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_x128(size_t elements,const float * input,float * output,float scale,float max)18 void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_x128(
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 >= 128 * sizeof(float); elements -= 128 * sizeof(float)) {
42 // Load 128 (8x16) 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 const __m512 vi5 = _mm512_loadu_ps(input + 80);
49 const __m512 vi6 = _mm512_loadu_ps(input + 96);
50 const __m512 vi7 = _mm512_loadu_ps(input + 112);
51 input += 128;
52
53 // Subtract maximum input x := i - i_max.
54 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
55 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
56 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
57 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
58 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
59 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
60 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
61 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
62
63 // Compute reduced argument elements := round(x / log(2)).
64 __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
65 __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
66 __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
67 __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
68 __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
69 __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
70 __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
71 __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
72
73 // Compute reduced argument t := x - elements * log(2).
74 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
75 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
76 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
77 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
78 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
79 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
80 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
81 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
82 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
83
84 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
85 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
86 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
87 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
88 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
89 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
90 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
91 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
92
93 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
94 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
95 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
96 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
97 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
98 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
99 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
100 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
101 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
102
103 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
104 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
105 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
106 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
107 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
108 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
109 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
110 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
111
112 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
113 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
114 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
115 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
116 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
117 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
118 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
119 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
120
121 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
122 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
123 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
124 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
125 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
126 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
128 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
129
130 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
131 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
132 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
133 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
134 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
135 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
136 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
137 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
138
139 // Reconstruct the final f value:
140 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
141 // = 2**elements * p
142 __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
143 __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
144 __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
145 __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
146 __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
147 __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
148 __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
149 __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
150
151 // Multiply by scale.
152 vf0 = _mm512_mul_ps(vf0, vscale);
153 vf1 = _mm512_mul_ps(vf1, vscale);
154 vf2 = _mm512_mul_ps(vf2, vscale);
155 vf3 = _mm512_mul_ps(vf3, vscale);
156 vf4 = _mm512_mul_ps(vf4, vscale);
157 vf5 = _mm512_mul_ps(vf5, vscale);
158 vf6 = _mm512_mul_ps(vf6, vscale);
159 vf7 = _mm512_mul_ps(vf7, vscale);
160
161 // Store 128 (8x16) outputs at a time.
162 _mm512_storeu_ps(output, vf0);
163 _mm512_storeu_ps(output + 0, vf0);
164 _mm512_storeu_ps(output + 16, vf1);
165 _mm512_storeu_ps(output + 32, vf2);
166 _mm512_storeu_ps(output + 48, vf3);
167 _mm512_storeu_ps(output + 64, vf4);
168 _mm512_storeu_ps(output + 80, vf5);
169 _mm512_storeu_ps(output + 96, vf6);
170 _mm512_storeu_ps(output + 112, vf7);
171 output += 128;
172 }
173 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
174 // Load 16 inputs at a time.
175 const __m512 vi = _mm512_loadu_ps(input);
176 input += 16;
177
178 // Subtract maximum input x := i - i_max.
179 const __m512 vx = _mm512_sub_ps(vi, vi_max);
180
181 // Compute reduced argument elements := round(x / log(2)).
182 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
183
184 // Compute reduced argument t := x - elements * log(2).
185 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
186 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
187 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
188
189 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
190 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
191 vp = _mm512_fmadd_ps(vp, vt, vc3);
192 vp = _mm512_fmadd_ps(vp, vt, vc2);
193 vp = _mm512_fmadd_ps(vp, vt, vc1);
194 vp = _mm512_fmadd_ps(vp, vt, vc0);
195
196 // Reconstruct the final f value:
197 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
198 // = 2**elements * p
199 __m512 vf = _mm512_scalef_ps(vp, vn);
200
201 // Multiply by scale.
202 vf = _mm512_mul_ps(vf, vscale);
203
204 // Store 16 outputs at a time.
205 _mm512_storeu_ps(output, vf);
206 output += 16;
207 }
208 if (elements != 0) {
209 // Prepare mask for valid 32-bit elements (depends on elements).
210 elements >>= 2 /* log2(sizeof(float)) */;
211 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
212
213 // Load up to 15 inputs at a time.
214 const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input);
215
216 // Subtract maximum input x := i - i_max.
217 const __m512 vx = _mm512_sub_ps(vi, vi_max);
218
219 // Compute reduced argument elements := round(x / log(2)).
220 __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
221
222 // Compute reduced argument t := x - elements * log(2).
223 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
224 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
225 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
226
227 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
228 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
229 vp = _mm512_fmadd_ps(vp, vt, vc3);
230 vp = _mm512_fmadd_ps(vp, vt, vc2);
231 vp = _mm512_fmadd_ps(vp, vt, vc1);
232 vp = _mm512_fmadd_ps(vp, vt, vc0);
233
234 // Reconstruct the final f value:
235 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
236 // = 2**elements * p
237 __m512 vf = _mm512_scalef_ps(vp, vn);
238
239 // Multiply by scale.
240 vf = _mm512_mul_ps(vf, vscale);
241
242 // Store up to 15 outputs at a time.
243 _mm512_mask_storeu_ps(output, vmask, vf);
244 }
245 }
246