1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddexpminusmax/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/raddexpminusmax.h>
16
17
xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4(size_t elements,const float * input,float * sum,float max)18 void xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x128_acc4(
19 size_t elements,
20 const float* input,
21 float* sum,
22 float max)
23 {
24 assert(elements % sizeof(float) == 0);
25
26 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
27 const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f);
28 const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f);
29
30 const __m512 vc0 = _mm512_set1_ps(1.0f);
31 const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
32 const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
33 const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
34 const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
35 const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
36
37 const __m512 vi_max = _mm512_set1_ps(max);
38
39 __m512 vacc0 = _mm512_setzero_ps();
40 __m512 vacc1 = _mm512_setzero_ps();
41 __m512 vacc2 = _mm512_setzero_ps();
42 __m512 vacc3 = _mm512_setzero_ps();
43 for (; elements >= 128 * sizeof(float); elements -= 128 * sizeof(float)) {
44 // Load 128 (8x16) inputs at a time.
45 const __m512 vi0 = _mm512_loadu_ps(input);
46 const __m512 vi1 = _mm512_loadu_ps(input + 16);
47 const __m512 vi2 = _mm512_loadu_ps(input + 32);
48 const __m512 vi3 = _mm512_loadu_ps(input + 48);
49 const __m512 vi4 = _mm512_loadu_ps(input + 64);
50 const __m512 vi5 = _mm512_loadu_ps(input + 80);
51 const __m512 vi6 = _mm512_loadu_ps(input + 96);
52 const __m512 vi7 = _mm512_loadu_ps(input + 112);
53 input += 128;
54
55 // Subtract maximum input x := i - i_max.
56 const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
57 const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
58 const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
59 const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
60 const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
61 const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
62 const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
63 const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
64
65 // Compute reduced argument elements := round(x / log(2)).
66 const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
67 const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
68 const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
69 const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
70 const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
71 const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
72 const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
73 const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
74
75 // Compute reduced argument t := x - elements * log(2).
76 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
77 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
78 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
79 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
80 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
81 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
82 __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
83 __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
84 __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
85
86 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
87 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
88 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
89 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
90 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
91 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
92 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
93 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
94
95 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
96 __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
97 __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
98 __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
99 __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
100 __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
101 __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
102 __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
103 __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
104
105 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
106 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
107 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
108 vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
109 vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
110 vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
111 vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
112 vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
113
114 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
115 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
116 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
117 vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
118 vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
119 vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
120 vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
121 vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
122
123 vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
124 vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
125 vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
126 vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
127 vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
128 vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
129 vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
130 vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
131
132 vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
133 vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
134 vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
135 vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
136 vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
137 vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
138 vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
139 vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
140
141 // Reconstruct the final f value:
142 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
143 // = 2**elements * p
144 const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
145 const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
146 const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
147 const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
148 const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
149 const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
150 const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
151 const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
152
153 // Accumulate computed exponents.
154 vacc0 = _mm512_add_ps(vacc0, vf0);
155 vacc1 = _mm512_add_ps(vacc1, vf1);
156 vacc2 = _mm512_add_ps(vacc2, vf2);
157 vacc3 = _mm512_add_ps(vacc3, vf3);
158 vacc0 = _mm512_add_ps(vacc0, vf4);
159 vacc1 = _mm512_add_ps(vacc1, vf5);
160 vacc2 = _mm512_add_ps(vacc2, vf6);
161 vacc3 = _mm512_add_ps(vacc3, vf7);
162 }
163 // Add up all accumulators to vacc0
164 vacc0 = _mm512_add_ps(vacc0, vacc1);
165 vacc2 = _mm512_add_ps(vacc2, vacc3);
166 vacc0 = _mm512_add_ps(vacc0, vacc2);
167
168 __m512 vacc = vacc0;
169 for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
170 // Load 16 inputs at a time.
171 const __m512 vi = _mm512_loadu_ps(input);
172 input += 16;
173
174 // Subtract maximum input x := i - i_max.
175 const __m512 vx = _mm512_sub_ps(vi, vi_max);
176
177 // Compute reduced argument elements := round(x / log(2)).
178 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
179
180 // Compute reduced argument t := x - elements * log(2).
181 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
182 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
183 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
184
185 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
186 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
187 vp = _mm512_fmadd_ps(vp, vt, vc3);
188 vp = _mm512_fmadd_ps(vp, vt, vc2);
189 vp = _mm512_fmadd_ps(vp, vt, vc1);
190 vp = _mm512_fmadd_ps(vp, vt, vc0);
191
192 // Reconstruct the final f value:
193 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
194 // = 2**elements * p
195 const __m512 vf = _mm512_scalef_ps(vp, vn);
196
197 // Accumulate computed exponents.
198 vacc = _mm512_add_ps(vacc, vf);
199 }
200 if (elements != 0) {
201 // Prepare mask for valid 32-bit elements (depends on elements).
202 elements >>= 2 /* log2(sizeof(float)) */;
203 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
204
205 // Load up to 15 inputs at a time.
206 const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
207
208 // Subtract maximum input x := i - i_max.
209 const __m512 vx = _mm512_sub_ps(vi, vi_max);
210
211 // Compute reduced argument elements := round(x / log(2)).
212 const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
213
214 // Compute reduced argument t := x - elements * log(2).
215 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
216 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
217 vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
218
219 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
220 __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
221 vp = _mm512_fmadd_ps(vp, vt, vc3);
222 vp = _mm512_fmadd_ps(vp, vt, vc2);
223 vp = _mm512_fmadd_ps(vp, vt, vc1);
224 vp = _mm512_fmadd_ps(vp, vt, vc0);
225
226 // Reconstruct the final f value:
227 // f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
228 // = 2**elements * p
229 const __m512 vf = _mm512_scalef_ps(vp, vn);
230
231 // Accumulate computed exponents.
232 vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
233 }
234 *sum = _mm512_reduce_add_ps(vacc);
235 }
236