• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/avx-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 <immintrin.h>
13 
14 #include <xnnpack/common.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17 
18 
19 extern XNN_INTERNAL const int xnn_table_exp2minus_k_over_16[16];
20 
21 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
22 
xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24(
24     size_t n,
25     const float* x,
26     float* y,
27     const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
28 {
29   assert(n % sizeof(float) == 0);
30 
31   const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
32   const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
33   const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
34 
35   const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
36   const __m256 vmagic_bias = _mm256_set1_ps(0x1.800000p19f);
37   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
38   const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0xF));
39   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
40   const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
41   const __m256 vc3 = _mm256_set1_ps(0x1.55561Cp-3f);
42   const __m256 vc2 = _mm256_set1_ps(0x1.0001ECp-1f);
43   const __m256 vone = _mm256_set1_ps(1.0f);
44 
45   for (; n >= 24 * sizeof(float); n -= 24 * sizeof(float)) {
46     __m256 vx0 = _mm256_loadu_ps(x);
47     __m256 vx1 = _mm256_loadu_ps(x + 8);
48     __m256 vx2 = _mm256_loadu_ps(x + 16);
49     x += 24;
50 
51     const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
52     const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
53     const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
54 
55     __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
56     __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
57     __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
58 
59     const __m256 vidx0 = _mm256_and_ps(vn0, vindex_mask);
60 
61     const __m128i vidx0_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx0)), 2);
62     const __m128i vidx0_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx0, 1)), 2);
63     #if XNN_ARCH_X86_64
64       const uint64_t vidx0_ll = (uint64_t) _mm_cvtsi128_si64(vidx0_lo);
65       const uint64_t vidx0_lh = (uint64_t) _mm_extract_epi64(vidx0_lo, 1);
66       const uint64_t vidx0_hl = (uint64_t) _mm_cvtsi128_si64(vidx0_hi);
67       const uint64_t vidx0_hh = (uint64_t) _mm_extract_epi64(vidx0_hi, 1);
68       __m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0_ll));
69       __m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0_lh));
70       __m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0_hl));
71       __m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx0_hh));
72       vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx0_ll >> 32))), 1);
73       vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx0_lh >> 32))), 1);
74       vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx0_hl >> 32))), 1);
75       vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx0_hh >> 32))), 1);
76     #else
77       __m128i vl0_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_lo)));
78       __m128i vl0_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)));
79       __m128i vl0_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx0_hi)));
80       __m128i vl0_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)));
81       vl0_ll = _mm_insert_epi32(vl0_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1))), 1);
82       vl0_lh = _mm_insert_epi32(vl0_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3))), 1);
83       vl0_hl = _mm_insert_epi32(vl0_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1))), 1);
84       vl0_hh = _mm_insert_epi32(vl0_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3))), 1);
85     #endif
86     const __m128i vl0_lo = _mm_unpacklo_epi64(vl0_ll, vl0_lh);
87     const __m128i vl0_hi = _mm_unpacklo_epi64(vl0_hl, vl0_hh);
88     const __m256 vidx1 = _mm256_and_ps(vn1, vindex_mask);
89 
90     const __m128i vidx1_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx1)), 2);
91     const __m128i vidx1_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx1, 1)), 2);
92     #if XNN_ARCH_X86_64
93       const uint64_t vidx1_ll = (uint64_t) _mm_cvtsi128_si64(vidx1_lo);
94       const uint64_t vidx1_lh = (uint64_t) _mm_extract_epi64(vidx1_lo, 1);
95       const uint64_t vidx1_hl = (uint64_t) _mm_cvtsi128_si64(vidx1_hi);
96       const uint64_t vidx1_hh = (uint64_t) _mm_extract_epi64(vidx1_hi, 1);
97       __m128i vl1_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1_ll));
98       __m128i vl1_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1_lh));
99       __m128i vl1_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1_hl));
100       __m128i vl1_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx1_hh));
101       vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx1_ll >> 32))), 1);
102       vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx1_lh >> 32))), 1);
103       vl1_hl = _mm_insert_epi32(vl1_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx1_hl >> 32))), 1);
104       vl1_hh = _mm_insert_epi32(vl1_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx1_hh >> 32))), 1);
105     #else
106       __m128i vl1_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_lo)));
107       __m128i vl1_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)));
108       __m128i vl1_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx1_hi)));
109       __m128i vl1_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)));
110       vl1_ll = _mm_insert_epi32(vl1_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1))), 1);
111       vl1_lh = _mm_insert_epi32(vl1_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3))), 1);
112       vl1_hl = _mm_insert_epi32(vl1_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 1))), 1);
113       vl1_hh = _mm_insert_epi32(vl1_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 3))), 1);
114     #endif
115     const __m128i vl1_lo = _mm_unpacklo_epi64(vl1_ll, vl1_lh);
116     const __m128i vl1_hi = _mm_unpacklo_epi64(vl1_hl, vl1_hh);
117     const __m256 vidx2 = _mm256_and_ps(vn2, vindex_mask);
118 
