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_x8(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_x8(
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 >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
39 const __m128 vx0123 = _mm_loadu_ps(x);
40 const __m128 vx4567 = _mm_loadu_ps(x + 4);
41 x += 8;
42
43 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
44 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
45
46 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
47 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
48
49 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
50 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
51
52 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
53 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
54
55 #if XNN_ARCH_X86_64
56 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
57 const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
58 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
59 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
60 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))));
61 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
62 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))));
63 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
64 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
65 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
66 const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
67 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
68 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
69 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))));
70 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
71 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))));
72 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
73 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
74 #else // !XNN_ARCH_X86_64
75 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
76 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
77 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
78 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
79 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
80 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
81 const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)));
82 const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
83 const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)));
84 const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
85 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
86 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
87 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
88 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
89 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
90 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
91 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
92 const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)));
93 const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
94 const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)));
95 const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
96 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
97 #endif // XNN_ARCH_X86_64
98
99 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
100 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
101
102 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
103 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
104
105 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
106 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
107
108 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
109 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
110
111 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
112 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
113
114 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
115 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
116
117 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
118 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
119
120 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
121 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
122
123 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
124 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
125
126 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
127 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
128
129 vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
130 vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
131
132 _mm_storeu_ps(y, vf0123);
133 _mm_storeu_ps(y + 4, vf4567);
134 y += 8;
135 }
136 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
137 const __m128 vx = _mm_loadu_ps(x);
138 x += 4;
139
140 const __m128 vz = _mm_or_ps(vx, vsign_mask);
141
142 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
143 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
144
145 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
146 #if XNN_ARCH_X86_64
147 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
148 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
149 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
150 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
151 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
152 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
153 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
154 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
155 #else // !XNN_ARCH_X86_64
156 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
157 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
158 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
159 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
160 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
161 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
162 #endif // XNN_ARCH_X86_64
163 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
164
165 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
166 vn = _mm_sub_ps(vn, vmagic_bias);
167
168 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
169 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
170
171 __m128 vp = _mm_mul_ps(vt, vc2);
172 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
173
174 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
175
176 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
177 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
178 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
179 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
180
181 _mm_storeu_ps(y, vf);
182 y += 4;
183 }
184 if XNN_UNLIKELY(n != 0) {
185 const __m128 vx = _mm_loadu_ps(x);
186
187 const __m128 vz = _mm_or_ps(vx, vsign_mask);
188
189 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
190 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
191
192 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
193 #if XNN_ARCH_X86_64
194 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
195 const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
196 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
197 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
198 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
199 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
200 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
201 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
202 #else // !XNN_ARCH_X86_64
203 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
204 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
205 const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
206 const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
207 const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
208 const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
209 #endif // XNN_ARCH_X86_64
210 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
211
212 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
213 vn = _mm_sub_ps(vn, vmagic_bias);
214
215 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
216 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
217
218 __m128 vp = _mm_mul_ps(vt, vc2);
219 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
220
221 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
222
223 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
224 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
225 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
226 vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
227
228 if (n & (2 * sizeof(float))) {
229 _mm_storel_pi((__m64*) y, vf);
230 vf = _mm_movehl_ps(vf, vf);
231 y += 2;
232 }
233 if (n & (1 * sizeof(float))) {
234 _mm_store_ss(y, vf);
235 }
236 }
237 }
238