1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-vsigmoid/avx512f-rr2-lut32-p2-perm2-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_rr2_lut32_p2_perm2_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_rr2_lut32_p2_perm2_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_rr2_lut32_p2.sign_mask);
28   const __m512 vmagic_bias = _mm512_set1_ps(params->avx512_rr2_lut32_p2.magic_bias);
29   const __m512 vlog2e = _mm512_set1_ps(params->avx512_rr2_lut32_p2.log2e);
30   const __m512 vtable_lo = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_lo);
31   const __m512 vtable_hi = _mm512_load_ps(params->avx512_rr2_lut32_p2.table_hi);
32   const __m512 vminus_ln2_hi = _mm512_set1_ps(params->avx512_rr2_lut32_p2.minus_ln2_hi);
33   const __m512 vminus_ln2_lo = _mm512_set1_ps(params->avx512_rr2_lut32_p2.minus_ln2_lo);
34   const __m512 vc2 = _mm512_set1_ps(params->avx512_rr2_lut32_p2.c2);
35   const __m512 vc1 = _mm512_set1_ps(params->avx512_rr2_lut32_p2.c1);
36   const __m512 vone = _mm512_set1_ps(params->avx512_rr2_lut32_p2.one);
37 
38   for (; n >= 80 * sizeof(float); n -= 80 * sizeof(float)) {
39     const __m512 vx0 = _mm512_loadu_ps(x);
40     const __m512 vx1 = _mm512_loadu_ps(x + 16);
41     const __m512 vx2 = _mm512_loadu_ps(x + 32);
42     const __m512 vx3 = _mm512_loadu_ps(x + 48);
43     const __m512 vx4 = _mm512_loadu_ps(x + 64);
44     x += 80;
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 
52     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
53     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
54     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
55     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
56     __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
57 
58     const __m512 vl0 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn0), vtable_hi);
59     const __m512 vl1 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn1), vtable_hi);
60     const __m512 vl2 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn2), vtable_hi);
61     const __m512 vl3 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn3), vtable_hi);
62     const __m512 vl4 = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn4), vtable_hi);
63 
64     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
65     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
66     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
67     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
68     vn4 = _mm512_sub_ps(vn4, vmagic_bias);
69 
70     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vz0);
71     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_hi, vz1);
72     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_hi, vz2);
73     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_hi, vz3);
74     __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_hi, vz4);
75 
76     vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0);
77     vt1 = _mm512_fmadd_ps(vn1, vminus_ln2_lo, vt1);
78     vt2 = _mm512_fmadd_ps(vn2, vminus_ln2_lo, vt2);
79     vt3 = _mm512_fmadd_ps(vn3, vminus_ln2_lo, vt3);
80     vt4 = _mm512_fmadd_ps(vn4, vminus_ln2_lo, vt4);
81 
82     __m512 vp0 = _mm512_fmadd_ps(vt0, vc2, vc1);
83     __m512 vp1 = _mm512_fmadd_ps(vt1, vc2, vc1);
84     __m512 vp2 = _mm512_fmadd_ps(vt2, vc2, vc1);
85     __m512 vp3 = _mm512_fmadd_ps(vt3, vc2, vc1);
86     __m512 vp4 = _mm512_fmadd_ps(vt4, vc2, vc1);
87 
88     vt0 = _mm512_mul_ps(vt0, vl0);
89     vt1 = _mm512_mul_ps(vt1, vl1);
90     vt2 = _mm512_mul_ps(vt2, vl2);
91     vt3 = _mm512_mul_ps(vt3, vl3);
92     vt4 = _mm512_mul_ps(vt4, vl4);
93 
94     vp0 = _mm512_fmadd_ps(vt0, vp0, vl0);
95     vp1 = _mm512_fmadd_ps(vt1, vp1, vl1);
96     vp2 = _mm512_fmadd_ps(vt2, vp2, vl2);
97     vp3 = _mm512_fmadd_ps(vt3, vp3, vl3);
98     vp4 = _mm512_fmadd_ps(vt4, vp4, vl4);
99 
100     const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
101     const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
102     const __m512 ve2 = _mm512_scalef_ps(vp2, vn2);
103     const __m512 ve3 = _mm512_scalef_ps(vp3, vn3);
