• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_x144(size_t elements,const float * input,float * output,float * sum,float max)18 void xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x144(
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   for (; elements >= 144 * sizeof(float); elements -= 144 * sizeof(float)) {
42     // Load 144 (9x16) inputs at a time.
43     const __m512 vi0 = _mm512_loadu_ps(input);
44     const __m512 vi1 = _mm512_loadu_ps(input + 16);
45     const __m512 vi2 = _mm512_loadu_ps(input + 32);
46     const __m512 vi3 = _mm512_loadu_ps(input + 48);
47     const __m512 vi4 = _mm512_loadu_ps(input + 64);
48     const __m512 vi5 = _mm512_loadu_ps(input + 80);
49     const __m512 vi6 = _mm512_loadu_ps(input + 96);
50     const __m512 vi7 = _mm512_loadu_ps(input + 112);
51     const __m512 vi8 = _mm512_loadu_ps(input + 128);
52     input += 144;
53 
54     // Subtract maximum input x := i - i_max.
55     const __m512 vx0 = _mm512_sub_ps(vi0, vi_max);
56     const __m512 vx1 = _mm512_sub_ps(vi1, vi_max);
57     const __m512 vx2 = _mm512_sub_ps(vi2, vi_max);
58     const __m512 vx3 = _mm512_sub_ps(vi3, vi_max);
59     const __m512 vx4 = _mm512_sub_ps(vi4, vi_max);
60     const __m512 vx5 = _mm512_sub_ps(vi5, vi_max);
61     const __m512 vx6 = _mm512_sub_ps(vi6, vi_max);
62     const __m512 vx7 = _mm512_sub_ps(vi7, vi_max);
63     const __m512 vx8 = _mm512_sub_ps(vi8, 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     const __m512 vn8 = _mm512_roundscale_ps(_mm512_mul_ps(vx8, 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     __m512 vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_hi, vx8);
87 
88     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
89     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
90     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
91     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
92     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
93     vt5 = _mm512_fmadd_ps(vn5, vminus_ln2_lo, vt5);
94     vt6 = _mm512_fmadd_ps(vn6, vminus_ln2_lo, vt6);
95     vt7 = _mm512_fmadd_ps(vn7, vminus_ln2_lo, vt7);
96     vt8 = _mm512_fmadd_ps(vn8, vminus_ln2_lo, vt8);
97 
98     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
99     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
100     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
101     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
102     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
103     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
104     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
105     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
106     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
107     __m512 vp8 = _mm512_fmadd_ps(vc5, vt8, vc4);
108 
109     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
110     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
111     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
112     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
113     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
114     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
115     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
116     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
117     vp8 = _mm512_fmadd_ps(vp8, vt8, vc3);
118 
119     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
120     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
121     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
122     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
123     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
124     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
125     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
126     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
127     vp8 = _mm512_fmadd_ps(vp8, vt8, vc2);
128 
129     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
130     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
131     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
132     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
133     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
134     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
135     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
136     vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
137     vp8 = _mm512_fmadd_ps(vp8, vt8, vc1);
138 
139     vp0 = _mm512_fmadd_ps(vp0, vt0, vc0);
140     vp1 = _mm512_fmadd_ps(vp1, vt1, vc0);
141     vp2 = _mm512_fmadd_ps(vp2, vt2, vc0);
142     vp3 = _mm512_fmadd_ps(vp3, vt3, vc0);
143     vp4 = _mm512_fmadd_ps(vp4, vt4, vc0);
144     vp5 = _mm512_fmadd_ps(vp5, vt5, vc0);
145     vp6 = _mm512_fmadd_ps(vp6, vt6, vc0);
146     vp7 = _mm512_fmadd_ps(vp7, vt7, vc0);
147     vp8 = _mm512_fmadd_ps(vp8, vt8, vc0);
148 
149     // Reconstruct the final f value:
150     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
151     //     = 2**elements * p
152     const __m512 vf0 = _mm512_scalef_ps(vp0, vn0);
153     const __m512 vf1 = _mm512_scalef_ps(vp1, vn1);
154     const __m512 vf2 = _mm512_scalef_ps(vp2, vn2);
155     const __m512 vf3 = _mm512_scalef_ps(vp3, vn3);
156     const __m512 vf4 = _mm512_scalef_ps(vp4, vn4);
157     const __m512 vf5 = _mm512_scalef_ps(vp5, vn5);
158     const __m512 vf6 = _mm512_scalef_ps(vp6, vn6);
159     const __m512 vf7 = _mm512_scalef_ps(vp7, vn7);
160     const __m512 vf8 = _mm512_scalef_ps(vp8, vn8);
161 
162     // Store 144 (9x16) outputs at a time.
