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_x192(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x192(
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 >= 192 * sizeof(float); elements -= 192 * sizeof(float)) {
43 // Load 192 (12x16) 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 const __m512 vx11 = _mm512_loadu_ps(x + 176);
56 x += 192;
57
58 // Compute reduced argument elements := round(x / log(2)).
59 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
60 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
61 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
62 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
63 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
64 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
65 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
66 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
67 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
68 const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
69 const __m512 vn10 = _mm512_roundscale_ps(_mm512_mul_ps(vx10, vlog2e), 0);
70 const __m512 vn11 = _mm512_roundscale_ps(_mm512_mul_ps(vx11, vlog2e), 0);
71
72 // Compute reduced argument t := x - elements * log(2).
73 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
74 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
75 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
76 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
77 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
78 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
79 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
80 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
81 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
82 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
83 __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
84 __m512 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_hi, vx10);
85 __m512 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_hi, vx11);
86
87 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
88 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
89 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
90 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
91 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
92 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
93 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
94 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
95 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
96 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
97 vt10 = _mm512_fmadd_ps(vn10, vminus_ln2_lo, vt10);
98 vt11 = _mm512_fmadd_ps(vn11, vminus_ln2_lo, vt11);
99
100 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
101 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
102 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
103 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
104 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
105 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
106 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
107 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
108 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
109 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
110 __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
111 __m512 vp10 = _mm512_fmadd_ps(vc5, vt10, vc4);
112 __m512 vp11 = _mm512_fmadd_ps(vc5, vt11, vc4);
113
114 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
115 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
116 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
117 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
118 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
119 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
120 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
121 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
122 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
123 vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
124 vp10 = _mm512_fmadd_ps(vp10, vt10, vc3);
125 vp11 = _mm512_fmadd_ps(vp11, vt11, vc3);
126
127 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
128 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
129 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
130 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
131 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
132 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
133 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
134 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
135 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
136 vp9 = _mm512_fmadd_ps(vp9, vt9, vc2);
137 vp10 = _mm512_fmadd_ps(vp10, vt10, vc2);
138 vp11 = _mm512_fmadd_ps(vp11, vt11, vc2);
139
140 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
141 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
142 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
143 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
144 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
145 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
146 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
147 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
148 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
149 vp9 = _mm512_fmadd_ps(vp9, vt9, vc1);
150 vp10 = _mm512_fmadd_ps(vp10, vt10, vc1);
151 vp11 = _mm512_fmadd_ps(vp11, vt11, vc1);
152
153 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
154 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
155 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
156 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
157 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
158 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
159 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
160 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
161 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
162 vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
163 vp10 = _mm512_fmadd_ps(vp10, vt10, vc0);
164 vp11 = _mm512_fmadd_ps(vp11, vt11, vc0);
165
166 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
167 // - vnX is "exponent"
168 // - vpX is "mantissa"
169 //
170 // exp2(ae) * av * exp2(be) * bv =
171 // = exp2(ae + be) * (av * bv)
172 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
173 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
174 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
175 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
176 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
177 __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
178 __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
179 __m512 vf7 = _mm512_mul_ps(vp7, vscalev);
180 __m512 vf8 = _mm512_mul_ps(vp8, vscalev);
181 __m512 vf9 = _mm512_mul_ps(vp9, vscalev);
182 __m512 vf10 = _mm512_mul_ps(vp10, vscalev);
183 __m512 vf11 = _mm512_mul_ps(vp11, vscalev);
184
185 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
186 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
187 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
188 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
189 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
190 const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
191 const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
192 const __m512 ve7 = _mm512_add_ps(vn7, vscalee);
193 const __m512 ve8 = _mm512_add_ps(vn8, vscalee);
194 const __m512 ve9 = _mm512_add_ps(vn9, vscalee);
195 const __m512 ve10 = _mm512_add_ps(vn10, vscalee);
196 const __m512 ve11 = _mm512_add_ps(vn11, vscalee);
197
198 // Multiply "mantissa" by the exp2("exponent").
