1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/avx-p5.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/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__avx_rr2_p5_div_x16(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx_rr2_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_hi = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
33 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
34 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
35 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
36 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
37 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
38 const __m256 vone = _mm256_set1_ps(1.0f);
39 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
40
41 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
42 const __m256 vx0 = _mm256_loadu_ps(x);
43 const __m256 vx1 = _mm256_loadu_ps(x + 8);
44 x += 16;
45
46 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
47 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
48
49 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
50 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
51
52 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
53 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
54 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
55 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
56 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
57 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
58
59 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
60 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
61
62 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
63 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
64
65 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
66 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
67
68 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
69 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
70
71 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
72 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
73
74 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
75 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
76
77 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
78 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
79
80 vt0 = _mm256_mul_ps(vt0, vs0);
81 vt1 = _mm256_mul_ps(vt1, vs1);
82
83 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
84 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
85
86 const __m256 vd0 = _mm256_add_ps(ve0, vone);
87 const __m256 vd1 = _mm256_add_ps(ve1, vone);
88
89 __m256 vf0 = _mm256_div_ps(ve0, vd0);
90 __m256 vf1 = _mm256_div_ps(ve1, vd1);
91
92 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
93 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
94
95 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
96 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
97
98 _mm256_storeu_ps(y, vf0);
99 _mm256_storeu_ps(y + 8, vf1);
100 y += 16;
101 }
102 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
103 const __m256 vx = _mm256_loadu_ps(x);
104 x += 8;
105
106 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
107
108 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
109
110 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
111 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
112 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
113
114 vn = _mm256_sub_ps(vn, vmagic_bias);
115
116 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
117 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
118
119 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
120 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
121 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
122 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
123
124 vt = _mm256_mul_ps(vt, vs);
125 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
126
127 const __m256 vd = _mm256_add_ps(ve, vone);
128 __m256 vf = _mm256_div_ps(ve, vd);
129
130 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
131 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
132
133 _mm256_storeu_ps(y, vf);
134 y += 8;
135 }
136 if XNN_UNLIKELY(n != 0) {
137 assert(n >= 1 * sizeof(float));
138 assert(n <= 7 * sizeof(float));
139 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
140
141 const __m256 vx = _mm256_maskload_ps(x, vmask);
142
143 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
144
145 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
146 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
147 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
148 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
149
150 vn = _mm256_sub_ps(vn, vmagic_bias);
151
152 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
153 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
154
155 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
156 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
157 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
158 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
159
160 vt = _mm256_mul_ps(vt, vs);
161 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
162
163 const __m256 vd = _mm256_add_ps(ve, vone);
164 __m256 vf = _mm256_div_ps(ve, vd);
165
166 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
167 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
168
169 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
170 __m128 vf_lo = _mm256_castps256_ps128(vf);
171 if (n & (4 * sizeof(float))) {
172 _mm_storeu_ps(y, vf_lo);
173 vf_lo = _mm256_extractf128_ps(vf, 1);
174 y += 4;
175 }
176 if (n & (2 * sizeof(float))) {
177 _mm_storel_pi((__m64*) y, vf_lo);
178 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
179 y += 2;
180 }
181 if (n & (1 * sizeof(float))) {
182 _mm_store_ss(y, vf_lo);
183 }
184 }
185 }
186