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_x176(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x176(
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 >= 176 * sizeof(float); elements -= 176 * sizeof(float)) {
43 // Load 176 (11x16) 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 const __m512 vx10 = _mm512_loadu_ps(x + 160);
55 x += 176;
56
57 // Compute reduced argument elements := round(x / log(2)).
58 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
59 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
60 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
61 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
62 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
63 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
64 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
65 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
66 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
67 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
68 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
69
70 // Compute reduced argument t := x - elements * log(2).
71 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
72 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
73 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
74 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
75 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
76 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
77 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
78 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
79 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
80 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
81 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
82 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
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 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
93 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
94 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
95
96 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
97 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
98 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
99 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
100 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
101 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
102 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
103 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
104 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
105 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
106 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
107 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
108
109 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
110 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
111 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
112 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
113 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
114 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
115 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
116 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
117 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
118 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
119 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
120
121 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
122 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
123 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
124 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
125 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
126 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
127 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
128 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
129 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
130 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
131 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
132
133 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
134 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
135 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
136 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
137 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
138 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
139 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
140 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
141 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
142 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
143 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
144
145 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
146 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
147 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
148 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
149 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
150 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
151 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
152 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
153 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
154 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
155 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
156
157 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
158 // - vnX is "exponent"
159 // - vpX is "mantissa"
160 //
161 // exp2(ae) * av * exp2(be) * bv =
162 // = exp2(ae + be) * (av * bv)
163 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
164 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
165 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
166 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
167 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
168 __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
169 __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
170 __m512 vf7 = _mm512_mul_ps(vp7, vscalev);
171 __m512 vf8 = _mm512_mul_ps(vp8, vscalev);
172 __m512 vf9 = _mm512_mul_ps(vp9, vscalev);
173 __m512 vf10 = _mm512_mul_ps(vp10, vscalev);
174
175 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
176 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
177 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
178 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
179 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
180 const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
181 const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
182 const __m512 ve7 = _mm512_add_ps(vn7, vscalee);
183 const __m512 ve8 = _mm512_add_ps(vn8, vscalee);
184 const __m512 ve9 = _mm512_add_ps(vn9, vscalee);
185 const __m512 ve10 = _mm512_add_ps(vn10, vscalee);
186
187 // Multiply "mantissa" by the exp2("exponent").
188 vf0 = _mm512_scalef_ps(vf0, ve0);
189 vf1 = _mm512_scalef_ps(vf1, ve1);
190 vf2 = _mm512_scalef_ps(vf2, ve2);
191 vf3 = _mm512_scalef_ps(vf3, ve3);
192 vf4 = _mm512_scalef_ps(vf4, ve4);
193 vf5 = _mm512_scalef_ps(vf5, ve5);
194 vf6 = _mm512_scalef_ps(vf6, ve6);
195 vf7 = _mm512_scalef_ps(vf7, ve7);
196 vf8 = _mm512_scalef_ps(vf8, ve8);
197 vf9 = _mm512_scalef_ps(vf9, ve9);
198 vf10 = _mm512_scalef_ps(vf10, ve10);
199
200 // Store 128 (8x16) results at a time.
201 _mm512_storeu_ps(y, vf0);
202 _mm512_storeu_ps(y + 0, vf0);
203 _mm512_storeu_ps(y + 16, vf1);
204 _mm512_storeu_ps(y + 32, vf2);
205 _mm512_storeu_ps(y + 48, vf3);
206 _mm512_storeu_ps(y + 64, vf4);
207 _mm512_storeu_ps(y + 80, vf5);
208 _mm512_storeu_ps(y + 96, vf6);
209 _mm512_storeu_ps(y + 112, vf7);
210 _mm512_storeu_ps(y + 128, vf8);
211 _mm512_storeu_ps(y + 144, vf9);
212 _mm512_storeu_ps(y + 160, vf10);
213 y += 176;
214 }
215
216 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
217 // Load 16 inputs at a time.
218 const __m512 vx = _mm512_loadu_ps(x);
219 x += 16;
220
221 // Compute reduced argument elements := round(x / log(2)).
222 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
223
224 // Compute reduced argument t := x - elements * log(2).
225 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
226 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
227 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
228
229 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
230 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
231 vp = _mm512_fmadd_ps(vp, vt, vc3);
232 vp = _mm512_fmadd_ps(vp, vt, vc2);
233 vp = _mm512_fmadd_ps(vp, vt, vc1);
234 vp = _mm512_fmadd_ps(vp, vt, vc0);
235
236 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
237 __m512 vf = _mm512_mul_ps(vp, vscalev);
238 const __m512 ve = _mm512_add_ps(vn, vscalee);
239
240 // Multiply "mantissa" by the exp2("exponent").
241 vf = _mm512_scalef_ps(vf, ve);
242
243 // Store 16 results at a time.
244 _mm512_storeu_ps(y, vf);
245 y += 16;
246 }
247 if XNN_UNLIKELY(elements != 0) {
248 // Prepare mask for valid 32-bit elements (depends on elements).
249 elements >>= 2 /* log2(sizeof(float)) */;
250 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
251
252 // Load up to 15 inputs at a time.
253 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
254
255 // Compute reduced argument elements := round(x / log(2)).
256 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
257
258 // Compute reduced argument t := x - elements * log(2).
259 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
260 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
261 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
262
263 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
264 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
265 vp = _mm512_fmadd_ps(vp, vt, vc3);
266 vp = _mm512_fmadd_ps(vp, vt, vc2);
267 vp = _mm512_fmadd_ps(vp, vt, vc1);
268 vp = _mm512_fmadd_ps(vp, vt, vc0);
269
270 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
271 __m512 vf = _mm512_mul_ps(vp, vscalev);
272 const __m512 ve = _mm512_add_ps(vn, vscalee);
273
274 // Multiply "mantissa" by the exp2("exponent").
275 vf = _mm512_scalef_ps(vf, ve);
276
277 // Store up to 15 results at a time.
278 _mm512_mask_storeu_ps(y, vmask, vf);
279 }
280 _mm256_zeroupper();
281 }
282