xref: /aosp_15_r20/external/XNNPACK/src/f32-raddextexp/gen/avx2-p5-x80-acc5.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_x80_acc5(size_t elements,const float * x,float * sum)20 void xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5(
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 vacce0 = vminus_inf;
50   __m256 vacce1 = vminus_inf;
51   __m256 vacce2 = vminus_inf;
52   __m256 vacce3 = vminus_inf;
53   __m256 vacce4 = vminus_inf;
54   for (; elements >= 80 * sizeof(float); elements -= 80 * sizeof(float)) {
55     // Load 80 (10x8) inputs at a time.
56     const __m256 vx0 = _mm256_loadu_ps(x);
57     const __m256 vx1 = _mm256_loadu_ps(x + 8);
58     const __m256 vx2 = _mm256_loadu_ps(x + 16);
59     const __m256 vx3 = _mm256_loadu_ps(x + 24);
60     const __m256 vx4 = _mm256_loadu_ps(x + 32);
61     const __m256 vx5 = _mm256_loadu_ps(x + 40);
62     const __m256 vx6 = _mm256_loadu_ps(x + 48);
63     const __m256 vx7 = _mm256_loadu_ps(x + 56);
64     const __m256 vx8 = _mm256_loadu_ps(x + 64);
65     const __m256 vx9 = _mm256_loadu_ps(x + 72);
66     x += 80;
67 
68     // Compute reduced argument elements := round(x / log(2)).
69     const __m256 vn0 = _mm256_round_ps(_mm256_mul_ps(vx0, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
70     const __m256 vn1 = _mm256_round_ps(_mm256_mul_ps(vx1, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
71     const __m256 vn2 = _mm256_round_ps(_mm256_mul_ps(vx2, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
72     const __m256 vn3 = _mm256_round_ps(_mm256_mul_ps(vx3, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
73     const __m256 vn4 = _mm256_round_ps(_mm256_mul_ps(vx4, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
74     const __m256 vn5 = _mm256_round_ps(_mm256_mul_ps(vx5, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
75     const __m256 vn6 = _mm256_round_ps(_mm256_mul_ps(vx6, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
76     const __m256 vn7 = _mm256_round_ps(_mm256_mul_ps(vx7, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
77     const __m256 vn8 = _mm256_round_ps(_mm256_mul_ps(vx8, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
78     const __m256 vn9 = _mm256_round_ps(_mm256_mul_ps(vx9, 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 
93     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
94     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
95     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
96     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
97     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
98     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
99     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
100     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
101     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
102     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
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 
116     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
117     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
118     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
119     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
120     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
121     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
122     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
123     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
124     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
125     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
126 
127     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
128     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
129     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
130     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
131     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
132     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
133     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
134     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
135     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
136     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
137 
138     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
139     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
140     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
141     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
142     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
143     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
144     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
145     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
146     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
147     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
148 
149     vp0 = _mm256_fmadd_ps(vp0, vt0, vc0);
150     vp1 = _mm256_fmadd_ps(vp1, vt1, vc0);
151     vp2 = _mm256_fmadd_ps(vp2, vt2, vc0);
152     vp3 = _mm256_fmadd_ps(vp3, vt3, vc0);
153     vp4 = _mm256_fmadd_ps(vp4, vt4, vc0);
154     vp5 = _mm256_fmadd_ps(vp5, vt5, vc0);
155     vp6 = _mm256_fmadd_ps(vp6, vt6, vc0);
156     vp7 = _mm256_fmadd_ps(vp7, vt7, vc0);
157     vp8 = _mm256_fmadd_ps(vp8, vt8, vc0);
158     vp9 = _mm256_fmadd_ps(vp9, vt9, vc0);
159 
160     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation where
161     //  - vnX is "exponent"
162     //  - vpX is "mantissa"
163     //
164     // exp2(ae) * av + exp2(be) * bv =
165     //   = exp2(max(ae, be)) * exp2(ae - max(ae, be)) * av + exp2(max(ae, be)) * exp2(be - max(ae, be)) * bv
166     //   = exp2(max_e) * (exp2(ae - max_e) * av + exp2(be - max_e) * bv)
167     //   = exp2(max_e) * (exp2(delta_ae) * av + exp2(delta_be) * bv)
168     //
169     // For computational efficiency we may add several "extended" floating-point numbers at a time.
