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_x48(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x48(
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 >= 48 * sizeof(float); elements -= 48 * sizeof(float)) {
42 // Load 48 (3x16) inputs at a time.
43 const __m512 vi0 = _mm512_loadu_ps(input);
44 const __m512 vi1 = _mm512_loadu_ps(input + 16);
45 const __m512 vi2 = _mm512_loadu_ps(input + 32);
46 input += 48;
47
48 // Subtract maximum input x := i - i_max.
49 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
50 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
51 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
52
53 // Compute reduced argument elements := round(x / log(2)).
54 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
55 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
56 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
57
58 // Compute reduced argument t := x - elements * 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 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
63
64 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
65 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
66 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
67
68 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
69 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
70 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
71 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
72
73 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
74 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
75 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
76
77 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
78 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
79 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
80
81 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
82 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
83 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
84
85 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
86 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
87 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
88
89 // Reconstruct the final f value:
90 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
91 // = 2**elements * p
92 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
93 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
94 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
95
96 // Multiply by scale.
97 vf0 = _mm512_mul_ps(vf0, vscale);
98 vf1 = _mm512_mul_ps(vf1, vscale);
99 vf2 = _mm512_mul_ps(vf2, vscale);
100
101 // Store 48 (3x16) outputs at a time.
102 _mm512_storeu_ps(output, vf0);
103 _mm512_storeu_ps(output + 0, vf0);
104 _mm512_storeu_ps(output + 16, vf1);
105 _mm512_storeu_ps(output + 32, vf2);
106 output += 48;
107 }
108 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
109 // Load 16 inputs at a time.
110 const __m512 vi = _mm512_loadu_ps(input);
111 input += 16;
112
113 // Subtract maximum input x := i - i_max.
114 const __m512 vx = _mm512_sub_ps(vi, vi_max);
115
116 // Compute reduced argument elements := round(x / log(2)).
117 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
118
119 // Compute reduced argument t := x - elements * log(2).
120 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
121 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
122 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
123
124 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
125 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
126 vp = _mm512_fmadd_ps(vp, vt, vc3);
127 vp = _mm512_fmadd_ps(vp, vt, vc2);
128 vp = _mm512_fmadd_ps(vp, vt, vc1);
129 vp = _mm512_fmadd_ps(vp, vt, vc0);
130
131 // Reconstruct the final f value:
132 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
133 // = 2**elements * p
134 __m512 vf = _mm512_scalef_ps(vp, vn);
135
136 // Multiply by scale.
137 vf = _mm512_mul_ps(vf, vscale);
138
139 // Store 16 outputs at a time.
140 _mm512_storeu_ps(output, vf);
141 output += 16;
142 }
143 if (elements != 0) {
144 // Prepare mask for valid 32-bit elements (depends on elements).
145 elements >>= 2 /* log2(sizeof(float)) */;
146 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
147
148 // Load up to 15 inputs at a time.
149 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
150
151 // Subtract maximum input x := i - i_max.
152 const __m512 vx = _mm512_sub_ps(vi, vi_max);
153
154 // Compute reduced argument elements := round(x / log(2)).
155 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
156
157 // Compute reduced argument t := x - elements * log(2).
158 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
159 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
160 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
161
162 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
163 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
164 vp = _mm512_fmadd_ps(vp, vt, vc3);
165 vp = _mm512_fmadd_ps(vp, vt, vc2);
166 vp = _mm512_fmadd_ps(vp, vt, vc1);
167 vp = _mm512_fmadd_ps(vp, vt, vc0);
168
169 // Reconstruct the final f value:
170 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
171 // = 2**elements * p
172 __m512 vf = _mm512_scalef_ps(vp, vn);
173
174 // Multiply by scale.
175 vf = _mm512_mul_ps(vf, vscale);
176
177 // Store up to 15 outputs at a time.
178 _mm512_mask_storeu_ps(output, vmask, vf);
179 }
180 }
181