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_x144(size_t elements,const float * x,float * y,float scale_value,float scale_exp)19 void xnn_f32_vscaleextexp_ukernel__avx512f_p5_scalef_x144(
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 >= 144 * sizeof(float); elements -= 144 * sizeof(float)) {
43 // Load 144 (9x16) 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 x += 144;
54
55 // Compute reduced argument elements := round(x / log(2)).
56 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
57 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
58 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
59 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
60 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
61 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
62 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
63 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
64 const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
65
66 // Compute reduced argument t := x - elements * log(2).
67 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
68 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
69 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
70 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
71 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
72 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
73 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
74 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
75 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
76 __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
77
78 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
79 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
80 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
81 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
82 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
83 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
84 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
85 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
86 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
87
88 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
89 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
90 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
91 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
92 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
93 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
94 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
95 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
96 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
97 __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
98
99 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
100 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
101 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
102 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
103 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
104 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
105 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
106 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
107 vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
108
109 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
110 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
111 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
112 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
113 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
114 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
115 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
116 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
117 vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
118
119 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
120 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
121 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
122 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
123 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
124 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
125 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
126 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
127 vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
128
129 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
130 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
131 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
132 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
133 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
134 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
135 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
136 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
137 vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
138
139 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation where
140 // - vnX is "exponent"
141 // - vpX is "mantissa"
142 //
143 // exp2(ae) * av * exp2(be) * bv =
144 // = exp2(ae + be) * (av * bv)
145 __m512 vf0 = _mm512_mul_ps(vp0, vscalev);
146 __m512 vf1 = _mm512_mul_ps(vp1, vscalev);
147 __m512 vf2 = _mm512_mul_ps(vp2, vscalev);
148 __m512 vf3 = _mm512_mul_ps(vp3, vscalev);
149 __m512 vf4 = _mm512_mul_ps(vp4, vscalev);
150 __m512 vf5 = _mm512_mul_ps(vp5, vscalev);
151 __m512 vf6 = _mm512_mul_ps(vp6, vscalev);
152 __m512 vf7 = _mm512_mul_ps(vp7, vscalev);
153 __m512 vf8 = _mm512_mul_ps(vp8, vscalev);
154
155 const __m512 ve0 = _mm512_add_ps(vn0, vscalee);
156 const __m512 ve1 = _mm512_add_ps(vn1, vscalee);
157 const __m512 ve2 = _mm512_add_ps(vn2, vscalee);
158 const __m512 ve3 = _mm512_add_ps(vn3, vscalee);
159 const __m512 ve4 = _mm512_add_ps(vn4, vscalee);
160 const __m512 ve5 = _mm512_add_ps(vn5, vscalee);
161 const __m512 ve6 = _mm512_add_ps(vn6, vscalee);
162 const __m512 ve7 = _mm512_add_ps(vn7, vscalee);
163 const __m512 ve8 = _mm512_add_ps(vn8, vscalee);
164
165 // Multiply "mantissa" by the exp2("exponent").
166 vf0 = _mm512_scalef_ps(vf0, ve0);
167 vf1 = _mm512_scalef_ps(vf1, ve1);
168 vf2 = _mm512_scalef_ps(vf2, ve2);
169 vf3 = _mm512_scalef_ps(vf3, ve3);
170 vf4 = _mm512_scalef_ps(vf4, ve4);
171 vf5 = _mm512_scalef_ps(vf5, ve5);
172 vf6 = _mm512_scalef_ps(vf6, ve6);
173 vf7 = _mm512_scalef_ps(vf7, ve7);
174 vf8 = _mm512_scalef_ps(vf8, ve8);
175
176 // Store 128 (8x16) results at a time.
177 _mm512_storeu_ps(y, vf0);
178 _mm512_storeu_ps(y + 0, vf0);
179 _mm512_storeu_ps(y + 16, vf1);
180 _mm512_storeu_ps(y + 32, vf2);
181 _mm512_storeu_ps(y + 48, vf3);
182 _mm512_storeu_ps(y + 64, vf4);
183 _mm512_storeu_ps(y + 80, vf5);
184 _mm512_storeu_ps(y + 96, vf6);
185 _mm512_storeu_ps(y + 112, vf7);
186 _mm512_storeu_ps(y + 128, vf8);
187 y += 144;
188 }
189
190 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
191 // Load 16 inputs at a time.
192 const __m512 vx = _mm512_loadu_ps(x);
193 x += 16;
194
195 // Compute reduced argument elements := round(x / log(2)).
196 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
197
198 // Compute reduced argument t := x - elements * log(2).
199 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
200 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
201 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
202
203 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
204 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
205 vp = _mm512_fmadd_ps(vp, vt, vc3);
206 vp = _mm512_fmadd_ps(vp, vt, vc2);
207 vp = _mm512_fmadd_ps(vp, vt, vc1);
208 vp = _mm512_fmadd_ps(vp, vt, vc0);
209
210 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
211 __m512 vf = _mm512_mul_ps(vp, vscalev);
212 const __m512 ve = _mm512_add_ps(vn, vscalee);
213
214 // Multiply "mantissa" by the exp2("exponent").
215 vf = _mm512_scalef_ps(vf, ve);
216
217 // Store 16 results at a time.
218 _mm512_storeu_ps(y, vf);
219 y += 16;
220 }
221 if XNN_UNLIKELY(elements != 0) {
222 // Prepare mask for valid 32-bit elements (depends on elements).
223 elements >>= 2 /* log2(sizeof(float)) */;
224 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
225
226 // Load up to 15 inputs at a time.
227 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
228
229 // Compute reduced argument elements := round(x / log(2)).
230 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
231
232 // Compute reduced argument t := x - elements * log(2).
233 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
234 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
235 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
236
237 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
238 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
239 vp = _mm512_fmadd_ps(vp, vt, vc3);
240 vp = _mm512_fmadd_ps(vp, vt, vc2);
241 vp = _mm512_fmadd_ps(vp, vt, vc1);
242 vp = _mm512_fmadd_ps(vp, vt, vc0);
243
244 // Multiply "extended" floating-point numbers in ("mantissa", "exponent") representation.
245 __m512 vf = _mm512_mul_ps(vp, vscalev);
246 const __m512 ve = _mm512_add_ps(vn, vscalee);
247
248 // Multiply "mantissa" by the exp2("exponent").
249 vf = _mm512_scalef_ps(vf, ve);
250
251 // Store up to 15 results at a time.
252 _mm512_mask_storeu_ps(y, vmask, vf);
253 }
254 _mm256_zeroupper();
255 }
256