1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx-rr2-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
xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x16(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx_rr2_p5_div_x16(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx_rr2_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p5.log2e);
29 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_hi);
30 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p5.minus_ln2_lo);
31 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p5.c5);
32 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p5.c4);
33 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p5.c3);
34 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p5.c2);
35 const __m256 vc1 = _mm256_load_ps(params->avx_rr2_p5.c1);
36 const __m256 vone = _mm256_load_ps(params->avx_rr2_p5.one);
37 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx_rr2_p5.denorm_cutoff);
38
39 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
40 const __m256 vx0 = _mm256_loadu_ps(x);
41 const __m256 vx1 = _mm256_loadu_ps(x + 8);
42 x += 16;
43
44 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
45 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
46
47 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
48 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
49
50 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
51 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
52 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
53 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
54 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
55 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
56
57 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
58 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
59
60 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
61 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
62
63 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
64 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
65
66 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
67 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
68
69 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
70 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
71
72 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
73 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
74
75 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
76 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
77
78 vt0 = _mm256_mul_ps(vt0, vs0);
79 vt1 = _mm256_mul_ps(vt1, vs1);
80
81 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
82 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
83
84 const __m256 vd0 = _mm256_add_ps(ve0, vone);
85 const __m256 vd1 = _mm256_add_ps(ve1, vone);
86
87 __m256 vf0 = _mm256_div_ps(ve0, vd0);
88 __m256 vf1 = _mm256_div_ps(ve1, vd1);
89
90 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
91 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
92
93 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
94 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
95
96 _mm256_storeu_ps(y, vf0);
97 _mm256_storeu_ps(y + 8, vf1);
98 y += 16;
99 }
100 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
101 const __m256 vx = _mm256_loadu_ps(x);
102 x += 8;
103
104 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
105
106 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
107
108 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
109 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
110 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
111
112 vn = _mm256_sub_ps(vn, vmagic_bias);
113
114 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
115 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
116
117 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
118 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
119 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
120 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
121
122 vt = _mm256_mul_ps(vt, vs);
123 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
124
125 const __m256 vd = _mm256_add_ps(ve, vone);
126 __m256 vf = _mm256_div_ps(ve, vd);
127
128 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
129 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
130
131 _mm256_storeu_ps(y, vf);
132 y += 8;
133 }
134 if XNN_UNLIKELY(n != 0) {
135 assert(n >= 1 * sizeof(float));
136 assert(n <= 7 * sizeof(float));
137 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p5.mask_table[7] - n));
138
139 const __m256 vx = _mm256_maskload_ps(x, vmask);
140
141 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
142
143 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
144 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
145 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
146 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
147
148 vn = _mm256_sub_ps(vn, vmagic_bias);
149
150 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
151 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
152
153 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
154 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
155 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
156 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
157
158 vt = _mm256_mul_ps(vt, vs);
159 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
160
161 const __m256 vd = _mm256_add_ps(ve, vone);
162 __m256 vf = _mm256_div_ps(ve, vd);
163
164 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
165 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
166
167 __m128 vf_lo = _mm256_castps256_ps128(vf);
168 if (n & (4 * sizeof(float))) {
169 _mm_storeu_ps(y, vf_lo);
170 vf_lo = _mm256_extractf128_ps(vf, 1);
171 y += 4;
172 }
173 if (n & (2 * sizeof(float))) {
174 _mm_storel_pi((__m64*) y, vf_lo);
175 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
176 y += 2;
177 }
178 if (n & (1 * sizeof(float))) {
179 _mm_store_ss(y, vf_lo);
180 }
181 }
182 }
183