163     _mm512_storeu_ps(output, vf0);
164     _mm512_storeu_ps(output + 16, vf1);
165     _mm512_storeu_ps(output + 32, vf2);
166     _mm512_storeu_ps(output + 48, vf3);
167     _mm512_storeu_ps(output + 64, vf4);
168     _mm512_storeu_ps(output + 80, vf5);
169     _mm512_storeu_ps(output + 96, vf6);
170     _mm512_storeu_ps(output + 112, vf7);
171     _mm512_storeu_ps(output + 128, vf8);
172     output += 144;
173 
174     // Accumulate computed exponents.
175     vacc0 = _mm512_add_ps(vacc0, vf0);
176     vacc0 = _mm512_add_ps(vacc0, vf1);
177     vacc0 = _mm512_add_ps(vacc0, vf2);
178     vacc0 = _mm512_add_ps(vacc0, vf3);
179     vacc0 = _mm512_add_ps(vacc0, vf4);
180     vacc0 = _mm512_add_ps(vacc0, vf5);
181     vacc0 = _mm512_add_ps(vacc0, vf6);
182     vacc0 = _mm512_add_ps(vacc0, vf7);
183     vacc0 = _mm512_add_ps(vacc0, vf8);
184   }
185 
186   __m512 vacc = vacc0;
187   for (; elements >= 16 * sizeof(float); elements -= 16 * sizeof(float)) {
188     // Load 16 inputs at a time.
189     const __m512 vi = _mm512_loadu_ps(input);
190     input += 16;
191 
192     // Subtract maximum input x := i - i_max.
193     const __m512 vx = _mm512_sub_ps(vi, vi_max);
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     // Reconstruct the final f value:
211     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
212     //     = 2**elements * p
213     const __m512 vf = _mm512_scalef_ps(vp, vn);
214 
215     // Store 16 outputs at a time.
216     _mm512_storeu_ps(output, vf);
217     output += 16;
218 
219     // Accumulate computed exponents.
220     vacc = _mm512_add_ps(vacc, vf);
221   }
222   if (elements != 0) {
223     // Prepare mask for valid 32-bit elements (depends on elements).
224     elements >>= 2 /* log2(sizeof(float)) */;
225     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << elements) - UINT32_C(1)));
226 
227     // Load up to 15 inputs at a time.
228     const __m512 vi = _mm512_maskz_loadu_ps(vmask, input);
229 
230     // Subtract maximum input x := i - i_max.
231     const __m512 vx = _mm512_sub_ps(vi, vi_max);
232 
233     // Compute reduced argument elements := round(x / log(2)).
234     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0);
235 
236     // Compute reduced argument t := x - elements * log(2).
237     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
238     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx);
239     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
240 
241     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
242     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
243     vp = _mm512_fmadd_ps(vp, vt, vc3);
244     vp = _mm512_fmadd_ps(vp, vt, vc2);
245     vp = _mm512_fmadd_ps(vp, vt, vc1);
246     vp = _mm512_fmadd_ps(vp, vt, vc0);
247 
248     // Reconstruct the final f value:
249     //   f = 2**elements * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
250     //     = 2**elements * p
251     const __m512 vf = _mm512_scalef_ps(vp, vn);
252 
253     // Store up to 15 outputs at a time.
254     _mm512_mask_storeu_ps(output, vmask, vf);
255 
256     // Accumulate computed exponents.
257     vacc = _mm512_mask_add_ps(vacc, vmask, vacc, vf);
258   }
259   *sum = _mm512_reduce_add_ps(vacc);
260 }
261