1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vscaleextexp/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/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vscaleextexp.h>
17
18
xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x160(
20 size_t elements,
21 const float* x,
22 float* y,
23 float scale_value,
24 float scale_exp)
25 {
26 assert(elements % sizeof(float) == 0);
27
28 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
29 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
30 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
31
32 const __m512 vc0 = _mm512_set1_ps(1.0f);
33 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
34 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
35 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
36 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
37 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
38
39 const __m512 vscalev = _mm512_set1_ps(scale_value);
40 const __m512 vscalee = _mm512_set1_ps(scale_exp);
41
42 for (; elements >= 160 * sizeof(float); elements -= 160 * sizeof(float)) {
43 // Load 160 (10x16) inputs at a time.
44 const __m512 vx0 = _mm512_loadu_ps(x);
45 const __m512 vx1 = _mm512_loadu_ps(x + 16);
46 const __m512 vx2 = _mm512_loadu_ps(x + 32);
47 const __m512 vx3 = _mm512_loadu_ps(x + 48);
48 const __m512 vx4 = _mm512_loadu_ps(x + 64);
49 const __m512 vx5 = _mm512_loadu_ps(x + 80);
50 const __m512 vx6 = _mm512_loadu_ps(x + 96);
51 const __m512 vx7 = _mm512_loadu_ps(x + 112);
52 const __m512 vx8 = _mm512_loadu_ps(x + 128);
53 const __m512 vx9 = _mm512_loadu_ps(x + 144);
54 x += 160;
55
56 // Compute reduced argument elements := round(x / log(2)).
57 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
58 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
59 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
60 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
61 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
62 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
63 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
64 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
65 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
66 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
67
68 // Compute reduced argument t := x - elements * log(2).
69 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
70 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
71 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
72 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
73 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
74 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
75 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
76 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
77 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
78 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
79 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
80
81 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
82 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
83 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
84 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
85 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
86 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
87 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
88 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
89 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
90 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
91
92 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
93 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
94 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
95 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
96 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
97 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
98 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
99 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
100 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
101 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
102 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
103
104 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
105 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
106 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
107 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
108 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
109 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
110 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
111 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
112 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
113 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
114
115 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
116 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
117 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
118 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
119 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
120 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
121 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
122 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
123 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
124 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
125
126 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
127 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
128 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
129 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
130 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
131 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
132 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
133 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
134 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
135 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
136
137 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
138 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
139 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
140 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
141 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
142 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
143 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
144 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
145 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
146 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
147
148 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
149 // - vnX is "exponent"
150 // - vpX is "mantissa"
151 //
152 // exp2(ae) * av * exp2(be) * bv =
153 // = exp2(ae + be) * (av * bv)
154 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
155 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
156 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
157 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
158 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
159 __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
160 __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
161 __m512 vf7 = _mm512_mul_ps(vp7, vscalev);
162 __m512 vf8 = _mm512_mul_ps(vp8, vscalev);
163 __m512 vf9 = _mm512_mul_ps(vp9, vscalev);
164
165 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
166 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
167 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
168 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
169 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
170 const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
171 const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
172 const __m512 ve7 = _mm512_add_ps(vn7, vscalee);
173 const __m512 ve8 = _mm512_add_ps(vn8, vscalee);
174 const __m512 ve9 = _mm512_add_ps(vn9, vscalee);
175
176 // Multiply "mantissa" by the exp2("exponent").
177 vf0 = _mm512_scalef_ps(vf0, ve0);
178 vf1 = _mm512_scalef_ps(vf1, ve1);
179 vf2 = _mm512_scalef_ps(vf2, ve2);
180 vf3 = _mm512_scalef_ps(vf3, ve3);
181 vf4 = _mm512_scalef_ps(vf4, ve4);
182 vf5 = _mm512_scalef_ps(vf5, ve5);
183 vf6 = _mm512_scalef_ps(vf6, ve6);
184 vf7 = _mm512_scalef_ps(vf7, ve7);
185 vf8 = _mm512_scalef_ps(vf8, ve8);
186 vf9 = _mm512_scalef_ps(vf9, ve9);
187
188 // Store 128 (8x16) results at a time.
189 _mm512_storeu_ps(y, vf0);
190 _mm512_storeu_ps(y + 0, vf0);
191 _mm512_storeu_ps(y + 16, vf1);
192 _mm512_storeu_ps(y + 32, vf2);
193 _mm512_storeu_ps(y + 48, vf3);
194 _mm512_storeu_ps(y + 64, vf4);
195 _mm512_storeu_ps(y + 80, vf5);
196 _mm512_storeu_ps(y + 96, vf6);
197 _mm512_storeu_ps(y + 112, vf7);
198 _mm512_storeu_ps(y + 128, vf8);
199 _mm512_storeu_ps(y + 144, vf9);
200 y += 160;
201 }
202
203 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
204 // Load 16 inputs at a time.
205 const __m512 vx = _mm512_loadu_ps(x);
206 x += 16;
207
208 // Compute reduced argument elements := round(x / log(2)).
209 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
210
211 // Compute reduced argument t := x - elements * log(2).
212 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
213 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
214 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
215
216 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
217 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
218 vp = _mm512_fmadd_ps(vp, vt, vc3);
219 vp = _mm512_fmadd_ps(vp, vt, vc2);
220 vp = _mm512_fmadd_ps(vp, vt, vc1);
221 vp = _mm512_fmadd_ps(vp, vt, vc0);
222
223 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
224 __m512 vf = _mm512_mul_ps(vp, vscalev);
225 const __m512 ve = _mm512_add_ps(vn, vscalee);
226
227 // Multiply "mantissa" by the exp2("exponent").
228 vf = _mm512_scalef_ps(vf, ve);
229
230 // Store 16 results at a time.
231 _mm512_storeu_ps(y, vf);
232 y += 16;
233 }
234 if XNN_UNLIKELY(elements != 0) {
235 // Prepare mask for valid 32-bit elements (depends on elements).
236 elements >>= 2 /* log2(sizeof(float)) */;
237 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
238
239 // Load up to 15 inputs at a time.
240 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
241
242 // Compute reduced argument elements := round(x / log(2)).
243 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
244
245 // Compute reduced argument t := x - elements * log(2).
246 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
247 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
248 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
249
250 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
251 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
252 vp = _mm512_fmadd_ps(vp, vt, vc3);
253 vp = _mm512_fmadd_ps(vp, vt, vc2);
254 vp = _mm512_fmadd_ps(vp, vt, vc1);
255 vp = _mm512_fmadd_ps(vp, vt, vc0);
256
257 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
258 __m512 vf = _mm512_mul_ps(vp, vscalev);
259 const __m512 ve = _mm512_add_ps(vn, vscalee);
260
261 // Multiply "mantissa" by the exp2("exponent").
262 vf = _mm512_scalef_ps(vf, ve);
263
264 // Store up to 15 results at a time.
265 _mm512_mask_storeu_ps(y, vmask, vf);
266 }
267 _mm256_zeroupper();
268 }
269