• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/sse-rr2-lut16-p3.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/vunary.h>
15 #include <xnnpack/common.h>
16 
17 
18 extern XNN_INTERNAL const float xnn_table_exp2minus_k_over_16[16];
19 
xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x20(
21     size_t n,
22     const float* x,
23     float* y,
24     const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
25 {
26   assert(n != 0);
27   assert(n % sizeof(float) == 0);
28   assert(x != NULL);
29   assert(y != NULL);
30 
31   const __m128 vprescale = _mm_load_ps(params->sse2_rr2_lut16_p3.prescale);
32   const __m128 valpha = _mm_load_ps(params->sse2_rr2_lut16_p3.alpha);
33   const __m128 vbeta = _mm_load_ps(params->sse2_rr2_lut16_p3.beta);
34   const __m128 vsat_cutoff = _mm_load_ps(params->sse2_rr2_lut16_p3.sat_cutoff);
35   const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_lut16_p3.magic_bias);
36   const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_lut16_p3.log2e);
37   const __m128i vindex_mask = _mm_load_si128((const __m128i*) params->sse2_rr2_lut16_p3.index_mask);
38   const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_hi);
39   const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_lut16_p3.minus_ln2_lo);
40   const __m128 vc3 = _mm_load_ps(params->sse2_rr2_lut16_p3.c3);
41   const __m128 vc2 = _mm_load_ps(params->sse2_rr2_lut16_p3.c2);
42   const __m128 vone = _mm_load_ps(params->sse2_rr2_lut16_p3.one);
43 
44   for (; n >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
45     __m128 vx0123 = _mm_loadu_ps(x);
46     __m128 vx4567 = _mm_loadu_ps(x + 4);
47     __m128 vx89AB = _mm_loadu_ps(x + 8);
48     __m128 vxCDEF = _mm_loadu_ps(x + 12);
49     __m128 vxGHIJ = _mm_loadu_ps(x + 16);
50     x += 20;
51 
52     const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
53     const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
54     const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
55     const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
56     const __m128 vzGHIJ = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxGHIJ, vprescale));
57 
58     __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
59     __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
60     __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
61     __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
62     __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
63 
64     const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
65     const __m128i ven0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 19);
66     const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
67     const __m128i ven4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 19);
68     const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
69     const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19);
70     const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
71     const __m128i venCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 19);
72     const __m128i vidxGHIJ = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnGHIJ), vindex_mask), 2);
73     const __m128i venGHIJ = _mm_slli_epi32(_mm_castps_si128(vnGHIJ), 19);
74 
75     #if XNN_ARCH_X86_64
76       const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
77       const uint64_t vidx23 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx0123, vidx0123));
78       const __m128i vl0   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01)));
79       const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)));
80       const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32))));
81       const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
82       const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32))));
83       const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
84       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
85       const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
86       const uint64_t vidx67 = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx4567, vidx4567));
87       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45)));
88       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)));
89       const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32))));
90       const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
91       const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32))));
92       const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
93       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
94       const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
95       const uint64_t vidxAB = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx89AB, vidx89AB));
96       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89)));
97       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB)));
98       const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32))));
99       const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
100       const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32))));
101       const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
102       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
103       const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
104       const uint64_t vidxEF = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxCDEF, vidxCDEF));
105       const __m128i vlC   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD)));
106       const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF)));
107       const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32))));
108       const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
109       const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32))));
110       const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
111       const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
112       const uint64_t vidxGH = (uint64_t) _mm_cvtsi128_si64(vidxGHIJ);
113       const uint64_t vidxIJ = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidxGHIJ, vidxGHIJ));
114       const __m128i vlG   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxGH)));
115       const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxIJ)));
116       const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxGH >> 32))));
117       const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
118       const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxIJ >> 32))));
119       const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
