• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddextexp/avx2-p5.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2019 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 #include <math.h>
12 
13 #include <immintrin.h>
14 
15 #include <xnnpack/raddextexp.h>
16 
17 
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19 
xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6(
21     size_t elements,
22     const float* x,
23     float* sum)
24 {
25   assert(elements % sizeof(float) == 0);
26 
27   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
28   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
29   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
30 
31   // The smallest elements such that 2**elements is considered non-negligible.
32   // For smaller elements, 2**elements is replaced with zero.
33   const __m256 vmin_exponent = _mm256_set1_ps(-127.0f);
34   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
35   const __m256 vminus_inf = _mm256_set1_ps(-INFINITY);
36 
37   const __m256 vc0 = _mm256_set1_ps(1.0f);
38   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
39   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
40   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
41   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
42   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
43 
44   __m256 vaccv0 = _mm256_setzero_ps();
45   __m256 vaccv1 = _mm256_setzero_ps();
46   __m256 vaccv2 = _mm256_setzero_ps();
47   __m256 vaccv3 = _mm256_setzero_ps();
48   __m256 vaccv4 = _mm256_setzero_ps();
49   __m256 vaccv5 = _mm256_setzero_ps();
50   __m256 vacce0 = vminus_inf;
51   __m256 vacce1 = vminus_inf;
52   __m256 vacce2 = vminus_inf;
53   __m256 vacce3 = vminus_inf;
54   __m256 vacce4 = vminus_inf;
55   __m256 vacce5 = vminus_inf;
56   for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
57     // Load 96 (12x8) inputs at a time.
58     const __m256 vx0 = _mm256_loadu_ps(x);
59     const __m256 vx1 = _mm256_loadu_ps(x + 8);
60     const __m256 vx2 = _mm256_loadu_ps(x + 16);
61     const __m256 vx3 = _mm256_loadu_ps(x + 24);
62     const __m256 vx4 = _mm256_loadu_ps(x + 32);
63     const __m256 vx5 = _mm256_loadu_ps(x + 40);
64     const __m256 vx6 = _mm256_loadu_ps(x + 48);
65     const __m256 vx7 = _mm256_loadu_ps(x + 56);
66     const __m256 vx8 = _mm256_loadu_ps(x + 64);
67     const __m256 vx9 = _mm256_loadu_ps(x + 72);
68     const __m256 vx10 = _mm256_loadu_ps(x + 80);
69     const __m256 vx11 = _mm256_loadu_ps(x + 88);
70     x += 96;
71 
72     // Compute reduced argument elements := round(x / log(2)).
73     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
74     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
75     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
76     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
77     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
78     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
79     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
80     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
81     const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
82     const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
83     const __m256 vn10 = _mm256_round_ps(_mm256_mul_ps(vx10, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
84     const __m256 vn11 = _mm256_round_ps(_mm256_mul_ps(vx11, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
85 
86     // Compute reduced argument t := x - elements * log(2).
87     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
88     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
89     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
90     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
91     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
92     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
93     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
94     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
95     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
96     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
97     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
98     __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
99     __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
100 
101     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
102     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
103     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
104     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
105     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
106     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
107     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
108     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
109     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
110     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
111     vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
112     vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
113 
114     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
115     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
116     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
117     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
118     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
119     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
120     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
121     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
122     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
123     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
124     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
125     __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
126     __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
127 
128     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
129     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
130     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
131     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
132     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
133     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
134     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
135     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
136     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
137     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
138     vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
139     vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
140 
141     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
142     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
143     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
144     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
145     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
146     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
147     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
148     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
149     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
150     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
151     vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
152     vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
153 
154     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
155     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
156     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
157     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
158     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
159     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
160     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
161     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
162     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
163     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
164     vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
165     vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
166 
167     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
168     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
169     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
170     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
171     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
172     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
173     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
174     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
175     vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
176     vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
177     vp10 = _mm256_fmadd_ps(vp10, vt10, vc0);
178     vp11 = _mm256_fmadd_ps(vp11, vt11, vc0);
179 
180     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
181     //  - vnX is "exponent"
182     //  - vpX is "mantissa"
183     //
184     // exp2(ae) * av + exp2(be) * bv =
185     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
186     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
187     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
188     //
189     // For computational efficiency we may add several "extended" floating-point numbers at a time.
190     __m256 vmax_e0 = _mm256_max_ps(vacce0, vn0);
191     __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1);
192     __m256 vmax_e2 = _mm256_max_ps(vacce2, vn2);
193     __m256 vmax_e3 = _mm256_max_ps(vacce3, vn3);
194     __m256 vmax_e4 = _mm256_max_ps(vacce4, vn4);
195     __m256 vmax_e5 = _mm256_max_ps(vacce5, vn5);
196     vmax_e0 = _mm256_max_ps(vmax_e0, vn6);
197     vmax_e1 = _mm256_max_ps(vmax_e1, vn7);
198     vmax_e2 = _mm256_max_ps(vmax_e2, vn8);
199     vmax_e3 = _mm256_max_ps(vmax_e3, vn9);
200     vmax_e4 = _mm256_max_ps(vmax_e4, vn10);
201     vmax_e5 = _mm256_max_ps(vmax_e5, vn11);
202 
203     // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
204     // This replacement is done in two steps:
205     // 1. Clamp minimum delta_e at -127.0.
