1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx512f-rr1-lut16-p3-perm-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_lut16_p3_perm_scalef_nr1fma_x80(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_lut16_p3_perm_scalef_nr1fma_x80(
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_lut16_p3.sign_mask);
28   const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr1_lut16_p3.magic_bias);
29   const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr1_lut16_p3.log2e);
30   const __m512 vtable = _mm512_load_ps(params->avx512_rr1_lut16_p3.table);
31   const __m512 vminus_ln2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.minus_ln2);
32   const __m512 vc3 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c3);
33   const __m512 vc2 = _mm512_set1_ps(params->avx512_rr1_lut16_p3.c2);
34   const __m512 vone = _mm512_set1_ps(params->avx512_rr1_lut16_p3.one);
35 
36   for (; n >= 80 * sizeof(float); n -= 80 * sizeof(float)) {
37     const __m512 vx0 = _mm512_loadu_ps(x);
38     const __m512 vx1 = _mm512_loadu_ps(x + 16);
39     const __m512 vx2 = _mm512_loadu_ps(x + 32);
40     const __m512 vx3 = _mm512_loadu_ps(x + 48);
41     const __m512 vx4 = _mm512_loadu_ps(x + 64);
42     x += 80;
43 
44     const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
45     const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
46     const __m512 vz2 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx2), vsign_mask));
47     const __m512 vz3 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx3), vsign_mask));
48     const __m512 vz4 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx4), vsign_mask));
49 
50     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
51     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
52     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
53     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
54     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
55 
56     const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
57     const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
58     const __m512 vl2 = _mm512_permutexvar_ps(_mm512_castps_si512(vn2), vtable);
59     const __m512 vl3 = _mm512_permutexvar_ps(_mm512_castps_si512(vn3), vtable);
60     const __m512 vl4 = _mm512_permutexvar_ps(_mm512_castps_si512(vn4), vtable);
61 
62     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
63     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
64     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
65     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
66     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
67 
68     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
69     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
70     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
71     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
72     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
73 
74     __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
75     __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
76     __m512 vp2 = _mm512_fmadd_ps(vt2, vc3, vc2);
77     __m512 vp3 = _mm512_fmadd_ps(vt3, vc3, vc2);
78     __m512 vp4 = _mm512_fmadd_ps(vt4, vc3, vc2);
79 
80     vp0 = _mm512_mul_ps(vp0, vt0);
81     vp1 = _mm512_mul_ps(vp1, vt1);
82     vp2 = _mm512_mul_ps(vp2, vt2);
83     vp3 = _mm512_mul_ps(vp3, vt3);
84     vp4 = _mm512_mul_ps(vp4, vt4);
85 
86     vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
87     vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
88     vp2 = _mm512_fmadd_ps(vt2, vp2, vt2);
89     vp3 = _mm512_fmadd_ps(vt3, vp3, vt3);
90     vp4 = _mm512_fmadd_ps(vt4, vp4, vt4);
91 
92     vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
93     vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
94     vp2 = _mm512_fmadd_ps(vl2, vp2, vl2);
95     vp3 = _mm512_fmadd_ps(vl3, vp3, vl3);
96     vp4 = _mm512_fmadd_ps(vl4, vp4, vl4);
97 
98     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
99     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
100     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
101     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
102     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
103 
104     const __m512 vd0 = _mm512_add_ps(ve0, vone);
105     const __m512 vd1 = _mm512_add_ps(ve1, vone);
106     const __m512 vd2 = _mm512_add_ps(ve2, vone);
107     const __m512 vd3 = _mm512_add_ps(ve3, vone);
108     const __m512 vd4 = _mm512_add_ps(ve4, vone);
109 
110     __m512 vr0 = _mm512_rcp14_ps(vd0);
111     __m512 vr1 = _mm512_rcp14_ps(vd1);
112     __m512 vr2 = _mm512_rcp14_ps(vd2);
113     __m512 vr3 = _mm512_rcp14_ps(vd3);
114     __m512 vr4 = _mm512_rcp14_ps(vd4);
115 
116     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
117     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
118     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
119     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
120     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
121 
122     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
123     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
124     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
125     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
126     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
127 
128 
129     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
130     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
131     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
132     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
133     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
134 
135     _mm512_storeu_ps(y, vf0);
136     _mm512_storeu_ps(y + 16, vf1);
137     _mm512_storeu_ps(y + 32, vf2);
138     _mm512_storeu_ps(y + 48, vf3);
139     _mm512_storeu_ps(y + 64, vf4);
140     y += 80;
141   }
142   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
143     const __m512 vx = _mm512_loadu_ps(x);
144     x += 16;
145 
146     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
147 
148     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
149     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
150     vn = _mm512_sub_ps(vn, vmagic_bias);
151 
152     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
153 
154     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
155     vp = _mm512_mul_ps(vp, vt);
156     vp = _mm512_fmadd_ps(vt, vp, vt);
157     vp = _mm512_fmadd_ps(vl, vp, vl);
158 
159     const __m512 ve = _mm512_scalef_ps(vp, vn);
160     const __m512 vd = _mm512_add_ps(ve, vone);
161 
162     __m512 vr = _mm512_rcp14_ps(vd);
163     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
164 
165     __m512 vf = _mm512_mul_ps(ve, vr);
166 
167     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
168 
169     _mm512_storeu_ps(y, vf);
170     y += 16;
171   }
172   if XNN_UNLIKELY(n != 0) {
173     assert(n >= 1 * sizeof(float));
174     assert(n <= 15 * sizeof(float));
175 
176     // Prepare mask for valid 32-bit elements (depends on n).
177     n >>= 2 /* log2(sizeof(float)) */;
178     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
179 
180     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
181     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
182 
183     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
184     const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
185     vn = _mm512_sub_ps(vn, vmagic_bias);
186 
187     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
188 
189     __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
190     vp = _mm512_mul_ps(vp, vt);
191     vp = _mm512_fmadd_ps(vt, vp, vt);
192     vp = _mm512_fmadd_ps(vl, vp, vl);
193 
194     const __m512 ve = _mm512_scalef_ps(vp, vn);
195     const __m512 vd = _mm512_add_ps(ve, vone);
196 
197     __m512 vr = _mm512_rcp14_ps(vd);
198     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
199 
200     __m512 vf = _mm512_mul_ps(ve, vr);
201 
202     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
203 
204     _mm512_mask_storeu_ps(y, vmask, vf);
205   }
206 }
207