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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 x += 16;
44
45 const __m128 vz0123 = _mm_or_ps(vx0123, vsign_mask);
46 const __m128 vz4567 = _mm_or_ps(vx4567, vsign_mask);
47 const __m128 vz89AB = _mm_or_ps(vx89AB, vsign_mask);
48 const __m128 vzCDEF = _mm_or_ps(vxCDEF, vsign_mask);
49
50 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
51 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
52 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
53 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
54
55 const __m128i ve0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 17);
56 const __m128i ve4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 17);
57 const __m128i ve89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 17);
58 const __m128i veCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 17);
59
60 const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
61 const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
62 const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
63 const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
64
65 #if XNN_ARCH_X86_64
66 const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
67 const uint64_t vidx23 = (uint64_t) _mm_extract_epi64(vidx0123, 1);
68 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx01)));
69 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx23)));
70 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))), 1);
71 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))), 1);
72 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
73 const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
74 const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
75 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
76 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
77 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))), 1);
78 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))), 1);
79 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
80 const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
81 const uint64_t vidxAB = (uint64_t) _mm_extract_epi64(vidx89AB, 1);
82 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
83 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
84 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))), 1);
85 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))), 1);
86 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
87 const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
88 const uint64_t vidxEF = (uint64_t) _mm_extract_epi64(vidxCDEF, 1);
89 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxCD)));
90 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxEF)));
91 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxCD >> 32))), 1);
92 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxEF >> 32))), 1);
93 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
94 #else // !XNN_ARCH_X86_64
95 const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
96 const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
97 const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
98 const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
99 const __m128i vl0 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
100 const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
101 const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)), 1);
102 const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)), 1);
103 const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
104 const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
105 const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
106 const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
107 const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
108 const __m128i vl4 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
109 const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
110 const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)), 1);
111 const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)), 1);
112 const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
113 const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
114 const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
115 const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
116 const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
117 const __m128i vl8 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
118 const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
119 const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)), 1);
120 const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)), 1);
121 const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
122 const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
123 const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
124 const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
125 const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
126 const __m128i vlC = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxC)));
127 const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxE)));
128 const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxD)), 1);
129 const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxF)), 1);
130 const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
131 #endif // XNN_ARCH_X86_64
132
133 const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
134 const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
135 const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
136 const __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, veCDEF));
137
138 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
139 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
140 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
141 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
142
143 __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
144 __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
145 __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
146 __m128 vtCDEF = _mm_add_ps(vzCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_hi));
147
148 vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
149 vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
150 vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
151 vtCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vnCDEF, vminus_ln2_lo));
152
153 __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
154 __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
155 __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
156 __m128 vpCDEF = _mm_mul_ps(vtCDEF, vc2);
157
158 vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
159 vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
160 vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
161 vpCDEF = _mm_add_ps(vtCDEF, _mm_mul_ps(vpCDEF, vtCDEF));
162
163 const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
164 const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
165 const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
166 const __m128 vyCDEF = _mm_add_ps(vsCDEF, _mm_mul_ps(vsCDEF, vpCDEF));
167
168 __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
169 __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
170 __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
171 __m128 vfCDEF = _mm_div_ps(vyCDEF, _mm_add_ps(vyCDEF, vone));
172
173 vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
174 vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
175 vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
176 vfCDEF = _mm_andnot_ps(_mm_cmplt_ps(vzCDEF, vdenorm_cutoff), vfCDEF);
177
178 vf0123 = _mm_blendv_ps(_mm_sub_ps(vone, vf0123), vf0123, vx0123);
179 vf4567 = _mm_blendv_ps(_mm_sub_ps(vone, vf4567), vf4567, vx4567);
180 vf89AB = _mm_blendv_ps(_mm_sub_ps(vone, vf89AB), vf89AB, vx89AB);
181 vfCDEF = _mm_blendv_ps(_mm_sub_ps(vone, vfCDEF), vfCDEF, vxCDEF);
182
183 _mm_storeu_ps(y, vf0123);
184 _mm_storeu_ps(y + 4, vf4567);
185 _mm_storeu_ps(y + 8, vf89AB);
186 _mm_storeu_ps(y + 12, vfCDEF);
187 y += 16;
188 }
189 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
190 const __m128 vx = _mm_loadu_ps(x);
191 x += 4;
192
193 const __m128 vz = _mm_or_ps(vx, vsign_mask);
194
195 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
196 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
197
198 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
199 #if XNN_ARCH_X86_64
200 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
201 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
202 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
203 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
204 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);
205 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);
206 #else // !XNN_ARCH_X86_64
207 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
208 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
209 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);
210 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);
211 #endif // XNN_ARCH_X86_64
212 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
213
214 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
215 vn = _mm_sub_ps(vn, vmagic_bias);
216
217 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
218 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
219
220 __m128 vp = _mm_mul_ps(vt, vc2);
221 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
222
223 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
224
225 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
226 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
227 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
228
229 _mm_storeu_ps(y, vf);
230 y += 4;
231 }
232 if XNN_UNLIKELY(n != 0) {
233 const __m128 vx = _mm_loadu_ps(x);
234
235 const __m128 vz = _mm_or_ps(vx, vsign_mask);
236
237 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
238 const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
239
240 const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
241 #if XNN_ARCH_X86_64
242 const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
243 const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
244 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
245 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
246 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);
247 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);
248 #else // !XNN_ARCH_X86_64
249 const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
250 const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
251 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);
252 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);
253 #endif // XNN_ARCH_X86_64
254 const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
255
256 const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
257 vn = _mm_sub_ps(vn, vmagic_bias);
258
259 __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
260 vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
261
262 __m128 vp = _mm_mul_ps(vt, vc2);
263 vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
264
265 const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
266
267 __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
268 vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
269 vf = _mm_blendv_ps(_mm_sub_ps(vone, vf), vf, vx);
270
271 if (n & (2 * sizeof(float))) {
272 _mm_storel_pi((__m64*) y, vf);
273 vf = _mm_movehl_ps(vf, vf);
274 y += 2;
275 }
276 if (n & (1 * sizeof(float))) {
277 _mm_store_ss(y, vf);
278 }
279 }
280 }
281