206     // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
207     const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent);
208     const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent);
209     const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_e2), vmin_exponent);
210     const __m256 vdelta_acce3 = _mm256_max_ps(_mm256_sub_ps(vacce3, vmax_e3), vmin_exponent);
211     const __m256 vdelta_acce4 = _mm256_max_ps(_mm256_sub_ps(vacce4, vmax_e4), vmin_exponent);
212     const __m256 vdelta_acce5 = _mm256_max_ps(_mm256_sub_ps(vacce5, vmax_e5), vmin_exponent);
213     const __m256 vdelta_e0 = _mm256_max_ps(_mm256_sub_ps(vn0, vmax_e0), vmin_exponent);
214     const __m256 vdelta_e1 = _mm256_max_ps(_mm256_sub_ps(vn1, vmax_e1), vmin_exponent);
215     const __m256 vdelta_e2 = _mm256_max_ps(_mm256_sub_ps(vn2, vmax_e2), vmin_exponent);
216     const __m256 vdelta_e3 = _mm256_max_ps(_mm256_sub_ps(vn3, vmax_e3), vmin_exponent);
217     const __m256 vdelta_e4 = _mm256_max_ps(_mm256_sub_ps(vn4, vmax_e4), vmin_exponent);
218     const __m256 vdelta_e5 = _mm256_max_ps(_mm256_sub_ps(vn5, vmax_e5), vmin_exponent);
219     const __m256 vdelta_e6 = _mm256_max_ps(_mm256_sub_ps(vn6, vmax_e0), vmin_exponent);
220     const __m256 vdelta_e7 = _mm256_max_ps(_mm256_sub_ps(vn7, vmax_e1), vmin_exponent);
221     const __m256 vdelta_e8 = _mm256_max_ps(_mm256_sub_ps(vn8, vmax_e2), vmin_exponent);
222     const __m256 vdelta_e9 = _mm256_max_ps(_mm256_sub_ps(vn9, vmax_e3), vmin_exponent);
223     const __m256 vdelta_e10 = _mm256_max_ps(_mm256_sub_ps(vn10, vmax_e4), vmin_exponent);
224     const __m256 vdelta_e11 = _mm256_max_ps(_mm256_sub_ps(vn11, vmax_e5), vmin_exponent);
225 
226     // Convert delta-exponents into scale factors:
227     // - s = exp2(delta_e) when delta_e > -127.0
228     // - s = 0.0 when delta_e <= -127.0
229     //
230     // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
231     const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
232     const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
233     const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
234     const __m256 vaccs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce3, vmagic_bias)), 23));
235     const __m256 vaccs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce4, vmagic_bias)), 23));
236     const __m256 vaccs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce5, vmagic_bias)), 23));
237     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e0, vmagic_bias)), 23));
238     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e1, vmagic_bias)), 23));
239     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e2, vmagic_bias)), 23));
240     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e3, vmagic_bias)), 23));
241     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e4, vmagic_bias)), 23));
242     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e5, vmagic_bias)), 23));
243     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e6, vmagic_bias)), 23));
244     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e7, vmagic_bias)), 23));
245     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e8, vmagic_bias)), 23));
246     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e9, vmagic_bias)), 23));
247     const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e10, vmagic_bias)), 23));
248     const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e11, vmagic_bias)), 23));
249 
250     // Update accumulated "mantissa" and "exponent" values
251     vaccv0 = _mm256_mul_ps(vaccv0, vaccs0);
252     vaccv1 = _mm256_mul_ps(vaccv1, vaccs1);
253     vaccv2 = _mm256_mul_ps(vaccv2, vaccs2);
254     vaccv3 = _mm256_mul_ps(vaccv3, vaccs3);
255     vaccv4 = _mm256_mul_ps(vaccv4, vaccs4);
256     vaccv5 = _mm256_mul_ps(vaccv5, vaccs5);
257     vaccv0 = _mm256_fmadd_ps(vp0, vs0, vaccv0);
258     vaccv1 = _mm256_fmadd_ps(vp1, vs1, vaccv1);
259     vaccv2 = _mm256_fmadd_ps(vp2, vs2, vaccv2);
260     vaccv3 = _mm256_fmadd_ps(vp3, vs3, vaccv3);
261     vaccv4 = _mm256_fmadd_ps(vp4, vs4, vaccv4);
262     vaccv5 = _mm256_fmadd_ps(vp5, vs5, vaccv5);
263     vaccv0 = _mm256_fmadd_ps(vp6, vs6, vaccv0);
264     vaccv1 = _mm256_fmadd_ps(vp7, vs7, vaccv1);
265     vaccv2 = _mm256_fmadd_ps(vp8, vs8, vaccv2);
266     vaccv3 = _mm256_fmadd_ps(vp9, vs9, vaccv3);
267     vaccv4 = _mm256_fmadd_ps(vp10, vs10, vaccv4);
268     vaccv5 = _mm256_fmadd_ps(vp11, vs11, vaccv5);
269 
270     vacce0 = vmax_e0;
271     vacce1 = vmax_e1;
272     vacce2 = vmax_e2;
273     vacce3 = vmax_e3;
274     vacce4 = vmax_e4;
275     vacce5 = vmax_e5;
276   }
277 
278   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
279   const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1);
280   const __m256 vmax_acce23 = _mm256_max_ps(vacce2, vacce3);
281   const __m256 vmax_acce45 = _mm256_max_ps(vacce4, vacce5);
282   const __m256 vmax_acce0123 = _mm256_max_ps(vmax_acce01, vmax_acce23);
283   const __m256 vmax_acce012345 = _mm256_max_ps(vmax_acce0123, vmax_acce45);
284 
285   const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce012345), vmin_exponent);
286   const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce012345), vmin_exponent);
287   const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_acce012345), vmin_exponent);
