xref: /aosp_15_r20/external/XNNPACK/src/f32-raddextexp/gen/avx2-p5-x96.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx2_p5_x96(
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 vacce0 = vminus_inf;
46   for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
47     // Load 96 (12x8) inputs at a time.
48     const __m256 vx0 = _mm256_loadu_ps(x);
49     const __m256 vx1 = _mm256_loadu_ps(x + 8);
50     const __m256 vx2 = _mm256_loadu_ps(x + 16);
51     const __m256 vx3 = _mm256_loadu_ps(x + 24);
52     const __m256 vx4 = _mm256_loadu_ps(x + 32);
53     const __m256 vx5 = _mm256_loadu_ps(x + 40);
54     const __m256 vx6 = _mm256_loadu_ps(x + 48);
55     const __m256 vx7 = _mm256_loadu_ps(x + 56);
56     const __m256 vx8 = _mm256_loadu_ps(x + 64);
57     const __m256 vx9 = _mm256_loadu_ps(x + 72);
58     const __m256 vx10 = _mm256_loadu_ps(x + 80);
59     const __m256 vx11 = _mm256_loadu_ps(x + 88);
60     x += 96;
61 
62     // Compute reduced argument elements := round(x / log(2)).
63     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
64     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
65     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
66     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
67     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
69     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
70     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
71     const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
72     const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
73     const __m256 vn10 = _mm256_round_ps(_mm256_mul_ps(vx10, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
74     const __m256 vn11 = _mm256_round_ps(_mm256_mul_ps(vx11, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
75 
76     // Compute reduced argument t := x - elements * log(2).
77     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
78     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
79     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
80     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
81     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
82     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
83     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
84     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
85     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
86     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
87     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
88     __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
89     __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
90 
91     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
92     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
93     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
94     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
95     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
96     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
97     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
98     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
99     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
100     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
101     vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
102     vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
103 
104     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
105     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
106     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
107     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
108     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
109     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
110     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
111     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
112     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
113     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
114     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
115     __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
116     __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
117 
118     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
119     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
120     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
121     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
122     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
123     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
124     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
125     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
126     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
127     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
128     vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
129     vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
130 
131     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
132     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
133     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
134     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
135     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
136     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
137     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
138     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
139     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
140     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
141     vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
142     vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
143 
144     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
145     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
146     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
147     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
148     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
149     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
150     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
151     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
152     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
153     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
154     vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
155     vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
156 
157     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
158     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
159     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
160     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
161     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
162     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
163     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
164     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
165     vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
166     vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
167     vp10 = _mm256_fmadd_ps(vp10, vt10, vc0);
168     vp11 = _mm256_fmadd_ps(vp11, vt11, vc0);
169 
170     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
171     //  - vnX is "exponent"
172     //  - vpX is "mantissa"
173     //
174     // exp2(ae) * av + exp2(be) * bv =
175     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
176     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
177     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
178     //
179     // For computational efficiency we may add several "extended" floating-point numbers at a time.
180     __m256 vmax_e0 = _mm256_max_ps(vacce0, vn0);
181     vmax_e0 = _mm256_max_ps(vmax_e0, vn1);
182     vmax_e0 = _mm256_max_ps(vmax_e0, vn2);
183     vmax_e0 = _mm256_max_ps(vmax_e0, vn3);
184     vmax_e0 = _mm256_max_ps(vmax_e0, vn4);
185     vmax_e0 = _mm256_max_ps(vmax_e0, vn5);
186     vmax_e0 = _mm256_max_ps(vmax_e0, vn6);
187     vmax_e0 = _mm256_max_ps(vmax_e0, vn7);
188     vmax_e0 = _mm256_max_ps(vmax_e0, vn8);
189     vmax_e0 = _mm256_max_ps(vmax_e0, vn9);
190     vmax_e0 = _mm256_max_ps(vmax_e0, vn10);
191     vmax_e0 = _mm256_max_ps(vmax_e0, vn11);
192 
193     // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
194     // This replacement is done in two steps:
195     // 1. Clamp minimum delta_e at -127.0.
