• 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_acc3(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3(
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 vacce0 = vminus_inf;
48   __m256 vacce1 = vminus_inf;
49   __m256 vacce2 = vminus_inf;
50   for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
51     // Load 96 (12x8) inputs at a time.
52     const __m256 vx0 = _mm256_loadu_ps(x);
53     const __m256 vx1 = _mm256_loadu_ps(x + 8);
54     const __m256 vx2 = _mm256_loadu_ps(x + 16);
55     const __m256 vx3 = _mm256_loadu_ps(x + 24);
56     const __m256 vx4 = _mm256_loadu_ps(x + 32);
57     const __m256 vx5 = _mm256_loadu_ps(x + 40);
58     const __m256 vx6 = _mm256_loadu_ps(x + 48);
59     const __m256 vx7 = _mm256_loadu_ps(x + 56);
60     const __m256 vx8 = _mm256_loadu_ps(x + 64);
61     const __m256 vx9 = _mm256_loadu_ps(x + 72);
62     const __m256 vx10 = _mm256_loadu_ps(x + 80);
63     const __m256 vx11 = _mm256_loadu_ps(x + 88);
64     x += 96;
65 
66     // Compute reduced argument elements := round(x / log(2)).
67     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
68     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
69     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
70     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
71     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
72     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
73     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
74     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
75     const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
76     const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
77     const __m256 vn10 = _mm256_round_ps(_mm256_mul_ps(vx10, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
78     const __m256 vn11 = _mm256_round_ps(_mm256_mul_ps(vx11, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
79 
80     // Compute reduced argument t := x - elements * log(2).
81     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
82     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
83     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
84     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
85     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
86     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
87     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
88     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
89     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
90     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
91     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
92     __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
93     __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
94 
95     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
96     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
97     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
98     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
99     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
100     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
101     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
102     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
103     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
104     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
105     vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
106     vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
107 
108     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
109     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
110     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
111     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
112     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
113     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
114     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
115     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
116     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
117     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
118     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
119     __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
120     __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
121 
122     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
123     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
124     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
125     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
126     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
127     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
128     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
129     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
130     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
131     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
132     vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
133     vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
134 
135     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
136     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
137     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
138     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
139     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
140     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
141     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
142     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
143     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
144     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
145     vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
146     vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
147 
148     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
149     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
150     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
151     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
152     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
153     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
154     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
155     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
156     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
157     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
158     vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
159     vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
160 
161     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
162     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
163     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
164     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
165     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
166     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
167     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
168     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
169     vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
170     vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
171     vp10 = _mm256_fmadd_ps(vp10, vt10, vc0);
172     vp11 = _mm256_fmadd_ps(vp11, vt11, vc0);
173 
174     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
175     //  - vnX is "exponent"
176     //  - vpX is "mantissa"
177     //
178     // exp2(ae) * av + exp2(be) * bv =
179     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
180     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
181     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
182     //
183     // For computational efficiency we may add several "extended" floating-point numbers at a time.
184     __m256 vmax_e0 = _mm256_max_ps(vacce0, vn0);
185     __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1);
186     __m256 vmax_e2 = _mm256_max_ps(vacce2, vn2);
187     vmax_e0 = _mm256_max_ps(vmax_e0, vn3);
188     vmax_e1 = _mm256_max_ps(vmax_e1, vn4);
189     vmax_e2 = _mm256_max_ps(vmax_e2, vn5);
190     vmax_e0 = _mm256_max_ps(vmax_e0, vn6);
191     vmax_e1 = _mm256_max_ps(vmax_e1, vn7);
192     vmax_e2 = _mm256_max_ps(vmax_e2, vn8);
193     vmax_e0 = _mm256_max_ps(vmax_e0, vn9);
194     vmax_e1 = _mm256_max_ps(vmax_e1, vn10);
195     vmax_e2 = _mm256_max_ps(vmax_e2, vn11);
196 
197     // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
198     // This replacement is done in two steps:
199     // 1. Clamp minimum delta_e at -127.0.
200     // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
201     const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent);
202     const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent);
