• 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 <smmintrin.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__sse41_rr2_lut16_p3_x16(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__sse41_rr2_lut16_p3_x16(
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_DISABLE_TSAN
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->sse.prescale);
32   const __m128 valpha = _mm_load_ps(params->sse.alpha);
33   const __m128 vbeta = _mm_load_ps(params->sse.beta);
34 
35   const __m128 vsat_cutoff = _mm_set1_ps(-0x1.154246p+4f);
36   const __m128 vmagic_bias = _mm_set1_ps(0x1.800000p19f);
37   const __m128 vlog2e = _mm_set1_ps(0x1.715476p+0f);
38   const __m128i vindex_mask = _mm_set1_epi32(0xF);
39   const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E400p-1f);
40   const __m128 vminus_ln2_lo = _mm_set1_ps(-0x1.7F7D1Cp-20f);
41   const __m128 vc3 = _mm_set1_ps(0x1.55561Cp-3f);
42   const __m128 vc2 = _mm_set1_ps(0x1.0001ECp-1f);
43   const __m128 vone = _mm_set1_ps(1.0f);
44 
45   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
46     __m128 vx0123 = _mm_loadu_ps(x);
47     __m128 vx4567 = _mm_loadu_ps(x + 4);
48     __m128 vx89AB = _mm_loadu_ps(x + 8);
49     __m128 vxCDEF = _mm_loadu_ps(x + 12);
50     x += 16;
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 
57     __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
58     __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
59     __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
60     __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
61 
62     const __m128i vidx0123 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn0123), vindex_mask), 2);
63     const __m128i ven0123 = _mm_slli_epi32(_mm_castps_si128(vn0123), 19);
64     const __m128i vidx4567 = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn4567), vindex_mask), 2);
65     const __m128i ven4567 = _mm_slli_epi32(_mm_castps_si128(vn4567), 19);
66     const __m128i vidx89AB = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn89AB), vindex_mask), 2);
67     const __m128i ven89AB = _mm_slli_epi32(_mm_castps_si128(vn89AB), 19);
68     const __m128i vidxCDEF = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vnCDEF), vindex_mask), 2);
69     const __m128i venCDEF = _mm_slli_epi32(_mm_castps_si128(vnCDEF), 19);
70 
71     #if XNN_ARCH_X86_64
72       const uint64_t vidx01 = (uint64_t) _mm_cvtsi128_si64(vidx0123);
73       const uint64_t vidx23 = (uint64_t) _mm_extract_epi64(vidx0123, 1);
74       const __m128i vl0   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx01)));
75       const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx23)));
76       const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx01 >> 32))), 1);
77       const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx23 >> 32))), 1);
78       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
79       const uint64_t vidx45 = (uint64_t) _mm_cvtsi128_si64(vidx4567);
80       const uint64_t vidx67 = (uint64_t) _mm_extract_epi64(vidx4567, 1);
81       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx45)));
82       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx67)));
83       const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx45 >> 32))), 1);
84       const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx67 >> 32))), 1);
85       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
86       const uint64_t vidx89 = (uint64_t) _mm_cvtsi128_si64(vidx89AB);
87       const uint64_t vidxAB = (uint64_t) _mm_extract_epi64(vidx89AB, 1);
88       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx89)));
89       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxAB)));
90       const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx89 >> 32))), 1);
91       const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxAB >> 32))), 1);
92       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
93       const uint64_t vidxCD = (uint64_t) _mm_cvtsi128_si64(vidxCDEF);
94       const uint64_t vidxEF = (uint64_t) _mm_extract_epi64(vidxCDEF, 1);
95       const __m128i vlC   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxCD)));
96       const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidxEF)));
97       const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxCD >> 32))), 1);
98       const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidxEF >> 32))), 1);
99       const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
100     #else  // !XNN_ARCH_X86_64
101       const uint32_t vidx0 = (uint32_t) _mm_cvtsi128_si32(vidx0123);
102       const uint32_t vidx1 = (uint32_t) _mm_extract_epi16(vidx0123, 2);
103       const uint32_t vidx2 = (uint32_t) _mm_extract_epi16(vidx0123, 4);
104       const uint32_t vidx3 = (uint32_t) _mm_extract_epi16(vidx0123, 6);
105       const __m128i vl0   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx0)));
106       const __m128i vl2 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx2)));
