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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 const __m128 vxKLMN = _mm_loadu_ps(x + 20);
45 x += 24;
46
47 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
48 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
49 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
50 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
51 const __m128 vzGHIJ = _mm_or_ps(vxGHIJ, vsign_mask);
52 const __m128 vzKLMN = _mm_or_ps(vxKLMN, vsign_mask);
53
54 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
55 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
56 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
57 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
58 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
59 __m128 vnKLMN = _mm_add_ps(_mm_mul_ps(vzKLMN, vlog2e), vmagic_bias);
60
61 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
62 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
63 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
64 const __m128i veCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 17);
65 const __m128i veGHIJ = _mm_slli_epi32(_mm_castps_si128(vnGHIJ), 17);
66 const __m128i veKLMN = _mm_slli_epi32(_mm_castps_si128(vnKLMN), 17);
67
68 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
69 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
70 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
71 const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
72 const __m128i vidxGHIJ = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnGHIJ), vindex_mask), 2);
73 const __m128i vidxKLMN = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnKLMN), vindex_mask), 2);
74
75 #if XNN_ARCH_X86_64
76 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
77 const uint64_t vidx23 = (uint64_t) _mm_extract_epi64(vidx0123, 1);
78 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
79 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
80 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))), 1);
81 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))), 1);
82 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
83 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
84 const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
85 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
86 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
87 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))), 1);
88 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))), 1);
89 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
90 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
91 const uint64_t vidxAB = (uint64_t) _mm_extract_epi64(vidx89AB, 1);
92 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
93 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
94 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))), 1);
95 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))), 1);
96 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
97 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
98 const uint64_t vidxEF = (uint64_t) _mm_extract_epi64(vidxCDEF, 1);
99 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
100 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
101 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))), 1);
102 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))), 1);
103 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
104 const uint64_t vidxGH = (uint64_t) _mm_cvtsi128_si64(vidxGHIJ);
105 const uint64_t vidxIJ = (uint64_t) _mm_extract_epi64(vidxGHIJ, 1);
106 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxGH)));
107 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxIJ)));
108 const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxGH >> 32))), 1);
109 const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxIJ >> 32))), 1);
110 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
111 const uint64_t vidxKL = (uint64_t) _mm_cvtsi128_si64(vidxKLMN);
112 const uint64_t vidxMN = (uint64_t) _mm_extract_epi64(vidxKLMN, 1);
113 const __m128i vlK = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxKL)));
114 const __m128i vlM = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxMN)));
115 const __m128i vlKL = _mm_insert_epi32(vlK, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxKL >> 32))), 1);
116 const __m128i vlMN = _mm_insert_epi32(vlM, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxMN >> 32))), 1);
117 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN);
118 #else // !XNN_ARCH_X86_64
119 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
120 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
121 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
122 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
123 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
124 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
125 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)), 1);
126 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)), 1);
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 vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)), 1);
135 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)), 1);
136 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
137 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
138 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
139 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
140 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
141 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
142 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
143 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)), 1);
144 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1);
145 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
146 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
147 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
148 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
149 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
150 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
151 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
152 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)), 1);
153 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)), 1);
154 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
155 const uint32_t vidxG = (uint32_t) _mm_cvtsi128_si32(vidxGHIJ);
156 const uint32_t vidxH = (uint32_t) _mm_extract_epi16(vidxGHIJ, 2);
157 const uint32_t vidxI = (uint32_t) _mm_extract_epi16(vidxGHIJ, 4);
158 const uint32_t vidxJ = (uint32_t) _mm_extract_epi16(vidxGHIJ, 6);
159 const __m128i vlG = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxG)));
160 const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxI)));
161 const __m128i vlGH = _mm_insert_epi32(vlG, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxH)), 1);
162 const __m128i vlIJ = _mm_insert_epi32(vlI, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxJ)), 1);
163 const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
164 const uint32_t vidxK = (uint32_t) _mm_cvtsi128_si32(vidxKLMN);
165 const uint32_t vidxL = (uint32_t) _mm_extract_epi16(vidxKLMN, 2);
166 const uint32_t vidxM = (uint32_t) _mm_extract_epi16(vidxKLMN, 4);
167 const uint32_t vidxN = (uint32_t) _mm_extract_epi16(vidxKLMN, 6);
168 const __m128i vlK = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxK)));
169 const __m128i vlM = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxM)));
170 const __m128i vlKL = _mm_insert_epi32(vlK, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxL)), 1);
