Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u112.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_u112(
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 >= 112 * sizeof(float); batch -= 112 * sizeof(float)) {
45 // Load 112 (7x16) 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 input += 112;
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
64 // Compute reduced argument batch := round(x / log(2)).
65 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
66 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
67 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
68 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
69 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
70 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
71 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
72
73 // Compute reduced argument t := x - batch * log(2).
74 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
75 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
76 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
77 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
78 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
79 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
80 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
82
83 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
84 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
85 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
86 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
87 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
88 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
89 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
90
91 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
92 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
93 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
94 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
95 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
96 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
97 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
98 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
99
100 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
101 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
102 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
103 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
104 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
105 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
106 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
107
108 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
109 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
110 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
111 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
112 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
113 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
114 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
115
116 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
117 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
118 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
119 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
120 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
121 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
122 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
123
124 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
125 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
126 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
127 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
128 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
129 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
130 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
131
132 // Reconstruct the final f value:
133 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
134 // = 2**batch * p
135 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
136 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
137 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
138 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
139 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
140 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
141 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
142
143 // Multiply by scale.
144 vf0 = _mm512_mul_ps(vf0, vscale);
145 vf1 = _mm512_mul_ps(vf1, vscale);
146 vf2 = _mm512_mul_ps(vf2, vscale);
147 vf3 = _mm512_mul_ps(vf3, vscale);
148 vf4 = _mm512_mul_ps(vf4, vscale);
149 vf5 = _mm512_mul_ps(vf5, vscale);
150 vf6 = _mm512_mul_ps(vf6, vscale);
151
152 // Store 112 (7x16) outputs at a time.
153 _mm512_storeu_ps(output, vf0);
154 _mm512_storeu_ps(output + 0, vf0);
155 _mm512_storeu_ps(output + 16, vf1);
156 _mm512_storeu_ps(output + 32, vf2);
157 _mm512_storeu_ps(output + 48, vf3);
158 _mm512_storeu_ps(output + 64, vf4);
159 _mm512_storeu_ps(output + 80, vf5);
160 _mm512_storeu_ps(output + 96, vf6);
161 output += 112;
162 }
163 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
164 // Load 16 inputs at a time.
165 const __m512 vi = _mm512_loadu_ps(input);
166 input += 16;
167
168 // Subtract maximum input x := i - i_max.
169 const __m512 vx = _mm512_sub_ps(vi, vi_max);
170
171 // Compute reduced argument batch := round(x / log(2)).
172 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
173
174 // Compute reduced argument t := x - batch * log(2).
175 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
176 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
177 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
178
179 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
180 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
181 vp = _mm512_fmadd_ps(vp, vt, vc3);
182 vp = _mm512_fmadd_ps(vp, vt, vc2);
183 vp = _mm512_fmadd_ps(vp, vt, vc1);
184 vp = _mm512_fmadd_ps(vp, vt, vc0);
185
186 // Reconstruct the final f value:
187 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
188 // = 2**batch * p
189 __m512 vf = _mm512_scalef_ps(vp, vn);
190
191 // Multiply by scale.
192 vf = _mm512_mul_ps(vf, vscale);
193
194 // Store 16 outputs at a time.
195 _mm512_storeu_ps(output, vf);
196 output += 16;
197 }
198 if (batch != 0) {
199 // Prepare mask for valid 32-bit batch (depends on batch).
200 batch >>= XNN_LOG2_SIZEOF_FLOAT;
201 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
202
203 // Load up to 15 inputs at a time.
204 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
205
206 // Subtract maximum input x := i - i_max.
207 const __m512 vx = _mm512_sub_ps(vi, vi_max);
208
209 // Compute reduced argument batch := round(x / log(2)).
210 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
211
212 // Compute reduced argument t := x - batch * log(2).
213 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
214 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
215 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
216
217 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
218 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
219 vp = _mm512_fmadd_ps(vp, vt, vc3);
220 vp = _mm512_fmadd_ps(vp, vt, vc2);
221 vp = _mm512_fmadd_ps(vp, vt, vc1);
222 vp = _mm512_fmadd_ps(vp, vt, vc0);
223
224 // Reconstruct the final f value:
225 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
226 // = 2**batch * p
227 __m512 vf = _mm512_scalef_ps(vp, vn);
228
229 // Multiply by scale.
230 vf = _mm512_mul_ps(vf, vscale);
231
232 // Store up to 15 outputs at a time.
233 _mm512_mask_storeu_ps(output, vmask, vf);
234 }
235 }