203     const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_e2), vmin_exponent);
204     const __m256 vdelta_e0 = _mm256_max_ps(_mm256_sub_ps(vn0, vmax_e0), vmin_exponent);
205     const __m256 vdelta_e1 = _mm256_max_ps(_mm256_sub_ps(vn1, vmax_e1), vmin_exponent);
206     const __m256 vdelta_e2 = _mm256_max_ps(_mm256_sub_ps(vn2, vmax_e2), vmin_exponent);
207     const __m256 vdelta_e3 = _mm256_max_ps(_mm256_sub_ps(vn3, vmax_e0), vmin_exponent);
208     const __m256 vdelta_e4 = _mm256_max_ps(_mm256_sub_ps(vn4, vmax_e1), vmin_exponent);
209     const __m256 vdelta_e5 = _mm256_max_ps(_mm256_sub_ps(vn5, vmax_e2), vmin_exponent);
210     const __m256 vdelta_e6 = _mm256_max_ps(_mm256_sub_ps(vn6, vmax_e0), vmin_exponent);
211     const __m256 vdelta_e7 = _mm256_max_ps(_mm256_sub_ps(vn7, vmax_e1), vmin_exponent);
212     const __m256 vdelta_e8 = _mm256_max_ps(_mm256_sub_ps(vn8, vmax_e2), vmin_exponent);
213     const __m256 vdelta_e9 = _mm256_max_ps(_mm256_sub_ps(vn9, vmax_e0), vmin_exponent);
214     const __m256 vdelta_e10 = _mm256_max_ps(_mm256_sub_ps(vn10, vmax_e1), vmin_exponent);
215     const __m256 vdelta_e11 = _mm256_max_ps(_mm256_sub_ps(vn11, vmax_e2), vmin_exponent);
216 
217     // Convert delta-exponents into scale factors:
218     // - s = exp2(delta_e) when delta_e > -127.0
219     // - s = 0.0 when delta_e <= -127.0
220     //
221     // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
222     const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
223     const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
224     const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
225     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e0, vmagic_bias)), 23));
226     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e1, vmagic_bias)), 23));
227     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e2, vmagic_bias)), 23));
228     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e3, vmagic_bias)), 23));
229     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e4, vmagic_bias)), 23));
230     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e5, vmagic_bias)), 23));
231     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e6, vmagic_bias)), 23));
232     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e7, vmagic_bias)), 23));
233     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e8, vmagic_bias)), 23));
234     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e9, vmagic_bias)), 23));
235     const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e10, vmagic_bias)), 23));
236     const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e11, vmagic_bias)), 23));
237 
238     // Update accumulated "mantissa" and "exponent" values
239     vaccv0 = _mm256_mul_ps(vaccv0, vaccs0);
240     vaccv1 = _mm256_mul_ps(vaccv1, vaccs1);
241     vaccv2 = _mm256_mul_ps(vaccv2, vaccs2);
242     vaccv0 = _mm256_fmadd_ps(vp0, vs0, vaccv0);
243     vaccv1 = _mm256_fmadd_ps(vp1, vs1, vaccv1);
244     vaccv2 = _mm256_fmadd_ps(vp2, vs2, vaccv2);
245     vaccv0 = _mm256_fmadd_ps(vp3, vs3, vaccv0);
246     vaccv1 = _mm256_fmadd_ps(vp4, vs4, vaccv1);
247     vaccv2 = _mm256_fmadd_ps(vp5, vs5, vaccv2);
248     vaccv0 = _mm256_fmadd_ps(vp6, vs6, vaccv0);
249     vaccv1 = _mm256_fmadd_ps(vp7, vs7, vaccv1);
250     vaccv2 = _mm256_fmadd_ps(vp8, vs8, vaccv2);
251     vaccv0 = _mm256_fmadd_ps(vp9, vs9, vaccv0);
252     vaccv1 = _mm256_fmadd_ps(vp10, vs10, vaccv1);
253     vaccv2 = _mm256_fmadd_ps(vp11, vs11, vaccv2);
254 
255     vacce0 = vmax_e0;
256     vacce1 = vmax_e1;
257     vacce2 = vmax_e2;
258   }
259 
260   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
261   const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1);
262   const __m256 vmax_acce2 = vacce2;
263   const __m256 vmax_acce012 = _mm256_max_ps(vmax_acce01, vmax_acce2);
264 
265   const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce012), vmin_exponent);
266   const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce012), vmin_exponent);
267   const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_acce012), vmin_exponent);
268 
269   const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
270   const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
271   const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
272 
273   __m256 vaccv = _mm256_mul_ps(vaccv0, vaccs0);
274   vaccv = _mm256_fmadd_ps(vaccv1, vaccs1, vaccv);
275   vaccv = _mm256_fmadd_ps(vaccv2, vaccs2, vaccv);
276   __m256 vacce = vmax_acce012;
277 
278   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
279     // Load 8 inputs at a time.
280     const __m256 vx = _mm256_loadu_ps(x);
281     x += 8;
282 
283     // Compute reduced argument elements := round(x / log(2)).
284     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
285 
286     // Compute reduced argument t := x - elements * log(2).
287     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
288     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
289     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
290 
291     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
292     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
293     vp = _mm256_fmadd_ps(vp, vt, vc3);
294     vp = _mm256_fmadd_ps(vp, vt, vc2);
295     vp = _mm256_fmadd_ps(vp, vt, vc1);
296     vp = _mm256_fmadd_ps(vp, vt, vc0);
297 
298     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
299     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
300 
301     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
302     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
303     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
304 
305     // Convert exponents into scale factors.
306     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
307     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
308 
309     // Update accumulated "mantissa" and "exponent" values.
310     vaccv = _mm256_mul_ps(vaccv, vaccs);
311     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
312 
313     vacce = vmax_e;
314   }
315   if XNN_UNLIKELY(elements != 0) {
316     assert(elements >= 1 * sizeof(float));
317     assert(elements <= 7 * sizeof(float));
318     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
319 
320     // Load up to 7 inputs at a time.
321     const __m256 vx = _mm256_maskload_ps(x, vmask);
322 
323     // Compute reduced argument elements := round(x / log(2)).
324     __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
325 
326     // Compute reduced argument t := x - elements * log(2).
327     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
328     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
329     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
330 
331     // Correct reduced argument elements for masked out elements.
332     vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
333 
334     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
335     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
336     vp = _mm256_fmadd_ps(vp, vt, vc3);
337     vp = _mm256_fmadd_ps(vp, vt, vc2);
338     vp = _mm256_fmadd_ps(vp, vt, vc1);
339     vp = _mm256_fmadd_ps(vp, vt, vc0);
340     vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
341 
342     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
343     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
344 
345     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
346     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
347     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
348 
349     // Convert exponents into scale factors.
350     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
351     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
352 
353     // Update accumulated "mantissa" and "exponent" values.
354     vaccv = _mm256_mul_ps(vaccv, vaccs);
355     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
356 
357     vacce = vmax_e;
358   }
359 
360   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
361   __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
362   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
363   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
364   const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
365   const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
366 
367   vaccv = _mm256_mul_ps(vaccv, vaccs);
368   __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
369   vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
370   vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
371 
372   _mm_store_ss(&sum[0], vaccv_sum);
373   _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
374 
375   _mm256_zeroupper();
376 }
377