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 <emmintrin.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__sse2_rr2_lut64_p2_div_x16(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__sse2_rr2_lut64_p2_div_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 x += 16;
44
45 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
46 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
47 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
48 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
49
50 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
51 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
52 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
53 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
54
55 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
56 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
57 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
58 const __m128i veCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 17);
59
60 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
61 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
62 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
63 const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
64
65 #if XNN_ARCH_X86_64
66 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
67 const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
68 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
69 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
70 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))));
71 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
72 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))));
73 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
74 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
75 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
76 const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
77 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
78 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
79 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))));
80 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
81 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))));
82 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
83 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
84 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
85 const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
86 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
87 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
88 const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))));
89 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
90 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))));
91 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
92 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
93 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
94 const uint64_t vidxEF = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxCDEF, vidxCDEF));
95 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
96 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
97 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))));
98 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
99 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))));
100 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
101 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
102 #else // !XNN_ARCH_X86_64
103 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
104 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
105 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
106 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
107 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
108 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
109 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)));
110 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
111 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)));
112 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
113 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
114 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
115 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
116 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
117 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
118 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
119 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
120 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)));
121 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
122 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)));
123 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
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 vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)));
132 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
133 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)));
134 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
135 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
136 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
137 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
138 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
139 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
140 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
141 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
142 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)));
143 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
144 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)));
145 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
146 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
147 #endif // XNN_ARCH_X86_64
148
149 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
150 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
151 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
152 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
153
154 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
155 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
156 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
157 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
158
159 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
160 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
161 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
162 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
163
164 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
165 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
166 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
167 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
168
169 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
170 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
171 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
172 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
173
174 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
175 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
176 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
177 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
178
179 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
180 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
181 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
182 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
183
184 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
185 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
186 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
187 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
188
189 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
190 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
191 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
192 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
193
194 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
195 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
196 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
197 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
198
199 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
200 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
201 vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
202 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF)));
203
204 _mm_storeu_ps(y, vf0123);
205 _mm_storeu_ps(y + 4, vf4567);
206 _mm_storeu_ps(y + 8, vf89AB);
207 _mm_storeu_ps(y + 12, vfCDEF);
208 y += 16;
209 }
210 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
211 const __m128 vx = _mm_loadu_ps(x);
212 x += 4;
213
214 const __m128 vz = _mm_or_ps(vx, vsign_mask);
215
216 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
217 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
218
219 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
220 #if XNN_ARCH_X86_64
221 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
222 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
223 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
224 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
225 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
226 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
227 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
228 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
229 #else // !XNN_ARCH_X86_64
230 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
231 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
232 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
233 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
234 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
235 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
236 #endif // XNN_ARCH_X86_64
237 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
238
239 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
240 vn = _mm_sub_ps(vn, vmagic_bias);
241
242 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
243 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
244
245 __m128 vp = _mm_mul_ps(vt, vc2);
246 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
247
248 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
249
250 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
251 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
252 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
253 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
254
255 _mm_storeu_ps(y, vf);
256 y += 4;
257 }
258 if XNN_UNLIKELY(n != 0) {
259 const __m128 vx = _mm_loadu_ps(x);
260
261 const __m128 vz = _mm_or_ps(vx, vsign_mask);
262
263 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
264 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
265
266 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
267 #if XNN_ARCH_X86_64
268 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
269 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
270 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
271 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
272 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
273 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
274 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
275 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
276 #else // !XNN_ARCH_X86_64
277 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
278 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
279 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
280 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
281 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
282 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
283 #endif // XNN_ARCH_X86_64
284 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
285
286 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
287 vn = _mm_sub_ps(vn, vmagic_bias);
288
289 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
290 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
291
292 __m128 vp = _mm_mul_ps(vt, vc2);
293 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
294
295 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
296
297 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
298 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
299 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
300 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
301
302 if (n & (2 * sizeof(float))) {
303 _mm_storel_pi((__m64*) y, vf);
304 vf = _mm_movehl_ps(vf, vf);
305 y += 2;
306 }
307 if (n & (1 * sizeof(float))) {
308 _mm_store_ss(y, vf);
309 }
310 }
311 }
312