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_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__sse2_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_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
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 vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))));
76 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
77 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))));
78 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
79 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
80 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
81 const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
82 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
83 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
84 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))));
85 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
86 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))));
87 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
88 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
89 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
90 const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
91 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
92 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
93 const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))));
94 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
95 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))));
96 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
97 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
98 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
99 const uint64_t vidxEF = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxCDEF, vidxCDEF));
100 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
101 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
102 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))));
103 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
104 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))));
105 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
106 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
107 const uint64_t vidxGH = (uint64_t) _mm_cvtsi128_si64(vidxGHIJ);
108 const uint64_t vidxIJ = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxGHIJ, vidxGHIJ));
109 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH)));
110 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ)));
111 const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32))));
112 const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
113 const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32))));
114 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
115 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
116 #else // !XNN_ARCH_X86_64
117 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
118 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
119 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
120 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
121 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
122 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
123 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)));
124 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
125 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)));
126 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
127 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
128 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
129 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
130 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
131 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
132 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
133 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
134 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)));
135 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
136 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)));
137 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
138 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
139 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
140 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
141 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
142 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
143 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
144 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
145 const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)));
146 const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
147 const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)));
148 const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
149 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
150 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
151 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
152 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
153 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
154 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
155 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
156 const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)));
157 const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
158 const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)));
159 const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
160 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
161 const uint32_t vidxG = (uint32_t) _mm_cvtsi128_si32(vidxGHIJ);
162 const uint32_t vidxH = (uint32_t) _mm_extract_epi16(vidxGHIJ, 2);
163 const uint32_t vidxI = (uint32_t) _mm_extract_epi16(vidxGHIJ, 4);
164 const uint32_t vidxJ = (uint32_t) _mm_extract_epi16(vidxGHIJ, 6);
165 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxG)));
166 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxI)));
167 const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxH)));
168 const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
169 const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxJ)));
170 const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
171 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
172 #endif // XNN_ARCH_X86_64
173
174 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
175 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
176 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
177 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
178 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_add_epi32(vlGHIJ, veGHIJ));
179
180 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
181 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
182 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
183 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
184 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
185
186 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
187 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
188 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
189 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
190 __m128 vtGHIJ = _mm_add_ps(vzGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_hi));
191
192 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
193 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
194 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
195 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
196 vtGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_lo));
197
198 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
199 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
200 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
201 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
202 __m128 vpGHIJ = _mm_mul_ps(vtGHIJ, vc2);
203
204 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
205 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
206 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
207 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
208 vpGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vpGHIJ, vtGHIJ));
209
210 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
211 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
212 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
213 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
214 const __m128 vyGHIJ = _mm_add_ps(vsGHIJ, _mm_mul_ps(vsGHIJ, vpGHIJ));
215
216 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
217 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
218 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
219 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
220 __m128 vfGHIJ = _mm_div_ps(vyGHIJ, _mm_add_ps(vyGHIJ, vone));
221
222 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
223 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
224 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
225 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
226 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
227
228 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
229 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
230 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
231 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
232 const __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ)));
233
234 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
235 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
236 vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
237 vfCDEF = _mm_or_ps(_mm_and_ps(vfCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, _mm_sub_ps(vone, vfCDEF)));
238 vfGHIJ = _mm_or_ps(_mm_and_ps(vfGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, _mm_sub_ps(vone, vfGHIJ)));
239
240 _mm_storeu_ps(y, vf0123);
241 _mm_storeu_ps(y + 4, vf4567);
242 _mm_storeu_ps(y + 8, vf89AB);
243 _mm_storeu_ps(y + 12, vfCDEF);
244 _mm_storeu_ps(y + 16, vfGHIJ);
245 y += 20;
246 }
247 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
248 const __m128 vx = _mm_loadu_ps(x);
249 x += 4;
250
251 const __m128 vz = _mm_or_ps(vx, vsign_mask);
252
253 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
254 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
255
256 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
257 #if XNN_ARCH_X86_64
258 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
259 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
260 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
261 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
262 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
263 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
264 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
265 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
266 #else // !XNN_ARCH_X86_64
267 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
268 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
269 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
270 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
271 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
272 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
273 #endif // XNN_ARCH_X86_64
274 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
275
276 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
277 vn = _mm_sub_ps(vn, vmagic_bias);
278
279 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
280 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
281
282 __m128 vp = _mm_mul_ps(vt, vc2);
283 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
284
285 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
286
287 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
288 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
289 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
290 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
291
292 _mm_storeu_ps(y, vf);
293 y += 4;
294 }
295 if XNN_UNLIKELY(n != 0) {
296 const __m128 vx = _mm_loadu_ps(x);
297
298 const __m128 vz = _mm_or_ps(vx, vsign_mask);
299
300 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
301 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
302
303 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
304 #if XNN_ARCH_X86_64
305 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
306 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
307 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
308 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
309 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
310 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
311 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
312 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
313 #else // !XNN_ARCH_X86_64
314 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
315 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
316 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
317 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
318 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
319 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
320 #endif // XNN_ARCH_X86_64
321 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
322
323 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
324 vn = _mm_sub_ps(vn, vmagic_bias);
325
326 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
327 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
328
329 __m128 vp = _mm_mul_ps(vt, vc2);
330 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
331
332 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
333
334 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
335 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
336 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
337 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
338
339 if (n & (2 * sizeof(float))) {
340 _mm_storel_pi((__m64*) y, vf);
341 vf = _mm_movehl_ps(vf, vf);
342 y += 2;
343 }
344 if (n & (1 * sizeof(float))) {
345 _mm_store_ss(y, vf);
346 }
347 }
348 }
349