1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/avx512f-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_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112(size_t n,const float * x,float * y,const void * params)19 void xnn_f32_sigmoid_ukernel__avx512f_rr1_p5_scalef_div_x112(
20     size_t n,
21     const float* x,
22     float* y,
23     const void* params)
24 {
25   assert(n % sizeof(float) == 0);
26 
27   const __m512i vsign_mask = _mm512_set1_epi32(0x80000000);
28   const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
29   const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
30   const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f);
31   const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f);
32   const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f);
33   const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f);
34   const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f);
35   const __m512 vone = _mm512_set1_ps(1.0f);
36 
37   for (; n >= 112 * sizeof(float); n -= 112 * 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     x += 112;
46 
47     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
48     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
49     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
50     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
51     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
52     const __m512 vz5 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx5), vsign_mask));
53     const __m512 vz6 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx6), vsign_mask));
54 
55     __m512 vn0 = _mm512_mul_ps(vz0, vlog2e);
56     __m512 vn1 = _mm512_mul_ps(vz1, vlog2e);
57     __m512 vn2 = _mm512_mul_ps(vz2, vlog2e);
58     __m512 vn3 = _mm512_mul_ps(vz3, vlog2e);
59     __m512 vn4 = _mm512_mul_ps(vz4, vlog2e);
60     __m512 vn5 = _mm512_mul_ps(vz5, vlog2e);
61     __m512 vn6 = _mm512_mul_ps(vz6, vlog2e);
62 
63     vn0 = _mm512_roundscale_ps(vn0, 0);
64     vn1 = _mm512_roundscale_ps(vn1, 0);
65     vn2 = _mm512_roundscale_ps(vn2, 0);
66     vn3 = _mm512_roundscale_ps(vn3, 0);
67     vn4 = _mm512_roundscale_ps(vn4, 0);
68     vn5 = _mm512_roundscale_ps(vn5, 0);
69     vn6 = _mm512_roundscale_ps(vn6, 0);
70 
71     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
72     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
73     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
74     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
75     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
76     __m512 vt5 = _mm512_fmadd_ps(vn5, vminus_ln2, vz5);
77     __m512 vt6 = _mm512_fmadd_ps(vn6, vminus_ln2, vz6);
78 
79     __m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4);
80     __m512 vp1 = _mm512_fmadd_ps(vc5, vt1, vc4);
81     __m512 vp2 = _mm512_fmadd_ps(vc5, vt2, vc4);
82     __m512 vp3 = _mm512_fmadd_ps(vc5, vt3, vc4);
83     __m512 vp4 = _mm512_fmadd_ps(vc5, vt4, vc4);
84     __m512 vp5 = _mm512_fmadd_ps(vc5, vt5, vc4);
85     __m512 vp6 = _mm512_fmadd_ps(vc5, vt6, vc4);
86 
87     vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
88     vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
89     vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
90     vp3 = _mm512_fmadd_ps(vp3, vt3, vc3);
91     vp4 = _mm512_fmadd_ps(vp4, vt4, vc3);
92     vp5 = _mm512_fmadd_ps(vp5, vt5, vc3);
93     vp6 = _mm512_fmadd_ps(vp6, vt6, vc3);
94 
95     vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
96     vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
97     vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
98     vp3 = _mm512_fmadd_ps(vp3, vt3, vc2);
99     vp4 = _mm512_fmadd_ps(vp4, vt4, vc2);
100     vp5 = _mm512_fmadd_ps(vp5, vt5, vc2);
101     vp6 = _mm512_fmadd_ps(vp6, vt6, vc2);
102 
103     vp0 = _mm512_fmadd_ps(vp0, vt0, vc1);
