xref: /aosp_15_r20/external/XNNPACK/src/f32-vsigmoid/gen/vsigmoid-sse2-rr2-lut64-p2-div-x12.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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_x12(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_x12(
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 >= 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_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
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 vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx01 >> 32))));
66       const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
67       const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx23 >> 32))));
68       const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
69       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
70       const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
71       const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
72       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx45)));
73       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx67)));
74       const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx45 >> 32))));
75       const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
76       const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx67 >> 32))));
77       const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
78       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
79       const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
80       const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
81       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx89)));
82       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidxAB)));
83       const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx89 >> 32))));
84       const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
85       const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidxAB >> 32))));
86       const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
87       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
88     #else  // !XNN_ARCH_X86_64
89       const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
90       const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
91       const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
92       const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
93       const __m128i vl0   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx0)));
94       const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx2)));
95       const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx1)));
96       const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
97       const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx3)));
98       const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
99       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
100       const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
101       const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
102       const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
103       const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
104       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx4)));
105       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx6)));
106       const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx5)));
107       const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
108       const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx7)));
109       const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
110       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
111       const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
112       const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
113       const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
114       const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
115       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx8)));
116       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxA)));
117       const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidx9)));
118       const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
119       const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + vidxB)));
120       const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
121       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
122     #endif  // XNN_ARCH_X86_64
123 
124     const __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ve0123));
125     const __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ve4567));
126     const __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ve89AB));
127 
128     vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
129     vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
130     vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
131 
132     __m128 vt0123 = _mm_add_ps(vz0123, _mm_mul_ps(vn0123, vminus_ln2_hi));
133     __m128 vt4567 = _mm_add_ps(vz4567, _mm_mul_ps(vn4567, vminus_ln2_hi));
134     __m128 vt89AB = _mm_add_ps(vz89AB, _mm_mul_ps(vn89AB, vminus_ln2_hi));
135 
136     vt0123 = _mm_add_ps(vt0123, _mm_mul_ps(vn0123, vminus_ln2_lo));
137     vt4567 = _mm_add_ps(vt4567, _mm_mul_ps(vn4567, vminus_ln2_lo));
138     vt89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vn89AB, vminus_ln2_lo));
139 
140     __m128 vp0123 = _mm_mul_ps(vt0123, vc2);
141     __m128 vp4567 = _mm_mul_ps(vt4567, vc2);
142     __m128 vp89AB = _mm_mul_ps(vt89AB, vc2);
143 
144     vp0123 = _mm_add_ps(vt0123, _mm_mul_ps(vp0123, vt0123));
145     vp4567 = _mm_add_ps(vt4567, _mm_mul_ps(vp4567, vt4567));
146     vp89AB = _mm_add_ps(vt89AB, _mm_mul_ps(vp89AB, vt89AB));
147 
148     const __m128 vy0123 = _mm_add_ps(vs0123, _mm_mul_ps(vs0123, vp0123));
