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_nr2_x24(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx_rr2_p5_nr2_x24(
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 vtwo = _mm256_set1_ps(2.0f);
40 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
41
42 for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
43 const __m256 vx0 = _mm256_loadu_ps(x);
44 const __m256 vx1 = _mm256_loadu_ps(x + 8);
45 const __m256 vx2 = _mm256_loadu_ps(x + 16);
46 x += 24;
47
48 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
49 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
50 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
51
52 __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
53 __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
54 __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
55
56 const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
57 const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
58 const __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo0), vs_hi0, 1);
59 const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
60 const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
61 const __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo1), vs_hi1, 1);
62 const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
63 const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
64 const __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo2), vs_hi2, 1);
65
66 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
67 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
68 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
69
70 __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
71 __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
72 __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
73
74 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
75 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
76 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
77
78 __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc5, vt0), vc4);
79 __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc5, vt1), vc4);
80 __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc5, vt2), vc4);
81
82 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
83 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
84 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
85
86 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
87 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
88 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
89
90 vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc1);
91 vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc1);
92 vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc1);
93
94 vt0 = _mm256_mul_ps(vt0, vs0);
95 vt1 = _mm256_mul_ps(vt1, vs1);
96 vt2 = _mm256_mul_ps(vt2, vs2);
97
98 const __m256 ve0 = _mm256_add_ps(_mm256_mul_ps(vt0, vp0), vs0);
99 const __m256 ve1 = _mm256_add_ps(_mm256_mul_ps(vt1, vp1), vs1);
100 const __m256 ve2 = _mm256_add_ps(_mm256_mul_ps(vt2, vp2), vs2);
101
102 const __m256 vd0 = _mm256_add_ps(ve0, vone);
103 const __m256 vd1 = _mm256_add_ps(ve1, vone);
104 const __m256 vd2 = _mm256_add_ps(ve2, vone);
105
106 __m256 vr0 = _mm256_rcp_ps(vd0);
107 __m256 vr1 = _mm256_rcp_ps(vd1);
108 __m256 vr2 = _mm256_rcp_ps(vd2);
109
110 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
111 vr0 = _mm256_mul_ps(vr0, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr0, vd0)));
112 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
113 vr1 = _mm256_mul_ps(vr1, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr1, vd1)));
114 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
115 vr2 = _mm256_mul_ps(vr2, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr2, vd2)));
116
117 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
118 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
119 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
120
121 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
122 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
123 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
124
125 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
126 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
127 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
128
129 _mm256_storeu_ps(y, vf0);
130 _mm256_storeu_ps(y + 8, vf1);
131 _mm256_storeu_ps(y + 16, vf2);
132 y += 24;
133 }
134 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
135 const __m256 vx = _mm256_loadu_ps(x);
136 x += 8;
137
138 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
139
140 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
141
142 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
143 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
144 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
145
146 vn = _mm256_sub_ps(vn, vmagic_bias);
147
148 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
149 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
150
151 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
152 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
153 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
154 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
155
156 vt = _mm256_mul_ps(vt, vs);
157 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
158
159 const __m256 vd = _mm256_add_ps(ve, vone);
160 __m256 vr = _mm256_rcp_ps(vd);
161 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
162 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
163 __m256 vf = _mm256_mul_ps(ve, vr);
164
165 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
166 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
167
168 _mm256_storeu_ps(y, vf);
169 y += 8;
170 }
171 if XNN_UNLIKELY(n != 0) {
172 assert(n >= 1 * sizeof(float));
173 assert(n <= 7 * sizeof(float));
174 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
175
176 const __m256 vx = _mm256_maskload_ps(x, vmask);
177
178 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
179
180 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
181 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
182 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
183 const __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
184
185 vn = _mm256_sub_ps(vn, vmagic_bias);
186
187 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
188 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
189
190 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc5, vt), vc4);
191 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
192 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
193 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc1);
194
195 vt = _mm256_mul_ps(vt, vs);
196 const __m256 ve = _mm256_add_ps(_mm256_mul_ps(vt, vp), vs);
197
198 const __m256 vd = _mm256_add_ps(ve, vone);
199 __m256 vr = _mm256_rcp_ps(vd);
200 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
201 vr = _mm256_mul_ps(vr, _mm256_sub_ps(vtwo, _mm256_mul_ps(vr, vd)));
202 __m256 vf = _mm256_mul_ps(ve, vr);
203
204 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
205 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
206
207 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
208 __m128 vf_lo = _mm256_castps256_ps128(vf);
209 if (n & (4 * sizeof(float))) {
210 _mm_storeu_ps(y, vf_lo);
211 vf_lo = _mm256_extractf128_ps(vf, 1);
212 y += 4;
213 }
214 if (n & (2 * sizeof(float))) {
215 _mm_storel_pi((__m64*) y, vf_lo);
216 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
217 y += 2;
218 }
219 if (n & (1 * sizeof(float))) {
220 _mm_store_ss(y, vf_lo);
221 }
222 }
223 }
224