171 const __m128i vlMN = _mm_insert_epi32(vlM, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxN)), 1);
172 const __m128i vlKLMN = _mm_unpacklo_epi64(vlKL, vlMN);
173 #endif // XNN_ARCH_X86_64
174
175 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
176 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
177 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
178 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
179 const __m128 vsGHIJ = _mm_castsi128_ps(_mm_add_epi32(vlGHIJ, veGHIJ));
180 const __m128 vsKLMN = _mm_castsi128_ps(_mm_add_epi32(vlKLMN, veKLMN));
181
182 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
183 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
184 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
185 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
186 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
187 vnKLMN = _mm_sub_ps(vnKLMN, vmagic_bias);
188
189 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
190 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
191 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
192 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
193 __m128 vtGHIJ = _mm_add_ps(vzGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_hi));
194 __m128 vtKLMN = _mm_add_ps(vzKLMN, _mm_mul_ps(vnKLMN, vminus_ln2_hi));
195
196 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
197 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
198 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
199 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
200 vtGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vnGHIJ, vminus_ln2_lo));
201 vtKLMN = _mm_add_ps(vtKLMN, _mm_mul_ps(vnKLMN, vminus_ln2_lo));
202
203 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
204 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
205 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
206 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
207 __m128 vpGHIJ = _mm_mul_ps(vtGHIJ, vc2);
208 __m128 vpKLMN = _mm_mul_ps(vtKLMN, vc2);
209
210 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
211 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
212 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
213 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
214 vpGHIJ = _mm_add_ps(vtGHIJ, _mm_mul_ps(vpGHIJ, vtGHIJ));
215 vpKLMN = _mm_add_ps(vtKLMN, _mm_mul_ps(vpKLMN, vtKLMN));
216
217 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
218 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
219 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
220 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
221 const __m128 vyGHIJ = _mm_add_ps(vsGHIJ, _mm_mul_ps(vsGHIJ, vpGHIJ));
222 const __m128 vyKLMN = _mm_add_ps(vsKLMN, _mm_mul_ps(vsKLMN, vpKLMN));
223
224 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
225 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
226 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
227 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
228 __m128 vfGHIJ = _mm_div_ps(vyGHIJ, _mm_add_ps(vyGHIJ, vone));
229 __m128 vfKLMN = _mm_div_ps(vyKLMN, _mm_add_ps(vyKLMN, vone));
230
231 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
232 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
233 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
234 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
235 vfGHIJ = _mm_andnot_ps(_mm_cmplt_ps(vzGHIJ, vdenorm_cutoff), vfGHIJ);
236 vfKLMN = _mm_andnot_ps(_mm_cmplt_ps(vzKLMN, vdenorm_cutoff), vfKLMN);
237
238 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
239 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
240 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
241 vfCDEF = _mm_blendv_ps(_mm_sub_ps(vone, vfCDEF), vfCDEF, vxCDEF);
242 vfGHIJ = _mm_blendv_ps(_mm_sub_ps(vone, vfGHIJ), vfGHIJ, vxGHIJ);
243 vfKLMN = _mm_blendv_ps(_mm_sub_ps(vone, vfKLMN), vfKLMN, vxKLMN);
244
245 _mm_storeu_ps(y, vf0123);
246 _mm_storeu_ps(y + 4, vf4567);
247 _mm_storeu_ps(y + 8, vf89AB);
248 _mm_storeu_ps(y + 12, vfCDEF);
249 _mm_storeu_ps(y + 16, vfGHIJ);
250 _mm_storeu_ps(y + 20, vfKLMN);
251 y += 24;
252 }
253 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
254 const __m128 vx = _mm_loadu_ps(x);
255 x += 4;
256
257 const __m128 vz = _mm_or_ps(vx, vsign_mask);
258
259 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
260 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
261
262 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
263 #if XNN_ARCH_X86_64
264 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
265 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
266 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
267 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
268 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);
269 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);
270 #else // !XNN_ARCH_X86_64
271 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
272 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
273 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);
274 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);
275 #endif // XNN_ARCH_X86_64
276 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
277
278 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
279 vn = _mm_sub_ps(vn, vmagic_bias);
280
281 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
282 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
283
284 __m128 vp = _mm_mul_ps(vt, vc2);
285 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
286
287 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
288
289 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
290 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
291 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
292
293 _mm_storeu_ps(y, vf);
294 y += 4;
295 }
296 if XNN_UNLIKELY(n != 0) {
297 const __m128 vx = _mm_loadu_ps(x);
298
299 const __m128 vz = _mm_or_ps(vx, vsign_mask);
300
301 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
302 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
303
304 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
305 #if XNN_ARCH_X86_64
306 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
307 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
308 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
309 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
310 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);
311 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);
312 #else // !XNN_ARCH_X86_64
313 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
314 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
315 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);
316 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);
317 #endif // XNN_ARCH_X86_64
318 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
319
320 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
321 vn = _mm_sub_ps(vn, vmagic_bias);
322
323 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
324 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
325
326 __m128 vp = _mm_mul_ps(vt, vc2);
327 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
328
329 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
330
331 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
332 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
333 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
334
335 if (n & (2 * sizeof(float))) {
336 _mm_storel_pi((__m64*) y, vf);
337 vf = _mm_movehl_ps(vf, vf);
338 y += 2;
339 }
340 if (n & (1 * sizeof(float))) {
341 _mm_store_ss(y, vf);
342 }
343 }
344 }
345