xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-avx512f-rr1-p5-scalef-nr1fma-x96.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_x96(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_x96(
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 >= 96 * sizeof(float); n -= 96 * 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     x += 96;
45 
46     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
47     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
48     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
49     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
50     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
51     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
52 
53     __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
54     __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
55     __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
56     __m512 vn3 = _mm512_mul_ps(vz3, vlog2e);
57     __m512 vn4 = _mm512_mul_ps(vz4, vlog2e);
58     __m512 vn5 = _mm512_mul_ps(vz5, vlog2e);
59 
60     vn0 = _mm512_roundscale_ps(vn0, 0);
61     vn1 = _mm512_roundscale_ps(vn1, 0);
62     vn2 = _mm512_roundscale_ps(vn2, 0);
63     vn3 = _mm512_roundscale_ps(vn3, 0);
64     vn4 = _mm512_roundscale_ps(vn4, 0);
65     vn5 = _mm512_roundscale_ps(vn5, 0);
66 
67     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
68     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
69     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
70     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
71     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
72     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
73 
74     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
75     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
76     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
77     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
78     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
79     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
80 
81     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
82     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
83     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
84     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
85     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
86     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
87 
88     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
89     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
90     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
91     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
92     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
93     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
94 
95     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
96     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
97     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
98     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
99     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
100     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
101 
102     vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
103     vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
104     vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
105     vp3 = _mm512_fmadd_ps(vp3, vt3, vone);
106     vp4 = _mm512_fmadd_ps(vp4, vt4, vone);
107     vp5 = _mm512_fmadd_ps(vp5, vt5, vone);
108 
109     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
110     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
111     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
112     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
113     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
114     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
115 
116     const __m512 vd0 = _mm512_add_ps(ve0, vone);
117     const __m512 vd1 = _mm512_add_ps(ve1, vone);
118     const __m512 vd2 = _mm512_add_ps(ve2, vone);
119     const __m512 vd3 = _mm512_add_ps(ve3, vone);
120     const __m512 vd4 = _mm512_add_ps(ve4, vone);
121     const __m512 vd5 = _mm512_add_ps(ve5, vone);
122 
123     __m512 vr0 = _mm512_rcp14_ps(vd0);
124     __m512 vr1 = _mm512_rcp14_ps(vd1);
125     __m512 vr2 = _mm512_rcp14_ps(vd2);
126     __m512 vr3 = _mm512_rcp14_ps(vd3);
127     __m512 vr4 = _mm512_rcp14_ps(vd4);
128     __m512 vr5 = _mm512_rcp14_ps(vd5);
129 
130     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
131     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
132     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
133     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
134     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
135     vr5 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
136 
137     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
138     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
139     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
140     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
141     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
142     __m512 vf5 = _mm512_mul_ps(ve5, vr5);
143 
144 
145     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
146     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
147     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
148     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
149     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
150     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
151 
152     _mm512_storeu_ps(y, vf0);
153     _mm512_storeu_ps(y + 16, vf1);
154     _mm512_storeu_ps(y + 32, vf2);
155     _mm512_storeu_ps(y + 48, vf3);
156     _mm512_storeu_ps(y + 64, vf4);
157     _mm512_storeu_ps(y + 80, vf5);
158     y += 96;
159   }
160   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
161     const __m512 vx = _mm512_loadu_ps(x);
162     x += 16;
163 
164     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
165 
166     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
167 
168     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
169 
170     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
171     vp = _mm512_fmadd_ps(vp, vt, vc3);
172     vp = _mm512_fmadd_ps(vp, vt, vc2);
173     vp = _mm512_fmadd_ps(vp, vt, vc1);
174     vp = _mm512_fmadd_ps(vp, vt, vone);
175 
176     const __m512 ve = _mm512_scalef_ps(vp, vn);
177     const __m512 vd = _mm512_add_ps(ve, vone);
178 
179     __m512 vr = _mm512_rcp14_ps(vd);
180     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
181 
182     __m512 vf = _mm512_mul_ps(ve, vr);
183 
184     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
185 
186     _mm512_storeu_ps(y, vf);
187     y += 16;
188   }
189   if XNN_UNLIKELY(n != 0) {
190     assert(n >= 1 * sizeof(float));
191     assert(n <= 15 * sizeof(float));
192 
193     // Prepare mask for valid 32-bit elements (depends on n).
194     n >>= 2 /* log2(sizeof(float)) */;
195     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
196 
197     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
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_mask_storeu_ps(y, vmask, vf);
221   }
222 }
223