170     __m256 vmax_e0 = _mm256_max_ps(vacce0, vn0);
171     __m256 vmax_e1 = _mm256_max_ps(vacce1, vn1);
172     __m256 vmax_e2 = _mm256_max_ps(vacce2, vn2);
173     __m256 vmax_e3 = _mm256_max_ps(vacce3, vn3);
174     __m256 vmax_e4 = _mm256_max_ps(vacce4, vn4);
175     vmax_e0 = _mm256_max_ps(vmax_e0, vn5);
176     vmax_e1 = _mm256_max_ps(vmax_e1, vn6);
177     vmax_e2 = _mm256_max_ps(vmax_e2, vn7);
178     vmax_e3 = _mm256_max_ps(vmax_e3, vn8);
179     vmax_e4 = _mm256_max_ps(vmax_e4, vn9);
180 
181     // For computational efficiency, replace exp2(delta_e) with 0.0f when delta_e <= -127.0.
182     // This replacement is done in two steps:
183     // 1. Clamp minimum delta_e at -127.0.
184     // 2. Map delta_e to scale factor 0.0 when delta_e == -127.0
185     const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_e0), vmin_exponent);
186     const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_e1), vmin_exponent);
187     const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_e2), vmin_exponent);
188     const __m256 vdelta_acce3 = _mm256_max_ps(_mm256_sub_ps(vacce3, vmax_e3), vmin_exponent);
189     const __m256 vdelta_acce4 = _mm256_max_ps(_mm256_sub_ps(vacce4, vmax_e4), vmin_exponent);
190     const __m256 vdelta_e0 = _mm256_max_ps(_mm256_sub_ps(vn0, vmax_e0), vmin_exponent);
191     const __m256 vdelta_e1 = _mm256_max_ps(_mm256_sub_ps(vn1, vmax_e1), vmin_exponent);
192     const __m256 vdelta_e2 = _mm256_max_ps(_mm256_sub_ps(vn2, vmax_e2), vmin_exponent);
193     const __m256 vdelta_e3 = _mm256_max_ps(_mm256_sub_ps(vn3, vmax_e3), vmin_exponent);
194     const __m256 vdelta_e4 = _mm256_max_ps(_mm256_sub_ps(vn4, vmax_e4), vmin_exponent);
195     const __m256 vdelta_e5 = _mm256_max_ps(_mm256_sub_ps(vn5, vmax_e0), vmin_exponent);
196     const __m256 vdelta_e6 = _mm256_max_ps(_mm256_sub_ps(vn6, vmax_e1), vmin_exponent);
197     const __m256 vdelta_e7 = _mm256_max_ps(_mm256_sub_ps(vn7, vmax_e2), vmin_exponent);
198     const __m256 vdelta_e8 = _mm256_max_ps(_mm256_sub_ps(vn8, vmax_e3), vmin_exponent);
199     const __m256 vdelta_e9 = _mm256_max_ps(_mm256_sub_ps(vn9, vmax_e4), vmin_exponent);
200 
201     // Convert delta-exponents into scale factors:
202     // - s = exp2(delta_e) when delta_e > -127.0
203     // - s = 0.0 when delta_e <= -127.0
204     //
205     // Note: delta-exponents can not exceed 0.0, thus scale factors can not exceed 1.0.
