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_x32(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_x32(
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 >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
37 const __m512 vx0 = _mm512_loadu_ps(x);
38 const __m512 vx1 = _mm512_loadu_ps(x + 16);
39 x += 32;
40
41 const __m512 vz0 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx0), vsign_mask));
42 const __m512 vz1 = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx1), vsign_mask));
43
44 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
45 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
46
47 const __m512 vl0 = _mm512_permutexvar_ps(_mm512_castps_si512(vn0), vtable);
48 const __m512 vl1 = _mm512_permutexvar_ps(_mm512_castps_si512(vn1), vtable);
49
50 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
51 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
52
53 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
54 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
55
56 __m512 vp0 = _mm512_fmadd_ps(vt0, vc3, vc2);
57 __m512 vp1 = _mm512_fmadd_ps(vt1, vc3, vc2);
58
59 vp0 = _mm512_mul_ps(vp0, vt0);
60 vp1 = _mm512_mul_ps(vp1, vt1);
61
62 vp0 = _mm512_fmadd_ps(vt0, vp0, vt0);
63 vp1 = _mm512_fmadd_ps(vt1, vp1, vt1);
64
65 vp0 = _mm512_fmadd_ps(vl0, vp0, vl0);
66 vp1 = _mm512_fmadd_ps(vl1, vp1, vl1);
67
68 const __m512 ve0 = _mm512_scalef_ps(vp0, vn0);
69 const __m512 ve1 = _mm512_scalef_ps(vp1, vn1);
70
71 const __m512 vd0 = _mm512_add_ps(ve0, vone);
72 const __m512 vd1 = _mm512_add_ps(ve1, vone);
73
74 __m512 vr0 = _mm512_rcp14_ps(vd0);
75 __m512 vr1 = _mm512_rcp14_ps(vd1);
76
77 vr0 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
78 vr1 = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
79
80 __m512 vf0 = _mm512_mul_ps(ve0, vr0);
81 __m512 vf1 = _mm512_mul_ps(ve1, vr1);
82
83
84 vf0 = _mm512_mask_sub_ps(vf0, _mm512_testn_epi32_mask(_mm512_castps_si512(vx0), vsign_mask), vone, vf0);
85 vf1 = _mm512_mask_sub_ps(vf1, _mm512_testn_epi32_mask(_mm512_castps_si512(vx1), vsign_mask), vone, vf1);
86
87 _mm512_storeu_ps(y, vf0);
88 _mm512_storeu_ps(y + 16, vf1);
89 y += 32;
90 }
91 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
92 const __m512 vx = _mm512_loadu_ps(x);
93 x += 16;
94
95 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
96
97 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
98 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
99 vn = _mm512_sub_ps(vn, vmagic_bias);
100
101 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
102
103 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
104 vp = _mm512_mul_ps(vp, vt);
105 vp = _mm512_fmadd_ps(vt, vp, vt);
106 vp = _mm512_fmadd_ps(vl, vp, vl);
107
108 const __m512 ve = _mm512_scalef_ps(vp, vn);
109 const __m512 vd = _mm512_add_ps(ve, vone);
110
111 __m512 vr = _mm512_rcp14_ps(vd);
112 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
113
114 __m512 vf = _mm512_mul_ps(ve, vr);
115
116 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
117
118 _mm512_storeu_ps(y, vf);
119 y += 16;
120 }
121 if XNN_UNLIKELY(n != 0) {
122 assert(n >= 1 * sizeof(float));
123 assert(n <= 15 * sizeof(float));
124
125 // Prepare mask for valid 32-bit elements (depends on n).
126 n >>= 2 /* log2(sizeof(float)) */;
127 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
128
129 const __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
130 const __m512 vz = _mm512_castsi512_ps(_mm512_or_epi32(_mm512_castps_si512(vx), vsign_mask));
131
132 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
133 const __m512 vl = _mm512_permutexvar_ps(_mm512_castps_si512(vn), vtable);
134 vn = _mm512_sub_ps(vn, vmagic_bias);
135
136 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
137
138 __m512 vp = _mm512_fmadd_ps(vt, vc3, vc2);
139 vp = _mm512_mul_ps(vp, vt);
140 vp = _mm512_fmadd_ps(vt, vp, vt);
141 vp = _mm512_fmadd_ps(vl, vp, vl);
142
143 const __m512 ve = _mm512_scalef_ps(vp, vn);
144 const __m512 vd = _mm512_add_ps(ve, vone);
145
146 __m512 vr = _mm512_rcp14_ps(vd);
147 vr = _mm512_fmadd_ps(_mm512_fnmadd_ps(vr, vd, vone), vr, vr);
148
149 __m512 vf = _mm512_mul_ps(ve, vr);
150
151 vf = _mm512_mask_sub_ps(vf, _mm512_testn_epi32_mask(_mm512_castps_si512(vx), vsign_mask), vone, vf);
152
153 _mm512_mask_storeu_ps(y, vmask, vf);
154 }
155 }
156