xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx512f-rr1-p5-scalef.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/vunary.h>
17 
18 
xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vsigmoid_ukernel__avx512f_rr1_p5_scalef_nr1fma_x128(
20     size_t n,
21     const float* x,
22     float* y,
23     const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(n % sizeof(float) == 0);
26 
27   const __m512i vsign_mask = _mm512_set1_epi32((int) params->avx512_rr1_p5.sign_mask);
28   const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_p5.log2e);
29   const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_p5.minus_ln2);
30   const __m512 vc5 = _mm512_set1_ps(params->avx512_rr1_p5.c5);
31   const __m512 vc4 = _mm512_set1_ps(params->avx512_rr1_p5.c4);
32   const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_p5.c3);
33   const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_p5.c2);
34   const __m512 vc1 = _mm512_set1_ps(params->avx512_rr1_p5.c1);
35   const __m512 vone = _mm512_set1_ps(params->avx512_rr1_p5.one);
36 
37   for (; n >= 128 * sizeof(float); n -= 128 * sizeof(float)) {
38     const __m512 vx0 = _mm512_loadu_ps(x);
39     const __m512 vx1 = _mm512_loadu_ps(x + 16);
40     const __m512 vx2 = _mm512_loadu_ps(x + 32);
41     const __m512 vx3 = _mm512_loadu_ps(x + 48);
42     const __m512 vx4 = _mm512_loadu_ps(x + 64);
43     const __m512 vx5 = _mm512_loadu_ps(x + 80);
44     const __m512 vx6 = _mm512_loadu_ps(x + 96);
45     const __m512 vx7 = _mm512_loadu_ps(x + 112);
46     x += 128;
47 
48     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
49     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
50     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
51     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
52     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
53     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
54     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
55     const __m512 vz7 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx7), vsign_mask));
56 
57     __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
58     __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
59     __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
60     __m512 vn3 = _mm512_mul_ps(vz3, vlog2e);
61     __m512 vn4 = _mm512_mul_ps(vz4, vlog2e);
62     __m512 vn5 = _mm512_mul_ps(vz5, vlog2e);
63     __m512 vn6 = _mm512_mul_ps(vz6, vlog2e);
64     __m512 vn7 = _mm512_mul_ps(vz7, vlog2e);
65 
66     vn0 = _mm512_roundscale_ps(vn0, 0);
67     vn1 = _mm512_roundscale_ps(vn1, 0);
68     vn2 = _mm512_roundscale_ps(vn2, 0);
69     vn3 = _mm512_roundscale_ps(vn3, 0);
70     vn4 = _mm512_roundscale_ps(vn4, 0);
71     vn5 = _mm512_roundscale_ps(vn5, 0);
72     vn6 = _mm512_roundscale_ps(vn6, 0);
73     vn7 = _mm512_roundscale_ps(vn7, 0);
74 
75     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
76     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
77     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
78     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
79     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
80     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
81     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
82     __m512 vt7 = _mm512_fmadd_ps(vn7, vminus_ln2, vz7);
83 
84     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
85     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
86     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
87     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
88     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
89     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
90     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
91     __m512 vp7 = _mm512_fmadd_ps(vc5, vt7, vc4);
92 
93     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
94     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
95     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
96     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
97     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
98     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
99     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
100     vp7 = _mm512_fmadd_ps(vp7, vt7, vc3);
101 
102     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
103     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
104     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
105     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
106     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
107     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
108     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
109     vp7 = _mm512_fmadd_ps(vp7, vt7, vc2);
110 
111     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
112     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
113     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
114     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
115     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
116     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
117     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
118     vp7 = _mm512_fmadd_ps(vp7, vt7, vc1);
119 
120     vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