120       const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
121     #else  // !XNN_ARCH_X86_64
122       const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
123       const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
124       const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
125       const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
126       const __m128i vl0   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx0)));
127       const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx2)));
128       const __m128i vl1 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx1)));
129       const __m128i vl01 = _mm_unpacklo_epi32(vl0, vl1);
130       const __m128i vl3 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx3)));
131       const __m128i vl23 = _mm_unpacklo_epi32(vl2, vl3);
132       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
133       const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
134       const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
135       const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
136       const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
137       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx4)));
138       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx6)));
139       const __m128i vl5 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx5)));
140       const __m128i vl45 = _mm_unpacklo_epi32(vl4, vl5);
141       const __m128i vl7 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx7)));
142       const __m128i vl67 = _mm_unpacklo_epi32(vl6, vl7);
143       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
144       const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
145       const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
146       const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
147       const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
148       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx8)));
149       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxA)));
150       const __m128i vl9 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx9)));
151       const __m128i vl89 = _mm_unpacklo_epi32(vl8, vl9);
152       const __m128i vlB = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)));
153       const __m128i vlAB = _mm_unpacklo_epi32(vlA, vlB);
154       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
155       const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
156       const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
157       const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
158       const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
159       const __m128i vlC   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxC)));
160       const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxE)));
161       const __m128i vlD = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxD)));
162       const __m128i vlCD = _mm_unpacklo_epi32(vlC, vlD);
163       const __m128i vlF = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxF)));
164       const __m128i vlEF = _mm_unpacklo_epi32(vlE, vlF);
165       const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
166       const uint32_t vidxG = (uint32_t) _mm_cvtsi128_si32(vidxGHIJ);
167       const uint32_t vidxH = (uint32_t) _mm_extract_epi16(vidxGHIJ, 2);
168       const uint32_t vidxI = (uint32_t) _mm_extract_epi16(vidxGHIJ, 4);
169       const uint32_t vidxJ = (uint32_t) _mm_extract_epi16(vidxGHIJ, 6);
170       const __m128i vlG   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxG)));
171       const __m128i vlI = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxI)));
172       const __m128i vlH = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxH)));
173       const __m128i vlGH = _mm_unpacklo_epi32(vlG, vlH);
174       const __m128i vlJ = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxJ)));
175       const __m128i vlIJ = _mm_unpacklo_epi32(vlI, vlJ);
176       const __m128i vlGHIJ = _mm_unpacklo_epi64(vlGH, vlIJ);
177     #endif  // XNN_ARCH_X86_64
178 
179     vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
180     __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ven0123));
181     vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
182     __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ven4567));
183     vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
184     __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB));
185     vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
186     __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, venCDEF));
187     vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
188     __m128 vsGHIJ = _mm_castsi128_ps(_mm_add_epi32(vlGHIJ, venGHIJ));
189 
190     __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
191     __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
192     __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
193     __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
194     __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
195 
196     vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
197     vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
198     vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
199     vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
200     vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
201 
202     __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc3, vt0123), vc2);
203     __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc3, vt4567), vc2);
204     __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc3, vt89AB), vc2);
205     __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc3, vtCDEF), vc2);
206     __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc3, vtGHIJ), vc2);
207 
208     vp0123 = _mm_mul_ps(vp0123, vt0123);
209     vp4567 = _mm_mul_ps(vp4567, vt4567);
210     vp89AB = _mm_mul_ps(vp89AB, vt89AB);
211     vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
212     vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ);
213 
214     vt0123 = _mm_mul_ps(vt0123, vs0123);
215     vs0123 = _mm_sub_ps(vs0123, vone);
216     vt4567 = _mm_mul_ps(vt4567, vs4567);
217     vs4567 = _mm_sub_ps(vs4567, vone);
218     vt89AB = _mm_mul_ps(vt89AB, vs89AB);
219     vs89AB = _mm_sub_ps(vs89AB, vone);
220     vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
221     vsCDEF = _mm_sub_ps(vsCDEF, vone);
222     vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
