Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u128.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_u128(
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 >= 128 * sizeof(float); batch -= 128 * sizeof(float)) {
45 // Load 128 (8x16) 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 input += 128;
55
56 // Subtract maximum input x := i - i_max.
57 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
58 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
59 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
60 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
61 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
62 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
63 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
64 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
65
66 // Compute reduced argument batch := round(x / log(2)).
67 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
68 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
69 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
70 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
71 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
72 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
73 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
74 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
75
76 // Compute reduced argument t := x - batch * log(2).
77 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
78 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
79 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
80 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
81 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
82 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
83 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
84 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
85 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
86
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
88 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
89 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
90 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
91 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
92 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
94 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
95
96 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
97 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
98 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
99 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
100 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
101 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
102 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
103 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
104 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
105
106 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
107 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
108 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
109 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
110 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
111 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
112 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
113 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
114
115 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
116 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
117 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
118 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
119 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
120 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
121 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
122 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
123
124 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
125 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
126 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
127 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
128 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
129 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
130 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
131 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
132
133 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
134 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
135 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
136 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
137 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
138 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
139 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
140 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
141
142 // Reconstruct the final f value:
143 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
144 // = 2**batch * p
145 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
146 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
147 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
148 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
149 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
150 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
151 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
152 __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
153
154 // Multiply by scale.
155 vf0 = _mm512_mul_ps(vf0, vscale);
156 vf1 = _mm512_mul_ps(vf1, vscale);
157 vf2 = _mm512_mul_ps(vf2, vscale);
158 vf3 = _mm512_mul_ps(vf3, vscale);
159 vf4 = _mm512_mul_ps(vf4, vscale);
160 vf5 = _mm512_mul_ps(vf5, vscale);
161 vf6 = _mm512_mul_ps(vf6, vscale);
162 vf7 = _mm512_mul_ps(vf7, vscale);
163
164 // Store 128 (8x16) outputs at a time.
165 _mm512_storeu_ps(output, vf0);
166 _mm512_storeu_ps(output + 0, vf0);
167 _mm512_storeu_ps(output + 16, vf1);
168 _mm512_storeu_ps(output + 32, vf2);
169 _mm512_storeu_ps(output + 48, vf3);
170 _mm512_storeu_ps(output + 64, vf4);
171 _mm512_storeu_ps(output + 80, vf5);
172 _mm512_storeu_ps(output + 96, vf6);
173 _mm512_storeu_ps(output + 112, vf7);
174 output += 128;
175 }
176 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
177 // Load 16 inputs at a time.
178 const __m512 vi = _mm512_loadu_ps(input);
179 input += 16;
180
181 // Subtract maximum input x := i - i_max.
182 const __m512 vx = _mm512_sub_ps(vi, vi_max);
183
184 // Compute reduced argument batch := round(x / log(2)).
185 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
186
187 // Compute reduced argument t := x - batch * log(2).
188 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
189 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
190 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
191
192 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
193 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
194 vp = _mm512_fmadd_ps(vp, vt, vc3);
195 vp = _mm512_fmadd_ps(vp, vt, vc2);
196 vp = _mm512_fmadd_ps(vp, vt, vc1);
197 vp = _mm512_fmadd_ps(vp, vt, vc0);
198
199 // Reconstruct the final f value:
200 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
201 // = 2**batch * p
202 __m512 vf = _mm512_scalef_ps(vp, vn);
203
204 // Multiply by scale.
205 vf = _mm512_mul_ps(vf, vscale);
206
207 // Store 16 outputs at a time.
208 _mm512_storeu_ps(output, vf);
209 output += 16;
210 }
211 if (batch != 0) {
212 // Prepare mask for valid 32-bit batch (depends on batch).
213 batch >>= XNN_LOG2_SIZEOF_FLOAT;
214 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
215
216 // Load up to 15 inputs at a time.
217 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
218
219 // Subtract maximum input x := i - i_max.
220 const __m512 vx = _mm512_sub_ps(vi, vi_max);
221
222 // Compute reduced argument batch := round(x / log(2)).
223 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
224
225 // Compute reduced argument t := x - batch * log(2).
226 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
227 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
228 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
229
230 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
231 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
232 vp = _mm512_fmadd_ps(vp, vt, vc3);
233 vp = _mm512_fmadd_ps(vp, vt, vc2);
234 vp = _mm512_fmadd_ps(vp, vt, vc1);
235 vp = _mm512_fmadd_ps(vp, vt, vc0);
236
237 // Reconstruct the final f value:
238 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
239 // = 2**batch * p
240 __m512 vf = _mm512_scalef_ps(vp, vn);
241
242 // Multiply by scale.
243 vf = _mm512_mul_ps(vf, vscale);
244
245 // Store up to 15 outputs at a time.
246 _mm512_mask_storeu_ps(output, vmask, vf);
247 }
248 }