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
xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x32(
19 size_t elements,
20 const float* input,
21 float* output,
22 float scale,
23 float max)
24 {
25 assert(elements % sizeof(float) == 0);
26
27 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
28 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
29 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
30
31 const __m512 vc0 = _mm512_set1_ps(1.0f);
32 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
33 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
34 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
35 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
36 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
37
38 const __m512 vscale = _mm512_set1_ps(scale);
39 const __m512 vi_max = _mm512_set1_ps(max);
40
41 for (; elements >= 32 * sizeof(float); elements -= 32 * sizeof(float)) {
42 // Load 32 (2x16) inputs at a time.
43 const __m512 vi0 = _mm512_loadu_ps(input);
44 const __m512 vi1 = _mm512_loadu_ps(input + 16);
45 input += 32;
46
47 // Subtract maximum input x := i - i_max.
48 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
49 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
50
51 // Compute reduced argument elements := round(x / log(2)).
52 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
53 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
54
55 // Compute reduced argument t := x - elements * log(2).
56 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
57 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
58 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
59
60 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
61 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
62
63 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
64 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
65 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
66
67 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
68 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
69
70 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
71 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
72
73 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
74 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
75
76 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
77 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
78
79 // Reconstruct the final f value:
80 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
81 // = 2**elements * p
82 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
83 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
84
85 // Multiply by scale.
86 vf0 = _mm512_mul_ps(vf0, vscale);
87 vf1 = _mm512_mul_ps(vf1, vscale);
88
89 // Store 32 (2x16) outputs at a time.
90 _mm512_storeu_ps(output, vf0);
91 _mm512_storeu_ps(output + 0, vf0);
92 _mm512_storeu_ps(output + 16, vf1);
93 output += 32;
94 }
95 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
96 // Load 16 inputs at a time.
97 const __m512 vi = _mm512_loadu_ps(input);
98 input += 16;
99
100 // Subtract maximum input x := i - i_max.
101 const __m512 vx = _mm512_sub_ps(vi, vi_max);
102
103 // Compute reduced argument elements := round(x / log(2)).
104 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
105
106 // Compute reduced argument t := x - elements * log(2).
107 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
108 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
109 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
110
111 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
112 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
113 vp = _mm512_fmadd_ps(vp, vt, vc3);
114 vp = _mm512_fmadd_ps(vp, vt, vc2);
115 vp = _mm512_fmadd_ps(vp, vt, vc1);
116 vp = _mm512_fmadd_ps(vp, vt, vc0);
117
118 // Reconstruct the final f value:
119 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
120 // = 2**elements * p
121 __m512 vf = _mm512_scalef_ps(vp, vn);
122
123 // Multiply by scale.
124 vf = _mm512_mul_ps(vf, vscale);
125
126 // Store 16 outputs at a time.
127 _mm512_storeu_ps(output, vf);
128 output += 16;
129 }
130 if (elements != 0) {
131 // Prepare mask for valid 32-bit elements (depends on elements).
132 elements >>= 2 /* log2(sizeof(float)) */;
133 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
134
135 // Load up to 15 inputs at a time.
136 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
137
138 // Subtract maximum input x := i - i_max.
139 const __m512 vx = _mm512_sub_ps(vi, vi_max);
140
141 // Compute reduced argument elements := round(x / log(2)).
142 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
143
144 // Compute reduced argument t := x - elements * log(2).
145 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
146 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
147 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
148
149 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
150 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
151 vp = _mm512_fmadd_ps(vp, vt, vc3);
152 vp = _mm512_fmadd_ps(vp, vt, vc2);
153 vp = _mm512_fmadd_ps(vp, vt, vc1);
154 vp = _mm512_fmadd_ps(vp, vt, vc0);
155
156 // Reconstruct the final f value:
157 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
158 // = 2**elements * p
159 __m512 vf = _mm512_scalef_ps(vp, vn);
160
161 // Multiply by scale.
162 vf = _mm512_mul_ps(vf, vscale);
163
164 // Store up to 15 outputs at a time.
165 _mm512_mask_storeu_ps(output, vmask, vf);
166 }
167 }
168