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_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__sse41_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_extract_epi64(vidx0123, 1);
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 vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))), 1);
61 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))), 1);
62 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
63 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
64 const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
65 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
66 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
67 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))), 1);
68 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))), 1);
69 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
70 #else // !XNN_ARCH_X86_64
71 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
72 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
73 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
74 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
75 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
76 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
77 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)), 1);
78 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)), 1);
79 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
80 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
81 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
82 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
83 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
84 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
85 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
86 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)), 1);
87 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)), 1);
88 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
89 #endif // XNN_ARCH_X86_64
90
91 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
92 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
93
94 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
95 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
96
97 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
98 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
99
100 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
101 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
102
103 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
104 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
105
106 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
107 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
108
109 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
110 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
111
112 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
113 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
114
115 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
116 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
117
118 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
119 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
120
121 _mm_storeu_ps(y, vf0123);
122 _mm_storeu_ps(y + 4, vf4567);
123 y += 8;
124 }
125 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
126 const __m128 vx = _mm_loadu_ps(x);
127 x += 4;
128
129 const __m128 vz = _mm_or_ps(vx, vsign_mask);
130
131 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
132 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
133
134 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
135 #if XNN_ARCH_X86_64
136 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
137 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
138 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
139 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
140 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);
141 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);
142 #else // !XNN_ARCH_X86_64
143 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
144 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
145 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);
146 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);
147 #endif // XNN_ARCH_X86_64
148 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
149
150 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
151 vn = _mm_sub_ps(vn, vmagic_bias);
152
153 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
154 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
155
156 __m128 vp = _mm_mul_ps(vt, vc2);
157 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
158
159 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
160
161 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
162 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
163 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
164
165 _mm_storeu_ps(y, vf);
166 y += 4;
167 }
168 if XNN_UNLIKELY(n != 0) {
169 const __m128 vx = _mm_loadu_ps(x);
170
171 const __m128 vz = _mm_or_ps(vx, vsign_mask);
172
173 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
174 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
175
176 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
177 #if XNN_ARCH_X86_64
178 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
179 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
180 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
181 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
182 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);
183 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);
184 #else // !XNN_ARCH_X86_64
185 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
186 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
187 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);
188 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);
189 #endif // XNN_ARCH_X86_64
190 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
191
192 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
193 vn = _mm_sub_ps(vn, vmagic_bias);
194
195 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
196 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
197
198 __m128 vp = _mm_mul_ps(vt, vc2);
199 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
200
201 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
202
203 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
204 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
205 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
206
207 if (n & (2 * sizeof(float))) {
208 _mm_storel_pi((__m64*) y, vf);
209 vf = _mm_movehl_ps(vf, vf);
210 y += 2;
211 }
212 if (n & (1 * sizeof(float))) {
213 _mm_store_ss(y, vf);
214 }
215 }
216 }
217