206     const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
207     const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
208     const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
209     const __m256 vaccs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce3, vmagic_bias)), 23));
210     const __m256 vaccs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce4, vmagic_bias)), 23));
211     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e0, vmagic_bias)), 23));
212     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e1, vmagic_bias)), 23));
213     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e2, vmagic_bias)), 23));
214     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e3, vmagic_bias)), 23));
215     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e4, vmagic_bias)), 23));
216     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e5, vmagic_bias)), 23));
217     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e6, vmagic_bias)), 23));
218     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e7, vmagic_bias)), 23));
219     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e8, vmagic_bias)), 23));
220     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e9, vmagic_bias)), 23));
221 
222     // Update accumulated "mantissa" and "exponent" values
223     vaccv0 = _mm256_mul_ps(vaccv0, vaccs0);
224     vaccv1 = _mm256_mul_ps(vaccv1, vaccs1);
225     vaccv2 = _mm256_mul_ps(vaccv2, vaccs2);
226     vaccv3 = _mm256_mul_ps(vaccv3, vaccs3);
227     vaccv4 = _mm256_mul_ps(vaccv4, vaccs4);
228     vaccv0 = _mm256_fmadd_ps(vp0, vs0, vaccv0);
229     vaccv1 = _mm256_fmadd_ps(vp1, vs1, vaccv1);
230     vaccv2 = _mm256_fmadd_ps(vp2, vs2, vaccv2);
231     vaccv3 = _mm256_fmadd_ps(vp3, vs3, vaccv3);
232     vaccv4 = _mm256_fmadd_ps(vp4, vs4, vaccv4);
233     vaccv0 = _mm256_fmadd_ps(vp5, vs5, vaccv0);
234     vaccv1 = _mm256_fmadd_ps(vp6, vs6, vaccv1);
235     vaccv2 = _mm256_fmadd_ps(vp7, vs7, vaccv2);
236     vaccv3 = _mm256_fmadd_ps(vp8, vs8, vaccv3);
237     vaccv4 = _mm256_fmadd_ps(vp9, vs9, vaccv4);
238 
239     vacce0 = vmax_e0;
240     vacce1 = vmax_e1;
241     vacce2 = vmax_e2;
242     vacce3 = vmax_e3;
243     vacce4 = vmax_e4;
244   }
245 
246   // Reduce partial sums of "extended" floating-point numbers into a single "extended" SIMD vector of sums.
247   const __m256 vmax_acce01 = _mm256_max_ps(vacce0, vacce1);
248   const __m256 vmax_acce23 = _mm256_max_ps(vacce2, vacce3);
249   const __m256 vmax_acce4 = vacce4;
250   const __m256 vmax_acce0123 = _mm256_max_ps(vmax_acce01, vmax_acce23);
251   const __m256 vmax_acce01234 = _mm256_max_ps(vmax_acce0123, vmax_acce4);
252 
253   const __m256 vdelta_acce0 = _mm256_max_ps(_mm256_sub_ps(vacce0, vmax_acce01234), vmin_exponent);
254   const __m256 vdelta_acce1 = _mm256_max_ps(_mm256_sub_ps(vacce1, vmax_acce01234), vmin_exponent);
255   const __m256 vdelta_acce2 = _mm256_max_ps(_mm256_sub_ps(vacce2, vmax_acce01234), vmin_exponent);
256   const __m256 vdelta_acce3 = _mm256_max_ps(_mm256_sub_ps(vacce3, vmax_acce01234), vmin_exponent);
257   const __m256 vdelta_acce4 = _mm256_max_ps(_mm256_sub_ps(vacce4, vmax_acce01234), vmin_exponent);
258 
259   const __m256 vaccs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce0, vmagic_bias)), 23));
260   const __m256 vaccs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce1, vmagic_bias)), 23));
261   const __m256 vaccs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce2, vmagic_bias)), 23));
262   const __m256 vaccs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce3, vmagic_bias)), 23));
263   const __m256 vaccs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce4, vmagic_bias)), 23));
264 
265   __m256 vaccv = _mm256_mul_ps(vaccv0, vaccs0);
266   vaccv = _mm256_fmadd_ps(vaccv1, vaccs1, vaccv);
267   vaccv = _mm256_fmadd_ps(vaccv2, vaccs2, vaccv);
268   vaccv = _mm256_fmadd_ps(vaccv3, vaccs3, vaccv);
269   vaccv = _mm256_fmadd_ps(vaccv4, vaccs4, vaccv);
270   __m256 vacce = vmax_acce01234;
271 
272   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
273     // Load 8 inputs at a time.
