• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x160_acc5(size_t elements,const float * input,float * sum,float max)18 void xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x160_acc5(
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   __m512 vacc4 = _mm512_setzero_ps();
44   for (; elements >= 160 * sizeof(float); elements -= 160 * sizeof(float)) {
45     // Load 160 (10x16) 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     const __m512 vi8 = _mm512_loadu_ps(input + 128);
55     const __m512 vi9 = _mm512_loadu_ps(input + 144);
56     input += 160;
57 
58     // Subtract maximum input x := i - i_max.
59     const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
60     const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
61     const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
62     const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
63     const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
64     const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
65     const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
66     const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
67     const __m512 vx8 = _mm512_sub_ps(vi8, vi_max);
68     const __m512 vx9 = _mm512_sub_ps(vi9, vi_max);
69 
70     // Compute reduced argument elements := round(x / log(2)).
71     const __m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0);
72     const __m512 vn1 = _mm512_roundscale_ps(_mm512_mul_ps(vx1, vlog2e), 0);
73     const __m512 vn2 = _mm512_roundscale_ps(_mm512_mul_ps(vx2, vlog2e), 0);
74     const __m512 vn3 = _mm512_roundscale_ps(_mm512_mul_ps(vx3, vlog2e), 0);
75     const __m512 vn4 = _mm512_roundscale_ps(_mm512_mul_ps(vx4, vlog2e), 0);
76     const __m512 vn5 = _mm512_roundscale_ps(_mm512_mul_ps(vx5, vlog2e), 0);
77     const __m512 vn6 = _mm512_roundscale_ps(_mm512_mul_ps(vx6, vlog2e), 0);
78     const __m512 vn7 = _mm512_roundscale_ps(_mm512_mul_ps(vx7, vlog2e), 0);
79     const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, vlog2e), 0);
80     const __m512 vn9 = _mm512_roundscale_ps(_mm512_mul_ps(vx9, vlog2e), 0);
81 
82     // Compute reduced argument t := x - elements * log(2).
83     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
84     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0);
85     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vx1);
86     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vx2);
87     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vx3);
88     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vx4);
89     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_hi, vx5);
90     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_hi, vx6);
91     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_hi, vx7);
92     __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
93     __m512 vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_hi, vx9);
94 
95     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
96     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
97     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
98     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
99     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
100     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
101     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
102     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
103     vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
104     vt9 = _mm512_fmadd_ps(vn9, vminus_ln2_lo, vt9);
105 
106     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
107     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
108     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
109     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
110     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
111     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
112     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
113     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
114     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
115     __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
116     __m512 vp9 = _mm512_fmadd_ps(vc5, vt9, vc4);
117 
118     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
119     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
120     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
121     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
122     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
123     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
124     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
125     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
126     vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
127     vp9 = _mm512_fmadd_ps(vp9, vt9, vc3);
128 
129     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
130     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
131     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
132     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
133     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
134     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
135     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
136     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
137     vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
138     vp9 = _mm512_fmadd_ps(vp9, vt9, 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 
151     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
152     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
153     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
154     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
155     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
156     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
157     vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
158     vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
159     vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
160     vp9 = _mm512_fmadd_ps(vp9, vt9, vc0);
161 
162     // Reconstruct the final f value:
163     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
164     //     = 2**elements * p
165     const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
166     const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
167     const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
168     const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
169     const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
170     const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
171     const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
172     const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
173     const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
174     const __m512 vf9 = _mm512_scalef_ps(vp9, vn9);
175 
176     // Accumulate computed exponents.
177     vacc0 = _mm512_add_ps(vacc0, vf0);
178     vacc1 = _mm512_add_ps(vacc1, vf1);
179     vacc2 = _mm512_add_ps(vacc2, vf2);
180     vacc3 = _mm512_add_ps(vacc3, vf3);
181     vacc4 = _mm512_add_ps(vacc4, vf4);
182     vacc0 = _mm512_add_ps(vacc0, vf5);
183     vacc1 = _mm512_add_ps(vacc1, vf6);
184     vacc2 = _mm512_add_ps(vacc2, vf7);
185     vacc3 = _mm512_add_ps(vacc3, vf8);
186     vacc4 = _mm512_add_ps(vacc4, vf9);
187   }
188   // Add up all accumulators to vacc0
189   vacc0 = _mm512_add_ps(vacc0, vacc1);
190   vacc2 = _mm512_add_ps(vacc2, vacc3);
191   vacc0 = _mm512_add_ps(vacc0, vacc2);
192   vacc0 = _mm512_add_ps(vacc0, vacc4);
193 
194   __m512 vacc = vacc0;
195   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
196     // Load 16 inputs at a time.
197     const __m512 vi = _mm512_loadu_ps(input);
198     input += 16;
199 
200     // Subtract maximum input x := i - i_max.
201     const __m512 vx = _mm512_sub_ps(vi, vi_max);
202 
203     // Compute reduced argument elements := round(x / log(2)).
204     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
205 
206     // Compute reduced argument t := x - elements * log(2).
207     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
208     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
209     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
210 
211     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
212     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
213     vp = _mm512_fmadd_ps(vp, vt, vc3);
214     vp = _mm512_fmadd_ps(vp, vt, vc2);
215     vp = _mm512_fmadd_ps(vp, vt, vc1);
216     vp = _mm512_fmadd_ps(vp, vt, vc0);
217 
218     // Reconstruct the final f value:
219     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
220     //     = 2**elements * p
221     const __m512 vf = _mm512_scalef_ps(vp, vn);
222 
223     // Accumulate computed exponents.
224     vacc = _mm512_add_ps(vacc, vf);
225   }
226   if (elements != 0) {
227     // Prepare mask for valid 32-bit elements (depends on elements).
228     elements >>= 2 /* log2(sizeof(float)) */;
229     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
230 
231     // Load up to 15 inputs at a time.
232     const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
233 
234     // Subtract maximum input x := i - i_max.
235     const __m512 vx = _mm512_sub_ps(vi, vi_max);
236 
237     // Compute reduced argument elements := round(x / log(2)).
238     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
239 
240     // Compute reduced argument t := x - elements * log(2).
241     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
242     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
243     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
244 
245     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
246     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
247     vp = _mm512_fmadd_ps(vp, vt, vc3);
248     vp = _mm512_fmadd_ps(vp, vt, vc2);
249     vp = _mm512_fmadd_ps(vp, vt, vc1);
250     vp = _mm512_fmadd_ps(vp, vt, vc0);
251 
252     // Reconstruct the final f value:
253     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
254     //     = 2**elements * p
255     const __m512 vf = _mm512_scalef_ps(vp, vn);
256 
257     // Accumulate computed exponents.
258     vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
259   }
260   *sum = _mm512_reduce_add_ps(vacc);
261 }
262