107       const __m128i vl01 = _mm_insert_epi32(vl0, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx1)), 1);
108       const __m128i vl23 = _mm_insert_epi32(vl2, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx3)), 1);
109       const __m128i vl0123 = _mm_unpacklo_epi64(vl01, vl23);
110       const uint32_t vidx4 = (uint32_t) _mm_cvtsi128_si32(vidx4567);
111       const uint32_t vidx5 = (uint32_t) _mm_extract_epi16(vidx4567, 2);
112       const uint32_t vidx6 = (uint32_t) _mm_extract_epi16(vidx4567, 4);
113       const uint32_t vidx7 = (uint32_t) _mm_extract_epi16(vidx4567, 6);
114       const __m128i vl4   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx4)));
115       const __m128i vl6 = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx6)));
116       const __m128i vl45 = _mm_insert_epi32(vl4, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx5)), 1);
117       const __m128i vl67 = _mm_insert_epi32(vl6, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx7)), 1);
118       const __m128i vl4567 = _mm_unpacklo_epi64(vl45, vl67);
119       const uint32_t vidx8 = (uint32_t) _mm_cvtsi128_si32(vidx89AB);
120       const uint32_t vidx9 = (uint32_t) _mm_extract_epi16(vidx89AB, 2);
121       const uint32_t vidxA = (uint32_t) _mm_extract_epi16(vidx89AB, 4);
122       const uint32_t vidxB = (uint32_t) _mm_extract_epi16(vidx89AB, 6);
123       const __m128i vl8   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx8)));
124       const __m128i vlA = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxA)));
125       const __m128i vl89 = _mm_insert_epi32(vl8, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidx9)), 1);
126       const __m128i vlAB = _mm_insert_epi32(vlA, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxB)), 1);
127       const __m128i vl89AB = _mm_unpacklo_epi64(vl89, vlAB);
128       const uint32_t vidxC = (uint32_t) _mm_cvtsi128_si32(vidxCDEF);
129       const uint32_t vidxD = (uint32_t) _mm_extract_epi16(vidxCDEF, 2);
130       const uint32_t vidxE = (uint32_t) _mm_extract_epi16(vidxCDEF, 4);
131       const uint32_t vidxF = (uint32_t) _mm_extract_epi16(vidxCDEF, 6);
132       const __m128i vlC   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxC)));
133       const __m128i vlE = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxE)));
134       const __m128i vlCD = _mm_insert_epi32(vlC, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxD)), 1);
135       const __m128i vlEF = _mm_insert_epi32(vlE, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + vidxF)), 1);
136       const __m128i vlCDEF = _mm_unpacklo_epi64(vlCD, vlEF);
137     #endif  // XNN_ARCH_X86_64
138 
139     vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
140     __m128 vs0123 = _mm_castsi128_ps(_mm_add_epi32(vl0123, ven0123));
141     vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
142     __m128 vs4567 = _mm_castsi128_ps(_mm_add_epi32(vl4567, ven4567));
143     vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
144     __m128 vs89AB = _mm_castsi128_ps(_mm_add_epi32(vl89AB, ven89AB));
145     vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
146     __m128 vsCDEF = _mm_castsi128_ps(_mm_add_epi32(vlCDEF, venCDEF));
147 
148     __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
149     __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
150     __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
151     __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
152 
153     vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
154     vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
155     vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
156     vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
157 
158     __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc3, vt0123), vc2);
159     __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc3, vt4567), vc2);
160     __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc3, vt89AB), vc2);
161     __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc3, vtCDEF), vc2);
162 
163     vp0123 = _mm_mul_ps(vp0123, vt0123);
164     vp4567 = _mm_mul_ps(vp4567, vt4567);
165     vp89AB = _mm_mul_ps(vp89AB, vt89AB);
166     vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
167 
168     vt0123 = _mm_mul_ps(vt0123, vs0123);
169     vs0123 = _mm_sub_ps(vs0123, vone);
170     vt4567 = _mm_mul_ps(vt4567, vs4567);
171     vs4567 = _mm_sub_ps(vs4567, vone);
172     vt89AB = _mm_mul_ps(vt89AB, vs89AB);
173     vs89AB = _mm_sub_ps(vs89AB, vone);
174     vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
175     vsCDEF = _mm_sub_ps(vsCDEF, vone);
176 
177     vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
178     vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
179     vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
180     vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
181 
182     const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
183     const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