274     const __m256 vx = _mm256_loadu_ps(x);
275     x += 8;
276 
277     // Compute reduced argument elements := round(x / log(2)).
278     const __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
279 
280     // Compute reduced argument t := x - elements * log(2).
281     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
282     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
283     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
284 
285     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
286     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
287     vp = _mm256_fmadd_ps(vp, vt, vc3);
288     vp = _mm256_fmadd_ps(vp, vt, vc2);
289     vp = _mm256_fmadd_ps(vp, vt, vc1);
290     vp = _mm256_fmadd_ps(vp, vt, vc0);
291 
292     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
293     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
294 
295     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
296     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
297     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
298 
299     // Convert exponents into scale factors.
300     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
301     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
302 
303     // Update accumulated "mantissa" and "exponent" values.
304     vaccv = _mm256_mul_ps(vaccv, vaccs);
305     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
306 
307     vacce = vmax_e;
308   }
309   if XNN_UNLIKELY(elements != 0) {
310     assert(elements >= 1 * sizeof(float));
311     assert(elements <= 7 * sizeof(float));
312     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
313 
314     // Load up to 7 inputs at a time.
315     const __m256 vx = _mm256_maskload_ps(x, vmask);
316 
317     // Compute reduced argument elements := round(x / log(2)).
318     __m256 vn = _mm256_round_ps(_mm256_mul_ps(vx, vlog2e), _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
319 
320     // Compute reduced argument t := x - elements * log(2).
321     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
322     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
323     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
324 
325     // Correct reduced argument elements for masked out elements.
326     vn = _mm256_blendv_ps(vacce, vn, _mm256_castsi256_ps(vmask));
327 
328     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
329     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
330     vp = _mm256_fmadd_ps(vp, vt, vc3);
331     vp = _mm256_fmadd_ps(vp, vt, vc2);
332     vp = _mm256_fmadd_ps(vp, vt, vc1);
333     vp = _mm256_fmadd_ps(vp, vt, vc0);
334     vp = _mm256_and_ps(vp, _mm256_castsi256_ps(vmask));
335 
336     // Accumulate "extended" floating-point numbers in ("mantissa", "exponent") representation.
337     const __m256 vmax_e = _mm256_max_ps(vacce, vn);
338 
339     // For computational efficiency, clamp minimum exp2(delta_e) at -127.0. It will be mapped to 0.0 scale factor later.
340     const __m256 vdelta_e = _mm256_max_ps(_mm256_sub_ps(vn, vmax_e), vmin_exponent);
341     const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_e), vmin_exponent);
342 
343     // Convert exponents into scale factors.
344     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_e, vmagic_bias)), 23));
345     const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
346 
347     // Update accumulated "mantissa" and "exponent" values.
348     vaccv = _mm256_mul_ps(vaccv, vaccs);
349     vaccv = _mm256_fmadd_ps(vp, vs, vaccv);
350 
351     vacce = vmax_e;
352   }
353 
354   // Reduce partial sums of "extended" floating-point numbers into a single "extended" floating-point sum.
355   __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
356   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, 2)));
357   vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, 1)));
358   const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
359   const __m256 vaccs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(_mm256_add_ps(vdelta_acce, vmagic_bias)), 23));
360 
361   vaccv = _mm256_mul_ps(vaccv, vaccs);
362   __m128 vaccv_sum = _mm_add_ps(_mm256_castps256_ps128(vaccv), _mm256_extractf128_ps(vaccv, 1));
363   vaccv_sum = _mm_add_ps(vaccv_sum, _mm_movehl_ps(vaccv_sum, vaccv_sum));
364   vaccv_sum = _mm_add_ss(vaccv_sum, _mm_movehdup_ps(vaccv_sum));
365 
366   _mm_store_ss(&sum[0], vaccv_sum);
367   _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
368 
369   _mm256_zeroupper();
370 }
371