Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u32.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_u32(
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 >= 32 * sizeof(float); batch -= 32 * sizeof(float)) {
45 // Load 32 (2x16) inputs at a time.
46 const __m512 vi0 = _mm512_loadu_ps(input);
47 const __m512 vi1 = _mm512_loadu_ps(input + 16);
48 input += 32;
49
50 // Subtract maximum input x := i - i_max.
51 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
52 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
53
54 // Compute reduced argument batch := round(x / log(2)).
55 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
56 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
57
58 // Compute reduced argument t := x - batch * log(2).
59 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
60 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
61 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
62
63 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
64 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
65
66 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
67 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
68 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
69
70 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
71 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
72
73 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
74 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
75
76 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
77 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
78
79 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
80 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
81
82 // Reconstruct the final f value:
83 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
84 // = 2**batch * p
85 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
86 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
87
88 // Multiply by scale.
89 vf0 = _mm512_mul_ps(vf0, vscale);
90 vf1 = _mm512_mul_ps(vf1, vscale);
91
92 // Store 32 (2x16) outputs at a time.
93 _mm512_storeu_ps(output, vf0);
94 _mm512_storeu_ps(output + 0, vf0);
95 _mm512_storeu_ps(output + 16, vf1);
96 output += 32;
97 }
98 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
99 // Load 16 inputs at a time.
100 const __m512 vi = _mm512_loadu_ps(input);
101 input += 16;
102
103 // Subtract maximum input x := i - i_max.
104 const __m512 vx = _mm512_sub_ps(vi, vi_max);
105
106 // Compute reduced argument batch := round(x / log(2)).
107 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
108
109 // Compute reduced argument t := x - batch * log(2).
110 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
111 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
112 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
113
114 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
115 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
116 vp = _mm512_fmadd_ps(vp, vt, vc3);
117 vp = _mm512_fmadd_ps(vp, vt, vc2);
118 vp = _mm512_fmadd_ps(vp, vt, vc1);
119 vp = _mm512_fmadd_ps(vp, vt, vc0);
120
121 // Reconstruct the final f value:
122 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
123 // = 2**batch * p
124 __m512 vf = _mm512_scalef_ps(vp, vn);
125
126 // Multiply by scale.
127 vf = _mm512_mul_ps(vf, vscale);
128
129 // Store 16 outputs at a time.
130 _mm512_storeu_ps(output, vf);
131 output += 16;
132 }
133 if (batch != 0) {
134 // Prepare mask for valid 32-bit batch (depends on batch).
135 batch >>= XNN_LOG2_SIZEOF_FLOAT;
136 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
137
138 // Load up to 15 inputs at a time.
139 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
140
141 // Subtract maximum input x := i - i_max.
142 const __m512 vx = _mm512_sub_ps(vi, vi_max);
143
144 // Compute reduced argument batch := round(x / log(2)).
145 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
146
147 // Compute reduced argument t := x - batch * log(2).
148 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
149 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
150 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
151
152 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
153 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
154 vp = _mm512_fmadd_ps(vp, vt, vc3);
155 vp = _mm512_fmadd_ps(vp, vt, vc2);
156 vp = _mm512_fmadd_ps(vp, vt, vc1);
157 vp = _mm512_fmadd_ps(vp, vt, vc0);
158
159 // Reconstruct the final f value:
160 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
161 // = 2**batch * p
162 __m512 vf = _mm512_scalef_ps(vp, vn);
163
164 // Multiply by scale.
165 vf = _mm512_mul_ps(vf, vscale);
166
167 // Store up to 15 outputs at a time.
168 _mm512_mask_storeu_ps(output, vmask, vf);
169 }
170 }