149     const __m128 vy4567 = _mm_add_ps(vs4567, _mm_mul_ps(vs4567, vp4567));
150     const __m128 vy89AB = _mm_add_ps(vs89AB, _mm_mul_ps(vs89AB, vp89AB));
151 
152     __m128 vf0123 = _mm_div_ps(vy0123, _mm_add_ps(vy0123, vone));
153     __m128 vf4567 = _mm_div_ps(vy4567, _mm_add_ps(vy4567, vone));
154     __m128 vf89AB = _mm_div_ps(vy89AB, _mm_add_ps(vy89AB, vone));
155 
156     vf0123 = _mm_andnot_ps(_mm_cmplt_ps(vz0123, vdenorm_cutoff), vf0123);
157     vf4567 = _mm_andnot_ps(_mm_cmplt_ps(vz4567, vdenorm_cutoff), vf4567);
158     vf89AB = _mm_andnot_ps(_mm_cmplt_ps(vz89AB, vdenorm_cutoff), vf89AB);
159 
160     const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
161     const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
162     const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
163 
164     vf0123 = _mm_or_ps(_mm_and_ps(vf0123, vm0123), _mm_andnot_ps(vm0123, _mm_sub_ps(vone, vf0123)));
165     vf4567 = _mm_or_ps(_mm_and_ps(vf4567, vm4567), _mm_andnot_ps(vm4567, _mm_sub_ps(vone, vf4567)));
166     vf89AB = _mm_or_ps(_mm_and_ps(vf89AB, vm89AB), _mm_andnot_ps(vm89AB, _mm_sub_ps(vone, vf89AB)));
167 
168     _mm_storeu_ps(y, vf0123);
169     _mm_storeu_ps(y + 4, vf4567);
170     _mm_storeu_ps(y + 8, vf89AB);
171     y += 12;
172   }
173   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
174     const __m128 vx = _mm_loadu_ps(x);
175     x += 4;
176 
177     const __m128 vz = _mm_or_ps(vx, vsign_mask);
178 
179     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
180     const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
181 
182     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
183     #if XNN_ARCH_X86_64
184       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
185       const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
186       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
187       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
188       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
189       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
190       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
191       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
192     #else  // !XNN_ARCH_X86_64
193       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
194       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
195       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
196       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
197       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
198       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
199     #endif  // XNN_ARCH_X86_64
200     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
201 
202     const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
203     vn = _mm_sub_ps(vn, vmagic_bias);
204 
205     __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
206     vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
207 
208     __m128 vp = _mm_mul_ps(vt, vc2);
209     vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
210 
211     const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
212 
213     __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
214     vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
215     const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
216     vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
217 
218     _mm_storeu_ps(y, vf);
219     y += 4;
220   }
221   if XNN_UNLIKELY(n != 0) {
222     const __m128 vx = _mm_loadu_ps(x);
223 
224     const __m128 vz = _mm_or_ps(vx, vsign_mask);
225 
226     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
227     const __m128i ve = _mm_slli_epi32(_mm_castps_si128(vn), 17);
228 
229     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
230     #if XNN_ARCH_X86_64
231       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
232       const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
233       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_lo)));
234       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) vidx_hi)));
235       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_lo >> 32))));
236       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
237       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) (vidx_hi >> 32))));
238       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
239     #else  // !XNN_ARCH_X86_64
240       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_cvtsi128_si32(vidx))));
241       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 4))));
242       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 2))));
243       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
244       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi16(vidx, 6))));
245       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
246     #endif  // XNN_ARCH_X86_64
247     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
248 
249     const __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ve));
250     vn = _mm_sub_ps(vn, vmagic_bias);
251 
252     __m128 vt = _mm_add_ps(vz, _mm_mul_ps(vn, vminus_ln2_hi));
253     vt = _mm_add_ps(vt, _mm_mul_ps(vn, vminus_ln2_lo));
254 
255     __m128 vp = _mm_mul_ps(vt, vc2);
256     vp = _mm_add_ps(vt, _mm_mul_ps(vp, vt));
257 
258     const __m128 vy = _mm_add_ps(vs, _mm_mul_ps(vs, vp));
259 
260     __m128 vf = _mm_div_ps(vy, _mm_add_ps(vy, vone));
261     vf = _mm_andnot_ps(_mm_cmplt_ps(vz, vdenorm_cutoff), vf);
262     const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
263     vf = _mm_or_ps(_mm_and_ps(vf, vm), _mm_andnot_ps(vm, _mm_sub_ps(vone, vf)));
264 
265     if (n & (2 * sizeof(float))) {
266       _mm_storel_pi((__m64*) y, vf);
267       vf = _mm_movehl_ps(vf, vf);
268       y += 2;
269     }
270     if (n & (1 * sizeof(float))) {
271       _mm_store_ss(y, vf);
272     }
273   }
274 }
275