Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u96.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_u96(
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 >= 96 * sizeof(float); batch -= 96 * sizeof(float)) {
45 // Load 96 (6x16) 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 input += 96;
53
54 // Subtract maximum input x := i - i_max.
55 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
56 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
57 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
58 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
59 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
60 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
61
62 // Compute reduced argument batch := round(x / log(2)).
63 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
64 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
65 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
66 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
67 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
68 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
69
70 // Compute reduced argument t := x - batch * log(2).
71 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
72 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
73 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
74 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
75 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
76 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
77 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
78
79 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
80 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
81 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
82 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
83 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
84 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
85
86 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
87 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
88 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
89 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
90 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
91 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
92 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
93
94 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
95 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
96 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
97 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
98 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
99 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
100
101 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
102 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
103 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
104 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
105 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
106 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
107
108 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
109 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
110 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
111 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
112 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
113 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
114
115 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
116 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
117 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
118 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
119 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
120 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
121
122 // Reconstruct the final f value:
123 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
124 // = 2**batch * p
125 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
126 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
127 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
128 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
129 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
130 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
131
132 // Multiply by scale.
133 vf0 = _mm512_mul_ps(vf0, vscale);
134 vf1 = _mm512_mul_ps(vf1, vscale);
135 vf2 = _mm512_mul_ps(vf2, vscale);
136 vf3 = _mm512_mul_ps(vf3, vscale);
137 vf4 = _mm512_mul_ps(vf4, vscale);
138 vf5 = _mm512_mul_ps(vf5, vscale);
139
140 // Store 96 (6x16) outputs at a time.
141 _mm512_storeu_ps(output, vf0);
142 _mm512_storeu_ps(output + 0, vf0);
143 _mm512_storeu_ps(output + 16, vf1);
144 _mm512_storeu_ps(output + 32, vf2);
145 _mm512_storeu_ps(output + 48, vf3);
146 _mm512_storeu_ps(output + 64, vf4);
147 _mm512_storeu_ps(output + 80, vf5);
148 output += 96;
149 }
150 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
151 // Load 16 inputs at a time.
152 const __m512 vi = _mm512_loadu_ps(input);
153 input += 16;
154
155 // Subtract maximum input x := i - i_max.
156 const __m512 vx = _mm512_sub_ps(vi, vi_max);
157
158 // Compute reduced argument batch := round(x / log(2)).
159 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
160
161 // Compute reduced argument t := x - batch * log(2).
162 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
163 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
164 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
165
166 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
167 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
168 vp = _mm512_fmadd_ps(vp, vt, vc3);
169 vp = _mm512_fmadd_ps(vp, vt, vc2);
170 vp = _mm512_fmadd_ps(vp, vt, vc1);
171 vp = _mm512_fmadd_ps(vp, vt, vc0);
172
173 // Reconstruct the final f value:
174 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
175 // = 2**batch * p
176 __m512 vf = _mm512_scalef_ps(vp, vn);
177
178 // Multiply by scale.
179 vf = _mm512_mul_ps(vf, vscale);
180
181 // Store 16 outputs at a time.
182 _mm512_storeu_ps(output, vf);
183 output += 16;
184 }
185 if (batch != 0) {
186 // Prepare mask for valid 32-bit batch (depends on batch).
187 batch >>= XNN_LOG2_SIZEOF_FLOAT;
188 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
189
190 // Load up to 15 inputs at a time.
191 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
192
193 // Subtract maximum input x := i - i_max.
194 const __m512 vx = _mm512_sub_ps(vi, vi_max);
195
196 // Compute reduced argument batch := round(x / log(2)).
197 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
198
199 // Compute reduced argument t := x - batch * log(2).
200 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
201 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
202 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
203
204 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
205 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
206 vp = _mm512_fmadd_ps(vp, vt, vc3);
207 vp = _mm512_fmadd_ps(vp, vt, vc2);
208 vp = _mm512_fmadd_ps(vp, vt, vc1);
209 vp = _mm512_fmadd_ps(vp, vt, vc0);
210
211 // Reconstruct the final f value:
212 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
213 // = 2**batch * p
214 __m512 vf = _mm512_scalef_ps(vp, vn);
215
216 // Multiply by scale.
217 vf = _mm512_mul_ps(vf, vscale);
218
219 // Store up to 15 outputs at a time.
220 _mm512_mask_storeu_ps(output, vmask, vf);
221 }
222 }