199 vf0 = _mm512_scalef_ps(vf0, ve0);
200 vf1 = _mm512_scalef_ps(vf1, ve1);
201 vf2 = _mm512_scalef_ps(vf2, ve2);
202 vf3 = _mm512_scalef_ps(vf3, ve3);
203 vf4 = _mm512_scalef_ps(vf4, ve4);
204 vf5 = _mm512_scalef_ps(vf5, ve5);
205 vf6 = _mm512_scalef_ps(vf6, ve6);
206 vf7 = _mm512_scalef_ps(vf7, ve7);
207 vf8 = _mm512_scalef_ps(vf8, ve8);
208 vf9 = _mm512_scalef_ps(vf9, ve9);
209 vf10 = _mm512_scalef_ps(vf10, ve10);
210 vf11 = _mm512_scalef_ps(vf11, ve11);
211
212 // Store 128 (8x16) results at a time.
213 _mm512_storeu_ps(y, vf0);
214 _mm512_storeu_ps(y + 0, vf0);
215 _mm512_storeu_ps(y + 16, vf1);
216 _mm512_storeu_ps(y + 32, vf2);
217 _mm512_storeu_ps(y + 48, vf3);
218 _mm512_storeu_ps(y + 64, vf4);
219 _mm512_storeu_ps(y + 80, vf5);
220 _mm512_storeu_ps(y + 96, vf6);
221 _mm512_storeu_ps(y + 112, vf7);
222 _mm512_storeu_ps(y + 128, vf8);
223 _mm512_storeu_ps(y + 144, vf9);
224 _mm512_storeu_ps(y + 160, vf10);
225 _mm512_storeu_ps(y + 176, vf11);
226 y += 192;
227 }
228
229 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
230 // Load 16 inputs at a time.
231 const __m512 vx = _mm512_loadu_ps(x);
232 x += 16;
233
234 // Compute reduced argument elements := round(x / log(2)).
235 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
236
237 // Compute reduced argument t := x - elements * log(2).
238 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
239 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
240 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
241
242 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
243 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
244 vp = _mm512_fmadd_ps(vp, vt, vc3);
245 vp = _mm512_fmadd_ps(vp, vt, vc2);
246 vp = _mm512_fmadd_ps(vp, vt, vc1);
247 vp = _mm512_fmadd_ps(vp, vt, vc0);
248
249 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
250 __m512 vf = _mm512_mul_ps(vp, vscalev);
251 const __m512 ve = _mm512_add_ps(vn, vscalee);
252
253 // Multiply "mantissa" by the exp2("exponent").
254 vf = _mm512_scalef_ps(vf, ve);
255
256 // Store 16 results at a time.
257 _mm512_storeu_ps(y, vf);
258 y += 16;
259 }
260 if XNN_UNLIKELY(elements != 0) {
261 // Prepare mask for valid 32-bit elements (depends on elements).
262 elements >>= 2 /* log2(sizeof(float)) */;
263 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
264
265 // Load up to 15 inputs at a time.
266 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
267
268 // Compute reduced argument elements := round(x / log(2)).
269 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
270
271 // Compute reduced argument t := x - elements * log(2).
272 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
273 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
274 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
275
276 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
277 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
278 vp = _mm512_fmadd_ps(vp, vt, vc3);
279 vp = _mm512_fmadd_ps(vp, vt, vc2);
280 vp = _mm512_fmadd_ps(vp, vt, vc1);
281 vp = _mm512_fmadd_ps(vp, vt, vc0);
282
283 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
284 __m512 vf = _mm512_mul_ps(vp, vscalev);
285 const __m512 ve = _mm512_add_ps(vn, vscalee);
286
287 // Multiply "mantissa" by the exp2("exponent").
288 vf = _mm512_scalef_ps(vf, ve);
289
290 // Store up to 15 results at a time.
291 _mm512_mask_storeu_ps(y, vmask, vf);
292 }
293 _mm256_zeroupper();
294 }
295