223     vsGHIJ = _mm_sub_ps(vsGHIJ, vone);
224 
225     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
226     vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
227     vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
228     vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
229     vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ);
230 
231     const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
232     const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
233     const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
234     const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
235     const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha);
236 
237     const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
238     vx0123 = _mm_mul_ps(vx0123, vbeta);
239     const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
240     vx4567 = _mm_mul_ps(vx4567, vbeta);
241     const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
242     vx89AB = _mm_mul_ps(vx89AB, vbeta);
243     const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
244     vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
245     const __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ)));
246     vxGHIJ = _mm_mul_ps(vxGHIJ, vbeta);
247 
248     const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
249     const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
250     const __m128 vy89AB = _mm_or_ps(_mm_and_ps(ve89AB, vm89AB), _mm_andnot_ps(vm89AB, vx89AB));
251     const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF));
252     const __m128 vyGHIJ = _mm_or_ps(_mm_and_ps(veGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, vxGHIJ));
253 
254     _mm_storeu_ps(y, vy0123);
255     _mm_storeu_ps(y + 4, vy4567);
256     _mm_storeu_ps(y + 8, vy89AB);
257     _mm_storeu_ps(y + 12, vyCDEF);
258     _mm_storeu_ps(y + 16, vyGHIJ);
259     y += 20;
260   }
261   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
262     __m128 vx = _mm_loadu_ps(x);
263     x += 4;
264 
265     const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
266 
267     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
268 
269     const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
270     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
271     #if XNN_ARCH_X86_64
272       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
273       const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
274       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
275       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
276       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
277       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
278       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
279       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
280     #else  // !XNN_ARCH_X86_64
281       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
282       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
283       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
284       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
285       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
286       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
287     #endif  // XNN_ARCH_X86_64
288     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
289     __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
290     vn = _mm_sub_ps(vn, vmagic_bias);
291 
292     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
293     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
294 
295     __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
296     vp = _mm_mul_ps(vp, vt);
297 
298     vt = _mm_mul_ps(vt, vs);
299     vs = _mm_sub_ps(vs, vone);
300     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
301     const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
302 
303     const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
304     vx = _mm_mul_ps(vx, vbeta);
305     const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
306 
307     _mm_storeu_ps(y, vy);
308     y += 4;
309   }
310   if XNN_UNLIKELY(n != 0) {
311     __m128 vx = _mm_loadu_ps(x);
312 
313     const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
314 
315     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
316 
317     const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
318     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
319     #if XNN_ARCH_X86_64
320       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
321       const uint64_t vidx_hi = (uint64_t) _mm_cvtsi128_si64(_mm_unpackhi_epi64(vidx, vidx));
322       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
323       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
324       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))));
325       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
326       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))));
327       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
328     #else  // !XNN_ARCH_X86_64
329       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
330       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
331       const __m128i vl_lh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))));
332       const __m128i vl_lo = _mm_unpacklo_epi32(vl_ll, vl_lh);
333       const __m128i vl_hh = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))));
334       const __m128i vl_hi = _mm_unpacklo_epi32(vl_hl, vl_hh);
335     #endif  // XNN_ARCH_X86_64
336     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
337     __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
338     vn = _mm_sub_ps(vn, vmagic_bias);
339 
340     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
341     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
342 
343     __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
344     vp = _mm_mul_ps(vp, vt);
345 
346     vt = _mm_mul_ps(vt, vs);
347     vs = _mm_sub_ps(vs, vone);
348     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
349     const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
350 
351     const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
352     vx = _mm_mul_ps(vx, vbeta);
353     __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
354 
355     if (n & (2 * sizeof(float))) {
356       _mm_storel_pi((__m64*) y, vy);
357       vy = _mm_movehl_ps(vy, vy);
358       y += 2;
359     }
360     if (n & (1 * sizeof(float))) {
361       _mm_store_ss(y, vy);
362     }
363   }
364 }
365