104     const __m512 ve4 = _mm512_scalef_ps(vp4, vn4);
105 
106     const __m512 vd0 = _mm512_add_ps(ve0, vone);
107     const __m512 vd1 = _mm512_add_ps(ve1, vone);
108     const __m512 vd2 = _mm512_add_ps(ve2, vone);
109     const __m512 vd3 = _mm512_add_ps(ve3, vone);
110     const __m512 vd4 = _mm512_add_ps(ve4, vone);
111 
112     __m512 vr0 = _mm512_rcp14_ps(vd0);
113     __m512 vr1 = _mm512_rcp14_ps(vd1);
114     __m512 vr2 = _mm512_rcp14_ps(vd2);
115     __m512 vr3 = _mm512_rcp14_ps(vd3);
116     __m512 vr4 = _mm512_rcp14_ps(vd4);
117 
118     vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
119     vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
120     vr2 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
121     vr3 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
122     vr4 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
123 
124     __m512 vf0 = _mm512_mul_ps(ve0, vr0);
125     __m512 vf1 = _mm512_mul_ps(ve1, vr1);
126     __m512 vf2 = _mm512_mul_ps(ve2, vr2);
127     __m512 vf3 = _mm512_mul_ps(ve3, vr3);
128     __m512 vf4 = _mm512_mul_ps(ve4, vr4);
129 
130 
131     vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
132     vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
133     vf2 = _mm512_mask_sub_ps(vf2, _mm512_testn_epi32_mask(_mm512_castps_si512(vx2), vsign_mask), vone, vf2);
134     vf3 = _mm512_mask_sub_ps(vf3, _mm512_testn_epi32_mask(_mm512_castps_si512(vx3), vsign_mask), vone, vf3);
135     vf4 = _mm512_mask_sub_ps(vf4, _mm512_testn_epi32_mask(_mm512_castps_si512(vx4), vsign_mask), vone, vf4);
136 
137     _mm512_storeu_ps(y, vf0);
138     _mm512_storeu_ps(y + 16, vf1);
139     _mm512_storeu_ps(y + 32, vf2);
140     _mm512_storeu_ps(y + 48, vf3);
141     _mm512_storeu_ps(y + 64, vf4);
142     y += 80;
143   }
144   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
145     const __m512 vx = _mm512_loadu_ps(x);
146     x += 16;
147 
148     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
149 
150     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
151     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
152     vn = _mm512_sub_ps(vn, vmagic_bias);
153 
154     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
155     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
156 
157     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
158     vt = _mm512_mul_ps(vt, vl);
159     vp = _mm512_fmadd_ps(vt, vp, vl);
160 
161     const __m512 ve = _mm512_scalef_ps(vp, vn);
162     const __m512 vd = _mm512_add_ps(ve, vone);
163 
164     __m512 vr = _mm512_rcp14_ps(vd);
165     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
166 
167     __m512 vf = _mm512_mul_ps(ve, vr);
168 
169     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
170 
171     _mm512_storeu_ps(y, vf);
172     y += 16;
173   }
174   if XNN_UNLIKELY(n != 0) {
175     assert(n >= 1 * sizeof(float));
176     assert(n <= 15 * sizeof(float));
177 
178     // Prepare mask for valid 32-bit elements (depends on n).
179     n >>= 2 /* log2(sizeof(float)) */;
180     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
181 
182     const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
183     const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
184 
185     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
186     const __m512 vl = _mm512_permutex2var_ps(vtable_lo, _mm512_castps_si512(vn), vtable_hi);
187     vn = _mm512_sub_ps(vn, vmagic_bias);
188 
189     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vz);
190     vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt);
191 
192     __m512 vp = _mm512_fmadd_ps(vt, vc2, vc1);
193     vt = _mm512_mul_ps(vt, vl);
194     vp = _mm512_fmadd_ps(vt, vp, vl);
195 
196     const __m512 ve = _mm512_scalef_ps(vp, vn);
197     const __m512 vd = _mm512_add_ps(ve, vone);
198 
199     __m512 vr = _mm512_rcp14_ps(vd);
200     vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
201 
202     __m512 vf = _mm512_mul_ps(ve, vr);
203 
204     vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
205 
206     _mm512_mask_storeu_ps(y, vmask, vf);
207   }
208 }
209