119     const __m128i vidx2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx2)), 2);
120     const __m128i vidx2_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx2, 1)), 2);
121     #if XNN_ARCH_X86_64
122       const uint64_t vidx2_ll = (uint64_t) _mm_cvtsi128_si64(vidx2_lo);
123       const uint64_t vidx2_lh = (uint64_t) _mm_extract_epi64(vidx2_lo, 1);
124       const uint64_t vidx2_hl = (uint64_t) _mm_cvtsi128_si64(vidx2_hi);
125       const uint64_t vidx2_hh = (uint64_t) _mm_extract_epi64(vidx2_hi, 1);
126       __m128i vl2_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2_ll));
127       __m128i vl2_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2_lh));
128       __m128i vl2_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2_hl));
129       __m128i vl2_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx2_hh));
130       vl2_ll = _mm_insert_epi32(vl2_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx2_ll >> 32))), 1);
131       vl2_lh = _mm_insert_epi32(vl2_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx2_lh >> 32))), 1);
132       vl2_hl = _mm_insert_epi32(vl2_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx2_hl >> 32))), 1);
133       vl2_hh = _mm_insert_epi32(vl2_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx2_hh >> 32))), 1);
134     #else
135       __m128i vl2_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_lo)));
136       __m128i vl2_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 2)));
137       __m128i vl2_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx2_hi)));
138       __m128i vl2_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_hi, 2)));
139       vl2_ll = _mm_insert_epi32(vl2_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 1))), 1);
140       vl2_lh = _mm_insert_epi32(vl2_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_lo, 3))), 1);
141       vl2_hl = _mm_insert_epi32(vl2_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_hi, 1))), 1);
142       vl2_hh = _mm_insert_epi32(vl2_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx2_hi, 3))), 1);
143     #endif
144     const __m128i vl2_lo = _mm_unpacklo_epi64(vl2_ll, vl2_lh);
145     const __m128i vl2_hi = _mm_unpacklo_epi64(vl2_hl, vl2_hh);
146 
147     const __m128i ven0_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 19);
148     const __m128i ven0_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 19);
149     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
150     const __m128 vs0_lo = _mm_castsi128_ps(_mm_add_epi32(vl0_lo, ven0_lo));
151     const __m128 vs0_hi = _mm_castsi128_ps(_mm_add_epi32(vl0_hi, ven0_hi));
152     const __m128i ven1_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 19);
153     const __m128i ven1_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 19);
154     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
155     const __m128 vs1_lo = _mm_castsi128_ps(_mm_add_epi32(vl1_lo, ven1_lo));
156     const __m128 vs1_hi = _mm_castsi128_ps(_mm_add_epi32(vl1_hi, ven1_hi));
157     const __m128i ven2_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 19);
158     const __m128i ven2_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 19);
159     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
160     const __m128 vs2_lo = _mm_castsi128_ps(_mm_add_epi32(vl2_lo, ven2_lo));
161     const __m128 vs2_hi = _mm_castsi128_ps(_mm_add_epi32(vl2_hi, ven2_hi));
162 
163     __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
164     __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
165     __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
166 
167     vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
168     __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
169     vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
170     __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
171     vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
172     __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1);
173 
174     __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc3, vt0), vc2);
175     __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc3, vt1), vc2);
176     __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc3, vt2), vc2);
177 
178     vp0 = _mm256_mul_ps(vp0, vt0);
179     vp1 = _mm256_mul_ps(vp1, vt1);
180     vp2 = _mm256_mul_ps(vp2, vt2);
181 
182     vt0 = _mm256_mul_ps(vt0, vs0);
183     vs0 = _mm256_sub_ps(vs0, vone);
184     vt1 = _mm256_mul_ps(vt1, vs1);
185     vs1 = _mm256_sub_ps(vs1, vone);
186     vt2 = _mm256_mul_ps(vt2, vs2);
187     vs2 = _mm256_sub_ps(vs2, vone);
188 
189     vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
190     vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
191     vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
192 
193     const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
194     vx0 = _mm256_mul_ps(vx0, vbeta);
195     const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
196     vx1 = _mm256_mul_ps(vx1, vbeta);
197     const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
198     vx2 = _mm256_mul_ps(vx2, vbeta);
199 
200     const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
201     const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
202     const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
203 
204     _mm256_storeu_ps(y, vy0);
205     _mm256_storeu_ps(y + 8, vy1);
206     _mm256_storeu_ps(y + 16, vy2);
207     y += 24;
208   }
209   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
210     __m256 vx = _mm256_loadu_ps(x);
211     x += 8;
212 
213     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
214 
215     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
216 
217     const __m256 vidx = _mm256_and_ps(vn, vindex_mask);
218 
219     const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
220     const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
221     #if XNN_ARCH_X86_64
222       const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
223       const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
224       const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
225       const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
226       __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_ll));
227       __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lh));
