Debian Codegraph @

src/f32-vscaleexpminusmax/gen/f32-vscaleexpminusmax-avx512f-p5-scalef-u64.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_u64(
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 >= 64 * sizeof(float); batch -= 64 * sizeof(float)) {
45 // Load 64 (4x16) 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 input += 64;
51
52 // Subtract maximum input x := i - i_max.
53 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
54 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
55 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
56 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
57
58 // Compute reduced argument batch := round(x / log(2)).
59 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
60 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
61 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
62 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
63
64 // Compute reduced argument t := x - batch * log(2).
65 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
66 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
67 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
68 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
69 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
70
71 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
72 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
73 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
74 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
75
76 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
77 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
78 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
79 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
80 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
81
82 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
83 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
84 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
85 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
86
87 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
88 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
89 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
90 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
91
92 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
93 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
94 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
95 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
96
97 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
98 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
99 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
100 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
101
102 // Reconstruct the final f value:
103 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
104 // = 2**batch * p
105 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
106 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
107 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
108 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
109
110 // Multiply by scale.
111 vf0 = _mm512_mul_ps(vf0, vscale);
112 vf1 = _mm512_mul_ps(vf1, vscale);
113 vf2 = _mm512_mul_ps(vf2, vscale);
114 vf3 = _mm512_mul_ps(vf3, vscale);
115
116 // Store 64 (4x16) outputs at a time.
117 _mm512_storeu_ps(output, vf0);
118 _mm512_storeu_ps(output + 0, vf0);
119 _mm512_storeu_ps(output + 16, vf1);
120 _mm512_storeu_ps(output + 32, vf2);
121 _mm512_storeu_ps(output + 48, vf3);
122 output += 64;
123 }
124 for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) {
125 // Load 16 inputs at a time.
126 const __m512 vi = _mm512_loadu_ps(input);
127 input += 16;
128
129 // Subtract maximum input x := i - i_max.
130 const __m512 vx = _mm512_sub_ps(vi, vi_max);
131
132 // Compute reduced argument batch := round(x / log(2)).
133 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
134
135 // Compute reduced argument t := x - batch * log(2).
136 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
137 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
138 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
139
140 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
141 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
142 vp = _mm512_fmadd_ps(vp, vt, vc3);
143 vp = _mm512_fmadd_ps(vp, vt, vc2);
144 vp = _mm512_fmadd_ps(vp, vt, vc1);
145 vp = _mm512_fmadd_ps(vp, vt, vc0);
146
147 // Reconstruct the final f value:
148 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
149 // = 2**batch * p
150 __m512 vf = _mm512_scalef_ps(vp, vn);
151
152 // Multiply by scale.
153 vf = _mm512_mul_ps(vf, vscale);
154
155 // Store 16 outputs at a time.
156 _mm512_storeu_ps(output, vf);
157 output += 16;
158 }
159 if (batch != 0) {
160 // Prepare mask for valid 32-bit batch (depends on batch).
161 batch >>= XNN_LOG2_SIZEOF_FLOAT;
162 const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1)));
163
164 // Load up to 15 inputs at a time.
165 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
166
167 // Subtract maximum input x := i - i_max.
168 const __m512 vx = _mm512_sub_ps(vi, vi_max);
169
170 // Compute reduced argument batch := round(x / log(2)).
171 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
172
173 // Compute reduced argument t := x - batch * log(2).
174 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
175 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
176 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
177
178 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
179 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
180 vp = _mm512_fmadd_ps(vp, vt, vc3);
181 vp = _mm512_fmadd_ps(vp, vt, vc2);
182 vp = _mm512_fmadd_ps(vp, vt, vc1);
183 vp = _mm512_fmadd_ps(vp, vt, vc0);
184
185 // Reconstruct the final f value:
186 // f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
187 // = 2**batch * p
188 __m512 vf = _mm512_scalef_ps(vp, vn);
189
190 // Multiply by scale.
191 vf = _mm512_mul_ps(vf, vscale);
192
193 // Store up to 15 outputs at a time.
194 _mm512_mask_storeu_ps(output, vmask, vf);
195 }
196 }