104     vp1 = _mm512_fmadd_ps(vp1, vt1, vc1);
105     vp2 = _mm512_fmadd_ps(vp2, vt2, vc1);
106     vp3 = _mm512_fmadd_ps(vp3, vt3, vc1);
107     vp4 = _mm512_fmadd_ps(vp4, vt4, vc1);
108     vp5 = _mm512_fmadd_ps(vp5, vt5, vc1);
109     vp6 = _mm512_fmadd_ps(vp6, vt6, vc1);
110 
111     vp0 = _mm512_fmadd_ps(vp0, vt0, vone);
112     vp1 = _mm512_fmadd_ps(vp1, vt1, vone);
113     vp2 = _mm512_fmadd_ps(vp2, vt2, vone);
114     vp3 = _mm512_fmadd_ps(vp3, vt3, vone);
115     vp4 = _mm512_fmadd_ps(vp4, vt4, vone);
116     vp5 = _mm512_fmadd_ps(vp5, vt5, vone);
117     vp6 = _mm512_fmadd_ps(vp6, vt6, vone);
118 
119     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
120     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
121     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
122     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
123     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
124     const __m512 ve5 = _mm512_scalef_ps(vp5, vn5);
125     const __m512 ve6 = _mm512_scalef_ps(vp6, vn6);
126 
127     const __m512 vd0 = _mm512_add_ps(ve0, vone);
128     const __m512 vd1 = _mm512_add_ps(ve1, vone);
129     const __m512 vd2 = _mm512_add_ps(ve2, vone);
130     const __m512 vd3 = _mm512_add_ps(ve3, vone);
131     const __m512 vd4 = _mm512_add_ps(ve4, vone);
132     const __m512 vd5 = _mm512_add_ps(ve5, vone);
133     const __m512 vd6 = _mm512_add_ps(ve6, vone);
134 
135     __m512 vf0 = _mm512_div_ps(ve0, vd0);
136     __m512 vf1 = _mm512_div_ps(ve1, vd1);
137     __m512 vf2 = _mm512_div_ps(ve2, vd2);
138     __m512 vf3 = _mm512_div_ps(ve3, vd3);
139     __m512 vf4 = _mm512_div_ps(ve4, vd4);
140     __m512 vf5 = _mm512_div_ps(ve5, vd5);
141     __m512 vf6 = _mm512_div_ps(ve6, vd6);
142 
143     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
144     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
145     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
146     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
147     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
148     vf5 = _mm512_mask_sub_ps(vf5, _mm512_testn_epi32_mask(_mm512_castps_si512(vx5), vsign_mask), vone, vf5);
149     vf6 = _mm512_mask_sub_ps(vf6, _mm512_testn_epi32_mask(_mm512_castps_si512(vx6), vsign_mask), vone, vf6);
150 
151     _mm512_storeu_ps(y, vf0);
152     _mm512_storeu_ps(y + 16, vf1);
153     _mm512_storeu_ps(y + 32, vf2);
154     _mm512_storeu_ps(y + 48, vf3);
155     _mm512_storeu_ps(y + 64, vf4);
156     _mm512_storeu_ps(y + 80, vf5);
157     _mm512_storeu_ps(y + 96, vf6);
158     y += 112;
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 vf = _mm512_div_ps(ve, vd);
180 
181     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
182 
183     _mm512_storeu_ps(y, vf);
184     y += 16;
185   }
186   if XNN_UNLIKELY(n != 0) {
187     assert(n >= 1 * sizeof(float));
188     assert(n <= 15 * sizeof(float));
189 
190     // Prepare mask for valid 32-bit elements (depends on n).
191     n >>= 2 /* log2(sizeof(float)) */;
192     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
193 
194     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
195     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
196 
197     const __m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vz, vlog2e), 0);
198 
199     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
200 
201     __m512 vp = _mm512_fmadd_ps(vc5, vt, vc4);
202     vp = _mm512_fmadd_ps(vp, vt, vc3);
203     vp = _mm512_fmadd_ps(vp, vt, vc2);
204     vp = _mm512_fmadd_ps(vp, vt, vc1);
205     vp = _mm512_fmadd_ps(vp, vt, vone);
206 
207     const __m512 ve = _mm512_scalef_ps(vp, vn);
208     const __m512 vd = _mm512_add_ps(ve, vone);
209 
210     __m512 vf = _mm512_div_ps(ve, vd);
211 
212     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
213 
214     _mm512_mask_storeu_ps(y, vmask, vf);
215   }
216 }
217