121     vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
122     vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
123     vp3 = _mm512_fmadd_ps(vp3, vt3, vone);
124     vp4 = _mm512_fmadd_ps(vp4, vt4, vone);
125     vp5 = _mm512_fmadd_ps(vp5, vt5, vone);
126     vp6 = _mm512_fmadd_ps(vp6, vt6, vone);
127     vp7 = _mm512_fmadd_ps(vp7, vt7, vone);
128 
129     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
130     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
131     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
132     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
133     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
134     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
135     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
136     const __m512 ve7 = _mm512_scalef_ps(vp7, vn7);
137 
138     const __m512 vd0 = _mm512_add_ps(ve0, vone);
139     const __m512 vd1 = _mm512_add_ps(ve1, vone);
140     const __m512 vd2 = _mm512_add_ps(ve2, vone);
141     const __m512 vd3 = _mm512_add_ps(ve3, vone);
142     const __m512 vd4 = _mm512_add_ps(ve4, vone);
143     const __m512 vd5 = _mm512_add_ps(ve5, vone);
144     const __m512 vd6 = _mm512_add_ps(ve6, vone);
145     const __m512 vd7 = _mm512_add_ps(ve7, vone);
146 
147     __m512 vr0 = _mm512_rcp14_ps(vd0);
148     __m512 vr1 = _mm512_rcp14_ps(vd1);
149     __m512 vr2 = _mm512_rcp14_ps(vd2);
150     __m512 vr3 = _mm512_rcp14_ps(vd3);
151     __m512 vr4 = _mm512_rcp14_ps(vd4);
152     __m512 vr5 = _mm512_rcp14_ps(vd5);
153     __m512 vr6 = _mm512_rcp14_ps(vd6);
154     __m512 vr7 = _mm512_rcp14_ps(vd7);
155 
156     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
157     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
158     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
159     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
160     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
161     vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
162     vr6 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
163     vr7 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr7, vd7, vone), vr7, vr7);
164 
165     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
166     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
167     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
168     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
169     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
170     __m512 vf5 = _mm512_mul_ps(ve5, vr5);
171     __m512 vf6 = _mm512_mul_ps(ve6, vr6);
172     __m512 vf7 = _mm512_mul_ps(ve7, vr7);
173 
174 
175     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
176     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
177     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
178     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
179     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
180     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
181     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
182     vf7 = _mm512_mask_sub_ps(vf7, _mm512_testn_epi32_mask(_mm512_castps_si512(vx7), vsign_mask), vone, vf7);
183 
184     _mm512_storeu_ps(y, vf0);
185     _mm512_storeu_ps(y + 16, vf1);
186     _mm512_storeu_ps(y + 32, vf2);
187     _mm512_storeu_ps(y + 48, vf3);
188     _mm512_storeu_ps(y + 64, vf4);
189     _mm512_storeu_ps(y + 80, vf5);
190     _mm512_storeu_ps(y + 96, vf6);
191     _mm512_storeu_ps(y + 112, vf7);
192     y += 128;
193   }
194   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
195     const __m512 vx = _mm512_loadu_ps(x);
196     x += 16;
197 
198     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
199 
200     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
201 
202     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
203 
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, vone);
209 
210     const __m512 ve = _mm512_scalef_ps(vp, vn);
211     const __m512 vd = _mm512_add_ps(ve, vone);
212 
213     __m512 vr = _mm512_rcp14_ps(vd);
214     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
215 
216     __m512 vf = _mm512_mul_ps(ve, vr);
217 
218     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
219 
220     _mm512_storeu_ps(y, vf);
221     y += 16;
222   }
223   if XNN_UNLIKELY(n != 0) {
224     assert(n >= 1 * sizeof(float));
225     assert(n <= 15 * sizeof(float));
226 
227     // Prepare mask for valid 32-bit elements (depends on n).
228     n >>= 2 /* log2(sizeof(float)) */;
229     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
230 
231     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
232     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
233 
234     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
235 
236     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
237 
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, vone);
243 
244     const __m512 ve = _mm512_scalef_ps(vp, vn);
245     const __m512 vd = _mm512_add_ps(ve, vone);
246 
247     __m512 vr = _mm512_rcp14_ps(vd);
248     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
249 
250     __m512 vf = _mm512_mul_ps(ve, vr);
251 
252     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
253 
254     _mm512_mask_storeu_ps(y, vmask, vf);
255   }
256 }
257