1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/sse-rr2-lut64-p2-div.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 <smmintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_64[64];
19
xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x20(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m128 vsign_mask = _mm_load_ps(params->sse2_rr2_lut64_p2.sign_mask);
29 const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut64_p2.magic_bias);
30 const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut64_p2.log2e);
31 const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut64_p2.index_mask);
32 const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_hi);
33 const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut64_p2.minus_ln2_lo);
34 const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut64_p2.c2);
35 const __m128 vone = _mm_load_ps(params->sse2_rr2_lut64_p2.one);
36 const __m128 vdenorm_cutoff = _mm_load_ps(params->sse2_rr2_lut64_p2.denorm_cutoff);
37
38 for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
39 const __m128 vx0123 = _mm_loadu_ps(x);
40 const __m128 vx4567 = _mm_loadu_ps(x + 4);
41 const __m128 vx89AB = _mm_loadu_ps(x + 8);
42 const __m128 vxCDEF = _mm_loadu_ps(x + 12);
43 const __m128 vxGHIJ = _mm_loadu_ps(x + 16);
44 x += 20;
45
46 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
47 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
48 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
49 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
50 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask);
51
52 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
53 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
54 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
55 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
56 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
57
58 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
59 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
60 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
61 const __m128i veCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 17);
62 const __m128i veGHIJ = _mm_slli_epi32(_mm_castps_si128(vnGHIJ), 17);
63
64 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
65 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
66 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
67 const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
68 const __m128i vidxGHIJ = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnGHIJ), vindex_mask), 2);
69
70 #if XNN_ARCH_X86_64
71 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
72 const uint64_t vidx23 = (uint64_t) _mm_extract_epi64(vidx0123, 1);
73 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
74 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
75 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))), 1);
76 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))), 1);
77 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
78 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
79 const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
80 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
81 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
82 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))), 1);
83 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))), 1);
84 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
85 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
86 const uint64_t vidxAB = (uint64_t) _mm_extract_epi64(vidx89AB, 1);
87 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
88 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
89 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))), 1);
90 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))), 1);
91 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
92 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
93 const uint64_t vidxEF = (uint64_t) _mm_extract_epi64(vidxCDEF, 1);
94 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
95 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
96 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))), 1);
97 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))), 1);
98 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
99 const uint64_t vidxGH = (uint64_t) _mm_cvtsi128_si64(vidxGHIJ);
100 const uint64_t vidxIJ = (uint64_t) _mm_extract_epi64(vidxGHIJ, 1);
101 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH)));
102 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ)));
103 const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32))), 1);
104 const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32))), 1);
105 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
106 #else // !XNN_ARCH_X86_64
107 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
108 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
109 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
110 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
111 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
112 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
113 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)), 1);
114 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)), 1);
115 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
116 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
117 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
118 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
119 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
120 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
121 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
122 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)), 1);
123 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)), 1);
124 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
125 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
126 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
127 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
128 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
129 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
130 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
131 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)), 1);
132 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1);
133 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
134 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
135 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
136 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
137 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
138 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
139 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
140 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)), 1);
141 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)), 1);
142 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
143 const uint32_t vidxG = (uint32_t) _mm_cvtsi128_si32(vidxGHIJ);
144 const uint32_t vidxH = (uint32_t) _mm_extract_epi16(vidxGHIJ, 2);
145 const uint32_t vidxI = (uint32_t) _mm_extract_epi16(vidxGHIJ, 4);
146 const uint32_t vidxJ = (uint32_t) _mm_extract_epi16(vidxGHIJ, 6);
147 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxG)));
148 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxI)));
149 const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxH)), 1);
150 const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxJ)), 1);
151 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
152 #endif // XNN_ARCH_X86_64
153
154 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
155 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
156 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
157 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
158 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_add_epi32(vlGHIJ, veGHIJ));
159
160 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
161 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
162 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
163 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
164 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
165
166 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
167 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
168 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
169 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
170 __m128 vtGHIJ = _mm_add_ps(vzGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_hi));
171
172 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
173 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
174 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
175 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
176 vtGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_lo));
177
178 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
179 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
180 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
181 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
182 __m128 vpGHIJ = _mm_mul_ps(vtGHIJ, vc2);
183
184 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
185 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
186 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
187 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
188 vpGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vpGHIJ, vtGHIJ));
189
190 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
191 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
192 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
193 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
194 const __m128 vyGHIJ = _mm_add_ps(vsGHIJ, _mm_mul_ps(vsGHIJ, vpGHIJ));
195
196 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
197 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
198 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
199 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
200 __m128 vfGHIJ = _mm_div_ps(vyGHIJ, _mm_add_ps(vyGHIJ, vone));
201
202 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
203 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
204 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
205 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
206 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
207
208 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
209 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
210 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
211 vfCDEF = _mm_blendv_ps(_mm_sub_ps(vone, vfCDEF), vfCDEF, vxCDEF);
212 vfGHIJ = _mm_blendv_ps(_mm_sub_ps(vone, vfGHIJ), vfGHIJ, vxGHIJ);
213
214 _mm_storeu_ps(y, vf0123);
215 _mm_storeu_ps(y + 4, vf4567);
216 _mm_storeu_ps(y + 8, vf89AB);
217 _mm_storeu_ps(y + 12, vfCDEF);
218 _mm_storeu_ps(y + 16, vfGHIJ);
219 y += 20;
220 }
221 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
222 const __m128 vx = _mm_loadu_ps(x);
223 x += 4;
224
225 const __m128 vz = _mm_or_ps(vx, vsign_mask);
226
227 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
228 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
229
230 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
231 #if XNN_ARCH_X86_64
232 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
233 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
234 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
235 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
236 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))), 1);
237 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))), 1);
238 #else // !XNN_ARCH_X86_64
239 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
240 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
241 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
242 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
243 #endif // XNN_ARCH_X86_64
244 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
245
246 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
247 vn = _mm_sub_ps(vn, vmagic_bias);
248
249 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
250 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
251
252 __m128 vp = _mm_mul_ps(vt, vc2);
253 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
254
255 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
256
257 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
258 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
259 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
260
261 _mm_storeu_ps(y, vf);
262 y += 4;
263 }
264 if XNN_UNLIKELY(n != 0) {
265 const __m128 vx = _mm_loadu_ps(x);
266
267 const __m128 vz = _mm_or_ps(vx, vsign_mask);
268
269 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
270 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
271
272 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
273 #if XNN_ARCH_X86_64
274 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
275 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
276 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
277 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
278 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))), 1);
279 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))), 1);
280 #else // !XNN_ARCH_X86_64
281 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
282 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
283 const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
284 const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
285 #endif // XNN_ARCH_X86_64
286 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
287
288 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
289 vn = _mm_sub_ps(vn, vmagic_bias);
290
291 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
292 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
293
294 __m128 vp = _mm_mul_ps(vt, vc2);
295 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
296
297 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
298
299 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
300 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
301 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
302
303 if (n & (2 * sizeof(float))) {
304 _mm_storel_pi((__m64*) y, vf);
305 vf = _mm_movehl_ps(vf, vf);
306 y += 2;
307 }
308 if (n & (1 * sizeof(float))) {
309 _mm_store_ss(y, vf);
310 }
311 }
312 }
313