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_x80(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x80(
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 >= 80 * sizeof(float); elements -= 80 * sizeof(float)) {
42 // Load 80 (5x16) 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 const __m512 vi3 = _mm512_loadu_ps(input + 48);
47 const __m512 vi4 = _mm512_loadu_ps(input + 64);
48 input += 80;
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 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
54 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
55 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
56
57 // Compute reduced argument elements := round(x / log(2)).
58 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
59 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
60 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
61 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
62 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
63
64 // Compute reduced argument t := x - elements * 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 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
71
72 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
73 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
74 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
75 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
76 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
77
78 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
79 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
80 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
81 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
82 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
83 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
84
85 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
86 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
87 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
88 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
89 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
90
91 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
92 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
93 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
94 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
95 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
96
97 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
98 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
99 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
100 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
101 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
102
103 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
104 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
105 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
106 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
107 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
108
109 // Reconstruct the final f value:
110 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
111 // = 2**elements * p
112 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
113 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
114 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
115 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
116 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
117
118 // Multiply by scale.
119 vf0 = _mm512_mul_ps(vf0, vscale);
120 vf1 = _mm512_mul_ps(vf1, vscale);
121 vf2 = _mm512_mul_ps(vf2, vscale);
122 vf3 = _mm512_mul_ps(vf3, vscale);
123 vf4 = _mm512_mul_ps(vf4, vscale);
124
125 // Store 80 (5x16) outputs at a time.
126 _mm512_storeu_ps(output, vf0);
127 _mm512_storeu_ps(output + 0, vf0);
128 _mm512_storeu_ps(output + 16, vf1);
129 _mm512_storeu_ps(output + 32, vf2);
130 _mm512_storeu_ps(output + 48, vf3);
131 _mm512_storeu_ps(output + 64, vf4);
132 output += 80;
133 }
134 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
135 // Load 16 inputs at a time.
136 const __m512 vi = _mm512_loadu_ps(input);
137 input += 16;
138
139 // Subtract maximum input x := i - i_max.
140 const __m512 vx = _mm512_sub_ps(vi, vi_max);
141
142 // Compute reduced argument elements := round(x / log(2)).
143 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
144
145 // Compute reduced argument t := x - elements * log(2).
146 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
147 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
148 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
149
150 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
151 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
152 vp = _mm512_fmadd_ps(vp, vt, vc3);
153 vp = _mm512_fmadd_ps(vp, vt, vc2);
154 vp = _mm512_fmadd_ps(vp, vt, vc1);
155 vp = _mm512_fmadd_ps(vp, vt, vc0);
156
157 // Reconstruct the final f value:
158 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
159 // = 2**elements * p
160 __m512 vf = _mm512_scalef_ps(vp, vn);
161
162 // Multiply by scale.
163 vf = _mm512_mul_ps(vf, vscale);
164
165 // Store 16 outputs at a time.
166 _mm512_storeu_ps(output, vf);
167 output += 16;
168 }
169 if (elements != 0) {
170 // Prepare mask for valid 32-bit elements (depends on elements).
171 elements >>= 2 /* log2(sizeof(float)) */;
172 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
173
174 // Load up to 15 inputs at a time.
175 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
176
177 // Subtract maximum input x := i - i_max.
178 const __m512 vx = _mm512_sub_ps(vi, vi_max);
179
180 // Compute reduced argument elements := round(x / log(2)).
181 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
182
183 // Compute reduced argument t := x - elements * log(2).
184 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
185 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
186 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
187
188 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
189 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
190 vp = _mm512_fmadd_ps(vp, vt, vc3);
191 vp = _mm512_fmadd_ps(vp, vt, vc2);
192 vp = _mm512_fmadd_ps(vp, vt, vc1);
193 vp = _mm512_fmadd_ps(vp, vt, vc0);
194
195 // Reconstruct the final f value:
196 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
197 // = 2**elements * p
198 __m512 vf = _mm512_scalef_ps(vp, vn);
199
200 // Multiply by scale.
201 vf = _mm512_mul_ps(vf, vscale);
202
203 // Store up to 15 outputs at a time.
204 _mm512_mask_storeu_ps(output, vmask, vf);
205 }
206 }
207