288   const __m256 vdelta_acce3 = _mm256_max_ps(_mm256_sub_ps(vacce3, vmax_acce012345), vmin_exponent);
289   const __m256 vdelta_acce4 = _mm256_max_ps(_mm256_sub_ps(vacce4, vmax_acce012345), vmin_exponent);
290   const __m256 vdelta_acce5 = _mm256_max_ps(_mm256_sub_ps(vacce5, vmax_acce012345), vmin_exponent);
291 
292   const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
293   const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
294   const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
295   const __m256 vaccs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce3, vmagic_bias)), 23));
296   const __m256 vaccs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce4, vmagic_bias)), 23));
297   const __m256 vaccs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce5, vmagic_bias)), 23));
298 
299   __m256 vaccv = _mm256_mul_ps(vaccv0, vaccs0);
300   vaccv = _mm256_fmadd_ps(vaccv1, vaccs1, vaccv);
301   vaccv = _mm256_fmadd_ps(vaccv2, vaccs2, vaccv);
302   vaccv = _mm256_fmadd_ps(vaccv3, vaccs3, vaccv);
303   vaccv = _mm256_fmadd_ps(vaccv4, vaccs4, vaccv);
304   vaccv = _mm256_fmadd_ps(vaccv5, vaccs5, vaccv);
305   __m256 vacce = vmax_acce012345;
306 
307   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
308     // Load 8 inputs at a time.
309     const __m256 vx = _mm256_loadu_ps(x);
310     x += 8;
311 
312     // Compute reduced argument elements := round(x / log(2)).
313     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
314 
315     // Compute reduced argument t := x - elements * log(2).
316     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
317     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
318     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
319 
320     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
321     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
322     vp = _mm256_fmadd_ps(vp, vt, vc3);
323     vp = _mm256_fmadd_ps(vp, vt, vc2);
324     vp = _mm256_fmadd_ps(vp, vt, vc1);
325     vp = _mm256_fmadd_ps(vp, vt, vc0);
326 
327     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
328     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
329 
330     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
331     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
332     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
333 
334     // Convert exponents into scale factors.
335     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
336     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
337 
338     // Update accumulated "mantissa" and "exponent" values.
339     vaccv = _mm256_mul_ps(vaccv, vaccs);
340     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
341 
342     vacce = vmax_e;
343   }
344   if XNN_UNLIKELY(elements != 0) {
345     assert(elements >= 1 * sizeof(float));
346     assert(elements <= 7 * sizeof(float));
347     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
348 
349     // Load up to 7 inputs at a time.
350     const __m256 vx = _mm256_maskload_ps(x, vmask);
351 
352     // Compute reduced argument elements := round(x / log(2)).
353     __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
354 
355     // Compute reduced argument t := x - elements * log(2).
356     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
357     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
358     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
359 
360     // Correct reduced argument elements for masked out elements.
361     vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
362 
363     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
364     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
365     vp = _mm256_fmadd_ps(vp, vt, vc3);
366     vp = _mm256_fmadd_ps(vp, vt, vc2);
367     vp = _mm256_fmadd_ps(vp, vt, vc1);
368     vp = _mm256_fmadd_ps(vp, vt, vc0);
369     vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
370 
371     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
372     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
373 
374     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
375     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
376     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
377 
378     // Convert exponents into scale factors.
379     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
380     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
381 
382     // Update accumulated "mantissa" and "exponent" values.
383     vaccv = _mm256_mul_ps(vaccv, vaccs);
384     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
385 
386     vacce = vmax_e;
387   }
388 
389   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
390   __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
391   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
392   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
393   const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
394   const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
395 
396   vaccv = _mm256_mul_ps(vaccv, vaccs);
397   __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
398   vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
399   vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
400 
401   _mm_store_ss(&sum[0], vaccv_sum);
402   _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
403 
404   _mm256_zeroupper();
405 }
406