Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u80.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_u80(
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 >= 80 * sizeof(float); batch -= 80 * sizeof(float)) {
45 // Load 80 (5x16) 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 input += 80;
52
53 // Subtract maximum input x := i - i_max.
54 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
55 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
56 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
57 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
58 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
59
60 // Compute reduced argument batch := round(x / log(2)).
61 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
62 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
63 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
64 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
65 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
66
67 // Compute reduced argument t := x - batch * log(2).
68 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
69 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
70 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
71 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
72 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
73 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
74
75 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
76 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
77 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
78 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
79 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
80
81 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
82 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
83 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
84 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
85 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
86 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
87
88 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
89 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
90 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
91 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
92 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
93
94 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
95 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
96 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
97 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
98 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
99
100 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
101 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
102 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
103 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
104 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
105
106 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
107 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
108 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
109 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
110 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
111
112 // Reconstruct the final f value:
113 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
114 // = 2**batch * p
115 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
116 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
117 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
118 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
119 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
120
121 // Multiply by scale.
122 vf0 = _mm512_mul_ps(vf0, vscale);
123 vf1 = _mm512_mul_ps(vf1, vscale);
124 vf2 = _mm512_mul_ps(vf2, vscale);
125 vf3 = _mm512_mul_ps(vf3, vscale);
126 vf4 = _mm512_mul_ps(vf4, vscale);
127
128 // Store 80 (5x16) outputs at a time.
129 _mm512_storeu_ps(output, vf0);
130 _mm512_storeu_ps(output + 0, vf0);
131 _mm512_storeu_ps(output + 16, vf1);
132 _mm512_storeu_ps(output + 32, vf2);
133 _mm512_storeu_ps(output + 48, vf3);
134 _mm512_storeu_ps(output + 64, vf4);
135 output += 80;
136 }
137 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
138 // Load 16 inputs at a time.
139 const __m512 vi = _mm512_loadu_ps(input);
140 input += 16;
141
142 // Subtract maximum input x := i - i_max.
143 const __m512 vx = _mm512_sub_ps(vi, vi_max);
144
145 // Compute reduced argument batch := round(x / log(2)).
146 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
147
148 // Compute reduced argument t := x - batch * log(2).
149 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
150 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
151 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
152
153 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
154 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
155 vp = _mm512_fmadd_ps(vp, vt, vc3);
156 vp = _mm512_fmadd_ps(vp, vt, vc2);
157 vp = _mm512_fmadd_ps(vp, vt, vc1);
158 vp = _mm512_fmadd_ps(vp, vt, vc0);
159
160 // Reconstruct the final f value:
161 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
162 // = 2**batch * p
163 __m512 vf = _mm512_scalef_ps(vp, vn);
164
165 // Multiply by scale.
166 vf = _mm512_mul_ps(vf, vscale);
167
168 // Store 16 outputs at a time.
169 _mm512_storeu_ps(output, vf);
170 output += 16;
171 }
172 if (batch != 0) {
173 // Prepare mask for valid 32-bit batch (depends on batch).
174 batch >>= XNN_LOG2_SIZEOF_FLOAT;
175 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
176
177 // Load up to 15 inputs at a time.
178 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
179
180 // Subtract maximum input x := i - i_max.
181 const __m512 vx = _mm512_sub_ps(vi, vi_max);
182
183 // Compute reduced argument batch := round(x / log(2)).
184 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
185
186 // Compute reduced argument t := x - batch * log(2).
187 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
188 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
189 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
190
191 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
192 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
193 vp = _mm512_fmadd_ps(vp, vt, vc3);
194 vp = _mm512_fmadd_ps(vp, vt, vc2);
195 vp = _mm512_fmadd_ps(vp, vt, vc1);
196 vp = _mm512_fmadd_ps(vp, vt, vc0);
197
198 // Reconstruct the final f value:
199 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
200 // = 2**batch * p
201 __m512 vf = _mm512_scalef_ps(vp, vn);
202
203 // Multiply by scale.
204 vf = _mm512_mul_ps(vf, vscale);
205
206 // Store up to 15 outputs at a time.
207 _mm512_mask_storeu_ps(output, vmask, vf);
208 }
209 }