228       __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hl));
229       __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh));
230       vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_ll >> 32))), 1);
231       vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lh >> 32))), 1);
232       vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hl >> 32))), 1);
233       vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
234     #else
235       __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo)));
236       __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2)));
237       __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi)));
238       __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2)));
239       vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
240       vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
241       vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
242       vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
243     #endif
244     const __m128i ven_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 19);
245     const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19);
246 
247     const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
248     const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
249 
250     vn = _mm256_sub_ps(vn, vmagic_bias);
251     const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ven_lo));
252     const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ven_hi));
253 
254     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
255     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
256     __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
257 
258     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc3, vt), vc2);
259     vp = _mm256_mul_ps(vp, vt);
260 
261     vt = _mm256_mul_ps(vt, vs);
262     vs = _mm256_sub_ps(vs, vone);
263     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
264 
265     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
266     vx = _mm256_mul_ps(vx, vbeta);
267     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
268 
269     _mm256_storeu_ps(y, vy);
270     y += 8;
271   }
272   if XNN_UNLIKELY(n != 0) {
273     assert(n >= 1 * sizeof(float));
274     assert(n <= 7 * sizeof(float));
275     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
276 
277     __m256 vx = _mm256_maskload_ps(x, vmask);
278 
279     const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
280 
281     __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
282 
283     const __m256 vidx = _mm256_and_ps(vn, vindex_mask);
284 
285     const __m128i vidx_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vidx)), 2);
286     const __m128i vidx_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vidx, 1)), 2);
287     #if XNN_ARCH_X86_64
288       const uint64_t vidx_ll = (uint64_t) _mm_cvtsi128_si64(vidx_lo);
289       const uint64_t vidx_lh = (uint64_t) _mm_extract_epi64(vidx_lo, 1);
290       const uint64_t vidx_hl = (uint64_t) _mm_cvtsi128_si64(vidx_hi);
291       const uint64_t vidx_hh = (uint64_t) _mm_extract_epi64(vidx_hi, 1);
292       __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_ll));
293       __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_lh));
294       __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hl));
295       __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) vidx_hh));
296       vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_ll >> 32))), 1);
297       vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_lh >> 32))), 1);
298       vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hl >> 32))), 1);
299       vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) (vidx_hh >> 32))), 1);
300     #else
301       __m128i vl_ll = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_lo)));
302       __m128i vl_lh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2)));
303       __m128i vl_hl = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_cvtsi128_si32(vidx_hi)));
304       __m128i vl_hh = _mm_loadu_si32((const void*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2)));
305       vl_ll = _mm_insert_epi32(vl_ll, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))), 1);
306       vl_lh = _mm_insert_epi32(vl_lh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))), 1);
307       vl_hl = _mm_insert_epi32(vl_hl, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))), 1);
308       vl_hh = _mm_insert_epi32(vl_hh, *((const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))), 1);
309     #endif
310     const __m128i ven_lo = _mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 19);
311     const __m128i ven_hi = _mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 19);
312 
313     const __m128i vl_lo = _mm_unpacklo_epi64(vl_ll, vl_lh);
314     const __m128i vl_hi = _mm_unpacklo_epi64(vl_hl, vl_hh);
315 
316     vn = _mm256_sub_ps(vn, vmagic_bias);
317     const __m128 vs_lo = _mm_castsi128_ps(_mm_add_epi32(vl_lo, ven_lo));
318     const __m128 vs_hi = _mm_castsi128_ps(_mm_add_epi32(vl_hi, ven_hi));
319 
320     __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
321     vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
322     __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
323 
324     __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc3, vt), vc2);
325     vp = _mm256_mul_ps(vp, vt);
326 
327     vt = _mm256_mul_ps(vt, vs);
328     vs = _mm256_sub_ps(vs, vone);
329     vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
330 
331     const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
332     vx = _mm256_mul_ps(vx, vbeta);
333     const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
334 
335     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
336     __m128 vy_lo = _mm256_castps256_ps128(vy);
337     if (n & (4 * sizeof(float))) {
338       _mm_storeu_ps(y, vy_lo);
339       vy_lo = _mm256_extractf128_ps(vy, 1);
340       y += 4;
341     }
342     if (n & (2 * sizeof(float))) {
343       _mm_storel_pi((__m64*) y, vy_lo);
344       vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
345       y += 2;
346     }
347     if (n & (1 * sizeof(float))) {
348       _mm_store_ss(y, vy_lo);
349     }
350   }
351 }
352