Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u48.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_u48(
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 >= 48 * sizeof(float); batch -= 48 * sizeof(float)) {
45 // Load 48 (3x16) 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 input += 48;
50
51 // Subtract maximum input x := i - i_max.
52 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
53 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
54 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
55
56 // Compute reduced argument batch := round(x / log(2)).
57 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
58 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
59 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
60
61 // Compute reduced argument t := x - batch * log(2).
62 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
63 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
64 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
65 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
66
67 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
68 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
69 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
70
71 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
72 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
73 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
74 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
75
76 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
77 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
78 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
79
80 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
81 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
82 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
83
84 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
85 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
86 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
87
88 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
89 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
90 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
91
92 // Reconstruct the final f value:
93 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
94 // = 2**batch * p
95 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
96 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
97 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
98
99 // Multiply by scale.
100 vf0 = _mm512_mul_ps(vf0, vscale);
101 vf1 = _mm512_mul_ps(vf1, vscale);
102 vf2 = _mm512_mul_ps(vf2, vscale);
103
104 // Store 48 (3x16) outputs at a time.
105 _mm512_storeu_ps(output, vf0);
106 _mm512_storeu_ps(output + 0, vf0);
107 _mm512_storeu_ps(output + 16, vf1);
108 _mm512_storeu_ps(output + 32, vf2);
109 output += 48;
110 }
111 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
112 // Load 16 inputs at a time.
113 const __m512 vi = _mm512_loadu_ps(input);
114 input += 16;
115
116 // Subtract maximum input x := i - i_max.
117 const __m512 vx = _mm512_sub_ps(vi, vi_max);
118
119 // Compute reduced argument batch := round(x / log(2)).
120 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
121
122 // Compute reduced argument t := x - batch * log(2).
123 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
124 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
125 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
126
127 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
128 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
129 vp = _mm512_fmadd_ps(vp, vt, vc3);
130 vp = _mm512_fmadd_ps(vp, vt, vc2);
131 vp = _mm512_fmadd_ps(vp, vt, vc1);
132 vp = _mm512_fmadd_ps(vp, vt, vc0);
133
134 // Reconstruct the final f value:
135 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
136 // = 2**batch * p
137 __m512 vf = _mm512_scalef_ps(vp, vn);
138
139 // Multiply by scale.
140 vf = _mm512_mul_ps(vf, vscale);
141
142 // Store 16 outputs at a time.
143 _mm512_storeu_ps(output, vf);
144 output += 16;
145 }
146 if (batch != 0) {
147 // Prepare mask for valid 32-bit batch (depends on batch).
148 batch >>= XNN_LOG2_SIZEOF_FLOAT;
149 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
150
151 // Load up to 15 inputs at a time.
152 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
153
154 // Subtract maximum input x := i - i_max.
155 const __m512 vx = _mm512_sub_ps(vi, vi_max);
156
157 // Compute reduced argument batch := round(x / log(2)).
158 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
159
160 // Compute reduced argument t := x - batch * log(2).
161 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
162 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
163 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
164
165 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
166 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
167 vp = _mm512_fmadd_ps(vp, vt, vc3);
168 vp = _mm512_fmadd_ps(vp, vt, vc2);
169 vp = _mm512_fmadd_ps(vp, vt, vc1);
170 vp = _mm512_fmadd_ps(vp, vt, vc0);
171
172 // Reconstruct the final f value:
173 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
174 // = 2**batch * p
175 __m512 vf = _mm512_scalef_ps(vp, vn);
176
177 // Multiply by scale.
178 vf = _mm512_mul_ps(vf, vscale);
179
180 // Store up to 15 outputs at a time.
181 _mm512_mask_storeu_ps(output, vmask, vf);
182 }
183 }