt, vc3);
218 vp = _mm512_fmadd_ps(vp, vt, vc2);
219 vp = _mm512_fmadd_ps(vp, vt, vc1);
220 vp = _mm512_fmadd_ps(vp, vt, vc0);
221
222 // Reconstruct the final f value:
223 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
224 // = 2**elements * p
225 __m512 vf = _mm512_scalef_ps(vp, vn);
226
227 // Multiply by scale.
228 vf = _mm512_mul_ps(vf, vscale);
229
230 // Store 16 outputs at a time.
231 _mm512_storeu_ps(output, vf);
232 output += 16;
233 }
234 if (elements != 0) {
235 // Prepare mask for valid 32-bit elements (depends on elements).
236 elements >>= 2 /* log2(sizeof(float)) */;
237 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
238
239 // Load up to 15 inputs at a time.
240 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
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 __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 approximation 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 __m512 vf = _mm512_scalef_ps(vp, vn);
264
265 // Multiply by scale.
266 vf = _mm512_mul_ps(vf, vscale);
267
268 // Store up to 15 outputs at a time.
269 _mm512_mask_storeu_ps(output, vmask, vf);
270 }
271 }
272