196     // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
197     const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent);
198     const __m256 vdelta_e0 = _mm256_max_ps(_mm256_sub_ps(vn0, vmax_e0), vmin_exponent);
199     const __m256 vdelta_e1 = _mm256_max_ps(_mm256_sub_ps(vn1, vmax_e0), vmin_exponent);
200     const __m256 vdelta_e2 = _mm256_max_ps(_mm256_sub_ps(vn2, vmax_e0), vmin_exponent);
201     const __m256 vdelta_e3 = _mm256_max_ps(_mm256_sub_ps(vn3, vmax_e0), vmin_exponent);
202     const __m256 vdelta_e4 = _mm256_max_ps(_mm256_sub_ps(vn4, vmax_e0), vmin_exponent);
203     const __m256 vdelta_e5 = _mm256_max_ps(_mm256_sub_ps(vn5, vmax_e0), vmin_exponent);
204     const __m256 vdelta_e6 = _mm256_max_ps(_mm256_sub_ps(vn6, vmax_e0), vmin_exponent);
205     const __m256 vdelta_e7 = _mm256_max_ps(_mm256_sub_ps(vn7, vmax_e0), vmin_exponent);
206     const __m256 vdelta_e8 = _mm256_max_ps(_mm256_sub_ps(vn8, vmax_e0), vmin_exponent);
207     const __m256 vdelta_e9 = _mm256_max_ps(_mm256_sub_ps(vn9, vmax_e0), vmin_exponent);
208     const __m256 vdelta_e10 = _mm256_max_ps(_mm256_sub_ps(vn10, vmax_e0), vmin_exponent);
209     const __m256 vdelta_e11 = _mm256_max_ps(_mm256_sub_ps(vn11, vmax_e0), vmin_exponent);
210 
211     // Convert delta-exponents into scale factors:
212     // - s = exp2(delta_e) when delta_e > -127.0
213     // - s = 0.0 when delta_e <= -127.0
214     //
215     // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
216     const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
217     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e0, vmagic_bias)), 23));
218     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e1, vmagic_bias)), 23));
219     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e2, vmagic_bias)), 23));
220     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e3, vmagic_bias)), 23));
221     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e4, vmagic_bias)), 23));
222     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e5, vmagic_bias)), 23));
223     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e6, vmagic_bias)), 23));
224     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e7, vmagic_bias)), 23));
225     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e8, vmagic_bias)), 23));
226     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e9, vmagic_bias)), 23));
227     const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e10, vmagic_bias)), 23));
228     const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e11, vmagic_bias)), 23));
229 
230     // Update accumulated "mantissa" and "exponent" values
231     vaccv0 = _mm256_mul_ps(vaccv0, vaccs0);
232     vaccv0 = _mm256_fmadd_ps(vp0, vs0, vaccv0);
233     vaccv0 = _mm256_fmadd_ps(vp1, vs1, vaccv0);
234     vaccv0 = _mm256_fmadd_ps(vp2, vs2, vaccv0);
235     vaccv0 = _mm256_fmadd_ps(vp3, vs3, vaccv0);
236     vaccv0 = _mm256_fmadd_ps(vp4, vs4, vaccv0);
237     vaccv0 = _mm256_fmadd_ps(vp5, vs5, vaccv0);
238     vaccv0 = _mm256_fmadd_ps(vp6, vs6, vaccv0);
239     vaccv0 = _mm256_fmadd_ps(vp7, vs7, vaccv0);
240     vaccv0 = _mm256_fmadd_ps(vp8, vs8, vaccv0);
241     vaccv0 = _mm256_fmadd_ps(vp9, vs9, vaccv0);
242     vaccv0 = _mm256_fmadd_ps(vp10, vs10, vaccv0);
243     vaccv0 = _mm256_fmadd_ps(vp11, vs11, vaccv0);
244 
245     vacce0 = vmax_e0;
246   }
247 
248   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
249   __m256 vaccv = vaccv0;
250   __m256 vacce = vacce0;
251 
252   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
253     // Load 8 inputs at a time.
254     const __m256 vx = _mm256_loadu_ps(x);
255     x += 8;
256 
257     // Compute reduced argument elements := round(x / log(2)).
258     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
259 
260     // Compute reduced argument t := x - elements * log(2).
261     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
262     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
263     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
264 
265     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
266     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
267     vp = _mm256_fmadd_ps(vp, vt, vc3);
268     vp = _mm256_fmadd_ps(vp, vt, vc2);
269     vp = _mm256_fmadd_ps(vp, vt, vc1);
270     vp = _mm256_fmadd_ps(vp, vt, vc0);
271 
272     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
273     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
274 
275     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
276     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
277     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
278 
279     // Convert exponents into scale factors.
280     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
281     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
282 
283     // Update accumulated "mantissa" and "exponent" values.
284     vaccv = _mm256_mul_ps(vaccv, vaccs);
285     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
286 
287     vacce = vmax_e;
288   }
289   if XNN_UNLIKELY(elements != 0) {
290     assert(elements >= 1 * sizeof(float));
291     assert(elements <= 7 * sizeof(float));
292     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
293 
294     // Load up to 7 inputs at a time.
295     const __m256 vx = _mm256_maskload_ps(x, vmask);
296 
297     // Compute reduced argument elements := round(x / log(2)).
298     __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
299 
300     // Compute reduced argument t := x - elements * log(2).
301     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
302     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
303     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
304 
305     // Correct reduced argument elements for masked out elements.
306     vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
307 
308     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
309     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
310     vp = _mm256_fmadd_ps(vp, vt, vc3);
311     vp = _mm256_fmadd_ps(vp, vt, vc2);
312     vp = _mm256_fmadd_ps(vp, vt, vc1);
313     vp = _mm256_fmadd_ps(vp, vt, vc0);
314     vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
315 
316     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
317     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
318 
319     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
320     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
321     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
322 
323     // Convert exponents into scale factors.
324     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
325     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
326 
327     // Update accumulated "mantissa" and "exponent" values.
328     vaccv = _mm256_mul_ps(vaccv, vaccs);
329     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
330 
331     vacce = vmax_e;
332   }
333 
334   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
335   __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
336   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
337   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
338   const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
339   const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
340 
341   vaccv = _mm256_mul_ps(vaccv, vaccs);
342   __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
343   vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
344   vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
345 
346   _mm_store_ss(&sum[0], vaccv_sum);
347   _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
348 
349   _mm256_zeroupper();
350 }
351