184     const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
185     const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
186 
187     vx0123 = _mm_mul_ps(vx0123, vbeta);
188     vx4567 = _mm_mul_ps(vx4567, vbeta);
189     vx89AB = _mm_mul_ps(vx89AB, vbeta);
190     vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
191 
192     const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
193     const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
194     const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
195     const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
196 
197     _mm_storeu_ps(y, vy0123);
198     _mm_storeu_ps(y + 4, vy4567);
199     _mm_storeu_ps(y + 8, vy89AB);
200     _mm_storeu_ps(y + 12, vyCDEF);
201     y += 16;
202   }
203   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
204     __m128 vx = _mm_loadu_ps(x);
205     x += 4;
206 
207     const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
208 
209     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
210 
211     const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
212     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
213     #if XNN_ARCH_X86_64
214       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
215       const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
216       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
217       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
218       const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))), 1);
219       const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))), 1);
220     #else  // !XNN_ARCH_X86_64
221       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
222       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
223       const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
224       const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
225     #endif  // XNN_ARCH_X86_64
226     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
227     __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
228     vn = _mm_sub_ps(vn, vmagic_bias);
229 
230     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
231     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
232 
233     __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
234     vp = _mm_mul_ps(vp, vt);
235 
236     vt = _mm_mul_ps(vt, vs);
237     vs = _mm_sub_ps(vs, vone);
238     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
239     const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
240 
241     vx = _mm_mul_ps(vx, vbeta);
242     const __m128 vy = _mm_blendv_ps(vx, ve, vx);
243 
244     _mm_storeu_ps(y, vy);
245     y += 4;
246   }
247   if XNN_UNLIKELY(n != 0) {
248     __m128 vx = _mm_loadu_ps(x);
249 
250     const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
251 
252     __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
253 
254     const __m128i ven = _mm_slli_epi32(_mm_castps_si128(vn), 19);
255     const __m128i vidx = _mm_slli_epi32(_mm_and_si128(_mm_castps_si128(vn), vindex_mask), 2);
256     #if XNN_ARCH_X86_64
257       const uint64_t vidx_lo = (uint64_t) _mm_cvtsi128_si64(vidx);
258       const uint64_t vidx_hi = (uint64_t) _mm_extract_epi64(vidx, 1);
259       const __m128i vl_ll   = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lo)));
260       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hi)));
261       const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lo >> 32))), 1);
262       const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hi >> 32))), 1);
263     #else  // !XNN_ARCH_X86_64
264       const __m128i vl_ll = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx))));
265       const __m128i vl_hl = _mm_cvtsi32_si128(*((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 4))));
266       const __m128i vl_lo = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 2))), 1);
267       const __m128i vl_hi = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi16(vidx, 6))), 1);
268     #endif  // XNN_ARCH_X86_64
269     const __m128i vl = _mm_unpacklo_epi64(vl_lo, vl_hi);
270     __m128 vs = _mm_castsi128_ps(_mm_add_epi32(vl, ven));
271     vn = _mm_sub_ps(vn, vmagic_bias);
272 
273     __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
274     vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
275 
276     __m128 vp = _mm_add_ps(_mm_mul_ps(vc3, vt), vc2);
277     vp = _mm_mul_ps(vp, vt);
278 
279     vt = _mm_mul_ps(vt, vs);
280     vs = _mm_sub_ps(vs, vone);
281     vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
282     const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
283 
284     vx = _mm_mul_ps(vx, vbeta);
285     __m128 vy = _mm_blendv_ps(vx, ve, vx);
286 
287     if (n & (2 * sizeof(float))) {
288       _mm_storel_pi((__m64*) y, vy);
289       vy = _mm_movehl_ps(vy, vy);
290       y += 2;
291     }
292     if (n & (1 * sizeof(float))) {
293       _mm_store_ss(y, vy);
294     }
295   }
296 }
297