1 // Auto-generated file. Do not edit!
2 // Template: src/f32-sigmoid/sse-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_sigmoid_ukernel__sse41_lut64_p2_div_x12(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__sse41_lut64_p2_div_x12(
21 size_t n,
22 const float* x,
23 float* y,
24 const void* params) XNN_DISABLE_TSAN
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m128 vsign_mask = _mm_set1_ps(-0.0f);
29 const __m128 vmagic_bias = _mm_set1_ps(0x1.800000p17f);
30 const __m128 vlog2e = _mm_set1_ps(0x1.715476p0f);
31 const __m128i vindex_mask = _mm_set1_epi32(INT32_C(0x3F));
32 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.630000p-1f);
33 const __m128 vminus_ln2_lo = _mm_set1_ps(0x1.BD0106p-13f);
34 const __m128 vc2 = _mm_set1_ps(0x1.FFFF0Ap-2f);
35 const __m128 vone = _mm_set1_ps(1.0f);
36 const __m128 vdenorm_cutoff = _mm_set1_ps(-0x1.5D589Ep+6f);
37
38 for (; n >= 12 * sizeof(float); n -= 12 * 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 x += 12;
43
44 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
45 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
46 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
47
48 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
49 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
50 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
51
52 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
53 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
54 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
55
56 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
57 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
58 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
59
60 #if XNN_ARCH_X86_64
61 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
62 const uint64_t vidx23 = (uint64_t) _mm_extract_epi64(vidx0123, 1);
63 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
64 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
65 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))), 1);
66 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))), 1);
67 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
68 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
69 const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
70 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
71 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
72 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))), 1);
73 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))), 1);
74 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
75 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
76 const uint64_t vidxAB = (uint64_t) _mm_extract_epi64(vidx89AB, 1);
77 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
78 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
79 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))), 1);
80 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))), 1);
81 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
82 #else // !XNN_ARCH_X86_64
83 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
84 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
85 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
86 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
87 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
88 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
89 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)), 1);
90 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)), 1);
91 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
92 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
93 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
94 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
95 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
96 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
97 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
98 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)), 1);
99 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)), 1);
100 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
101 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
102 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
103 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
104 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
105 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
106 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
107 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)), 1);
108 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1);
109 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
110 #endif // XNN_ARCH_X86_64
111
112 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
113 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
114 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
115
116 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
117 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
118 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
119
120 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
121 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
122 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
123
124 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
125 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
126 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
127
128 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
129 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
130 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
131
132 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
133 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
134 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
135
136 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
137 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
138 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
139
140 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
141 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
142 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
143
144 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
145 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
146 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
147
148 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
149 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
150 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
151
152 _mm_storeu_ps(y, vf0123);
153 _mm_storeu_ps(y + 4, vf4567);
154 _mm_storeu_ps(y + 8, vf89AB);
155 y += 12;
156 }
157 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
158 const __m128 vx = _mm_loadu_ps(x);
159 x += 4;
160
161 const __m128 vz = _mm_or_ps(vx, vsign_mask);
162
163 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
164 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
165
166 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
167 #if XNN_ARCH_X86_64
168 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
169 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
170 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
171 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
172 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);
173 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);
174 #else // !XNN_ARCH_X86_64
175 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
176 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
177 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);
178 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);
179 #endif // XNN_ARCH_X86_64
180 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
181
182 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
183 vn = _mm_sub_ps(vn, vmagic_bias);
184
185 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
186 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
187
188 __m128 vp = _mm_mul_ps(vt, vc2);
189 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
190
191 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
192
193 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
194 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
195 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
196
197 _mm_storeu_ps(y, vf);
198 y += 4;
199 }
200 if XNN_UNLIKELY(n != 0) {
201 const __m128 vx = _mm_loadu_ps(x);
202
203 const __m128 vz = _mm_or_ps(vx, vsign_mask);
204
205 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
206 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
207
208 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
209 #if XNN_ARCH_X86_64
210 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
211 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
212 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
213 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
214 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);
215 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);
216 #else // !XNN_ARCH_X86_64
217 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
218 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
219 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);
220 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);
221 #endif // XNN_ARCH_X86_64
222 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
223
224 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
225 vn = _mm_sub_ps(vn, vmagic_bias);
226
227 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
228 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
229
230 __m128 vp = _mm_mul_ps(vt, vc2);
231 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
232
233 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
234
235 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
236 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
237 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
238
239 if (n & (2 * sizeof(float))) {
240 _mm_storel_pi((__m64*) y, vf);
241 vf = _mm_movehl_ps(vf, vf);
242 y += 2;
243 }
244 if (n & (1 * sizeof(float))) {
245 _mm_store_ss(y, vf);
246 }
247 }
248 }
249