1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddstoreexpminusmax/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/raddstoreexpminusmax.h>
16
17
xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4(size_t elements,const float * input,float * output,float * sum,float max)18 void xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4(
19 size_t elements,
20 const float* input,
21 float* output,
22 float* sum,
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 vi_max = _mm512_set1_ps(max);
39
40 __m512 vacc0 = _mm512_setzero_ps();
41 __m512 vacc1 = _mm512_setzero_ps();
42 __m512 vacc2 = _mm512_setzero_ps();
43 __m512 vacc3 = _mm512_setzero_ps();
44 for (; elements >= 128 * sizeof(float); elements -= 128 * sizeof(float)) {
45 // Load 128 (8x16) inputs at a time.
46 const __m512 vi0 = _mm512_loadu_ps(input);
47 const __m512 vi1 = _mm512_loadu_ps(input + 16);
48 const __m512 vi2 = _mm512_loadu_ps(input + 32);
49 const __m512 vi3 = _mm512_loadu_ps(input + 48);
50 const __m512 vi4 = _mm512_loadu_ps(input + 64);
51 const __m512 vi5 = _mm512_loadu_ps(input + 80);
52 const __m512 vi6 = _mm512_loadu_ps(input + 96);
53 const __m512 vi7 = _mm512_loadu_ps(input + 112);
54 input += 128;
55
56 // Subtract maximum input x := i - i_max.
57 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
58 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
59 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
60 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
61 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
62 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
63 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
64 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
65
66 // Compute reduced argument elements := round(x / log(2)).
67 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
68 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
69 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
70 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
71 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
72 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
73 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
74 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
75
76 // Compute reduced argument t := x - elements * log(2).
77 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
78 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
79 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
80 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
81 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
82 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
83 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
84 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
85 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
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
96 // Compute degree-5 polynomial approxiatmion 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
106 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
107 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
108 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
109 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
110 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
111 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
112 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
113 vp7 = _mm512_fmadd_ps(vp7, vt7, 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
124 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
125 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
126 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
127 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
128 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
129 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
130 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
131 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
132
133 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
134 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
135 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
136 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
137 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
138 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
139 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
140 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
141
142 // Reconstruct the final f value:
143 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
144 // = 2**elements * p
145 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
146 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
147 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
148 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
149 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
150 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
151 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
152 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
153
154 // Store 128 (8x16) outputs at a time.
155 _mm512_storeu_ps(output, vf0);
156 _mm512_storeu_ps(output + 16, vf1);
157 _mm512_storeu_ps(output + 32, vf2);
158 _mm512_storeu_ps(output + 48, vf3);
159 _mm512_storeu_ps(output + 64, vf4);
160 _mm512_storeu_ps(output + 80, vf5);
161 _mm512_storeu_ps(output + 96, vf6);
162 _mm512_storeu_ps(output + 112, vf7);
163 output += 128;
164
165 // Accumulate computed exponents.
166 vacc0 = _mm512_add_ps(vacc0, vf0);
167 vacc1 = _mm512_add_ps(vacc1, vf1);
168 vacc2 = _mm512_add_ps(vacc2, vf2);
169 vacc3 = _mm512_add_ps(vacc3, vf3);
170 vacc0 = _mm512_add_ps(vacc0, vf4);
171 vacc1 = _mm512_add_ps(vacc1, vf5);
172 vacc2 = _mm512_add_ps(vacc2, vf6);
173 vacc3 = _mm512_add_ps(vacc3, vf7);
174 }
175 // Add up all accumulators to vacc0
176 vacc0 = _mm512_add_ps(vacc0, vacc1);
177 vacc2 = _mm512_add_ps(vacc2, vacc3);
178 vacc0 = _mm512_add_ps(vacc0, vacc2);
179
180 __m512 vacc = vacc0;
181 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
182 // Load 16 inputs at a time.
183 const __m512 vi = _mm512_loadu_ps(input);
184 input += 16;
185
186 // Subtract maximum input x := i - i_max.
187 const __m512 vx = _mm512_sub_ps(vi, vi_max);
188
189 // Compute reduced argument elements := round(x / log(2)).
190 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
191
192 // Compute reduced argument t := x - elements * log(2).
193 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
194 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
195 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
196
197 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
198 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
199 vp = _mm512_fmadd_ps(vp, vt, vc3);
200 vp = _mm512_fmadd_ps(vp, vt, vc2);
201 vp = _mm512_fmadd_ps(vp, vt, vc1);
202 vp = _mm512_fmadd_ps(vp, vt, vc0);
203
204 // Reconstruct the final f value:
205 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
206 // = 2**elements * p
207 const __m512 vf = _mm512_scalef_ps(vp, vn);
208
209 // Store 16 outputs at a time.
210 _mm512_storeu_ps(output, vf);
211 output += 16;
212
213 // Accumulate computed exponents.
214 vacc = _mm512_add_ps(vacc, vf);
215 }
216 if (elements != 0) {
217 // Prepare mask for valid 32-bit elements (depends on elements).
218 elements >>= 2 /* log2(sizeof(float)) */;
219 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
220
221 // Load up to 15 inputs at a time.
222 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
223
224 // Subtract maximum input x := i - i_max.
225 const __m512 vx = _mm512_sub_ps(vi, vi_max);
226
227 // Compute reduced argument elements := round(x / log(2)).
228 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
229
230 // Compute reduced argument t := x - elements * log(2).
231 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
232 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
233 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
234
235 // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
236 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
237 vp = _mm512_fmadd_ps(vp, vt, vc3);
238 vp = _mm512_fmadd_ps(vp, vt, vc2);
239 vp = _mm512_fmadd_ps(vp, vt, vc1);
240 vp = _mm512_fmadd_ps(vp, vt, vc0);
241
242 // Reconstruct the final f value:
243 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
244 // = 2**elements * p
245 const __m512 vf = _mm512_scalef_ps(vp, vn);
246
247 // Store up to 15 outputs at a time.
248 _mm512_mask_storeu_ps(output, vmask, vf);
249
250 // Accumulate computed exponents.
251 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
252 }
253 *sum = _mm512_reduce_add_ps(vacc);
254 }
255