1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/avx2-p5.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2019 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/vunary.h>
16
17
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x16(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx2_rr1_p5_div_x16(
21 size_t n,
22 const float* x,
23 float* y,
24 const void* params)
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
29 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
31 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
33 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
34 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
35 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
36 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37 const __m256 vone = _mm256_set1_ps(1.0f);
38 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
39
40 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
41 const __m256 vx0 = _mm256_loadu_ps(x);
42 const __m256 vx1 = _mm256_loadu_ps(x + 8);
43 x += 16;
44
45 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
46 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
47
48 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
49 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
50
51 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
52 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
53
54 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
55 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
56
57 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
58 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
59
60 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
61 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
62
63 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
64 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
65
66 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
67 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
68
69 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
70 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
71
72 vt0 = _mm256_mul_ps(vt0, vs0);
73 vt1 = _mm256_mul_ps(vt1, vs1);
74
75 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
76 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
77
78 const __m256 vd0 = _mm256_add_ps(ve0, vone);
79 const __m256 vd1 = _mm256_add_ps(ve1, vone);
80
81 __m256 vf0 = _mm256_div_ps(ve0, vd0);
82 __m256 vf1 = _mm256_div_ps(ve1, vd1);
83
84 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
85 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
86
87 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
88 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
89
90 _mm256_storeu_ps(y, vf0);
91 _mm256_storeu_ps(y + 8, vf1);
92 y += 16;
93 }
94 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
95 const __m256 vx = _mm256_loadu_ps(x);
96 x += 8;
97
98 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
99
100 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
101 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
102 vn = _mm256_sub_ps(vn, vmagic_bias);
103
104 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
105
106 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
107 vp = _mm256_fmadd_ps(vp, vt, vc3);
108 vp = _mm256_fmadd_ps(vp, vt, vc2);
109 vp = _mm256_fmadd_ps(vp, vt, vc1);
110
111 vt = _mm256_mul_ps(vt, vs);
112 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
113
114 const __m256 vd = _mm256_add_ps(ve, vone);
115 __m256 vf = _mm256_div_ps(ve, vd);
116
117 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
118 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
119
120 _mm256_storeu_ps(y, vf);
121 y += 8;
122 }
123 if XNN_UNLIKELY(n != 0) {
124 assert(n >= 1 * sizeof(float));
125 assert(n <= 7 * sizeof(float));
126 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
127
128 const __m256 vx = _mm256_maskload_ps(x, vmask);
129
130 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
131
132 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
133 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
134 vn = _mm256_sub_ps(vn, vmagic_bias);
135
136 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
137
138 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
139 vp = _mm256_fmadd_ps(vp, vt, vc3);
140 vp = _mm256_fmadd_ps(vp, vt, vc2);
141 vp = _mm256_fmadd_ps(vp, vt, vc1);
142
143 vt = _mm256_mul_ps(vt, vs);
144 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
145
146 const __m256 vd = _mm256_add_ps(ve, vone);
147 __m256 vf = _mm256_div_ps(ve, vd);
148
149 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
150 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
151
152 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
153 __m128 vf_lo = _mm256_castps256_ps128(vf);
154 if (n & (4 * sizeof(float))) {
155 _mm_storeu_ps(y, vf_lo);
156 vf_lo = _mm256_extractf128_ps(vf, 1);
157 y += 4;
158 }
159 if (n & (2 * sizeof(float))) {
160 _mm_storel_pi((__m64*) y, vf_lo);
161 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
162 y += 2;
163 }
164 if (n & (1 * sizeof(float))) {
165 _mm_store_ss(y, vf_lo);
166 }
167 }
168 }
169