• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddstoreexpminusmax/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 
12 #include <immintrin.h>
13 
14 #include <xnnpack/raddstoreexpminusmax.h>
15 
16 
17 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
18 
xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6(size_t elements,const float * input,float * output,float * sum,float max)19 void xnn_f32_raddstoreexpminusmax_ukernel__avx2_p5_x96_acc6(
20     size_t elements,
21     const float* input,
22     float* output,
23     float* sum,
24     float max)
25 {
26   assert(elements % sizeof(float) == 0);
27 
28   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
29   // The smallest x for which expf(x) is normalized.
30   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
31   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
32   const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
33   const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
34 
35   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
36   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
37   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
38   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
39   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
40 
41   const __m256 vi_max = _mm256_set1_ps(max);
42 
43   __m256 vacc0 = _mm256_setzero_ps();
44   __m256 vacc1 = _mm256_setzero_ps();
45   __m256 vacc2 = _mm256_setzero_ps();
46   __m256 vacc3 = _mm256_setzero_ps();
47   __m256 vacc4 = _mm256_setzero_ps();
48   __m256 vacc5 = _mm256_setzero_ps();
49   for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
50     // Load 96 (12x8) inputs at a time.
51     const __m256 vi0 = _mm256_loadu_ps(input);
52     const __m256 vi1 = _mm256_loadu_ps(input + 8);
53     const __m256 vi2 = _mm256_loadu_ps(input + 16);
54     const __m256 vi3 = _mm256_loadu_ps(input + 24);
55     const __m256 vi4 = _mm256_loadu_ps(input + 32);
56     const __m256 vi5 = _mm256_loadu_ps(input + 40);
57     const __m256 vi6 = _mm256_loadu_ps(input + 48);
58     const __m256 vi7 = _mm256_loadu_ps(input + 56);
59     const __m256 vi8 = _mm256_loadu_ps(input + 64);
60     const __m256 vi9 = _mm256_loadu_ps(input + 72);
61     const __m256 vi10 = _mm256_loadu_ps(input + 80);
62     const __m256 vi11 = _mm256_loadu_ps(input + 88);
63     input += 96;
64 
65     // Subtract maximum input x := i - i_max. This implies x <= 0.
66     const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
67     const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
68     const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
69     const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
70     const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
71     const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
72     const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
73     const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
74     const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
75     const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
76     const __m256 vx10 = _mm256_sub_ps(vi10, vi_max);
77     const __m256 vx11 = _mm256_sub_ps(vi11, vi_max);
78 
79     // Compute reduced argument elements := round(x / log(2)).
80     __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
81     __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
82     __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
83     __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
84     __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
85     __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
86     __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
87     __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
88     __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
89     __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
90     __m256 vn10 = _mm256_fmadd_ps(vx10, vlog2e, vmagic_bias);
91     __m256 vn11 = _mm256_fmadd_ps(vx11, vlog2e, vmagic_bias);
92 
93     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
94     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
95     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
96     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
97     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
98     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
99     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
100     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
101     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
102     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
103     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
104     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
105     const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn10), 23));
106     const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn11), 23));
107 
108     // Subtract the large number back to get final elements := round(x / log(2)).
109     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
110     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
111     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
112     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
113     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
114     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
115     vn6 = _mm256_sub_ps(vn6, vmagic_bias);
116     vn7 = _mm256_sub_ps(vn7, vmagic_bias);
117     vn8 = _mm256_sub_ps(vn8, vmagic_bias);
118     vn9 = _mm256_sub_ps(vn9, vmagic_bias);
119     vn10 = _mm256_sub_ps(vn10, vmagic_bias);
120     vn11 = _mm256_sub_ps(vn11, vmagic_bias);
121 
122     // Compute reduced argument t := x - elements * log(2).
123     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
124     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
125     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
126     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
127     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
128     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
129     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
130     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
131     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
132     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
133     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
134     __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
135     __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
136 
137     vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
138     vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
139     vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
140     vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
141     vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
142     vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
143     vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
144     vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
145     vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
146     vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
147     vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
148     vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
149 
150     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
151     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
152     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
153     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
154     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
155     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
156     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
157     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
158     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
159     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
160     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
161     __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
162     __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
163 
164     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
165     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
166     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
167     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
168     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
169     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
170     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
171     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
172     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
173     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
174     vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
175     vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
176 
177     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
178     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
179     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
180     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
181     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
182     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
183     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
184     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
185     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
186     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
187     vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
188     vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
189 
190     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
191     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
192     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
193     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
194     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
195     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
196     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
197     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
198     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
199     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
200     vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
201     vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
202 
203     // Reconstruct the final f value:
204     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
205     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
206     //     = s + (t * s) * p
207     vt0 = _mm256_mul_ps(vt0, vs0);
208     vt1 = _mm256_mul_ps(vt1, vs1);
209     vt2 = _mm256_mul_ps(vt2, vs2);
210     vt3 = _mm256_mul_ps(vt3, vs3);
211     vt4 = _mm256_mul_ps(vt4, vs4);
212     vt5 = _mm256_mul_ps(vt5, vs5);
213     vt6 = _mm256_mul_ps(vt6, vs6);
214     vt7 = _mm256_mul_ps(vt7, vs7);
215     vt8 = _mm256_mul_ps(vt8, vs8);
216     vt9 = _mm256_mul_ps(vt9, vs9);
217     vt10 = _mm256_mul_ps(vt10, vs10);
218     vt11 = _mm256_mul_ps(vt11, vs11);
219 
220     __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
221     __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
222     __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
223     __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
224     __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
225     __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
226     __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
227     __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
228     __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
229     __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
230     __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10);
231     __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11);
232 
233     // For inputs below zero cutoff, replace output with +0.0f.
234     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
235     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
236     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
237     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
238     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
239     vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
240     vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
241     vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
242     vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
243     vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
244     vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
245     vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10);
246     vf11 = _mm256_andnot_ps(_mm256_cmp_ps(vx11, vdenorm_cutoff, _CMP_LT_OS), vf11);
247 
248     // Store 96 (12x8) outputs at a time.
249     _mm256_storeu_ps(output, vf0);
250     _mm256_storeu_ps(output + 8, vf1);
251     _mm256_storeu_ps(output + 16, vf2);
252     _mm256_storeu_ps(output + 24, vf3);
253     _mm256_storeu_ps(output + 32, vf4);
254     _mm256_storeu_ps(output + 40, vf5);
255     _mm256_storeu_ps(output + 48, vf6);
256     _mm256_storeu_ps(output + 56, vf7);
257     _mm256_storeu_ps(output + 64, vf8);
258     _mm256_storeu_ps(output + 72, vf9);
259     _mm256_storeu_ps(output + 80, vf10);
260     _mm256_storeu_ps(output + 88, vf11);
261     output += 96;
262 
263     // Accumulate computed exponents.
264     vacc0 = _mm256_add_ps(vacc0, vf0);
265     vacc1 = _mm256_add_ps(vacc1, vf1);
266     vacc2 = _mm256_add_ps(vacc2, vf2);
267     vacc3 = _mm256_add_ps(vacc3, vf3);
268     vacc4 = _mm256_add_ps(vacc4, vf4);
269     vacc5 = _mm256_add_ps(vacc5, vf5);
270     vacc0 = _mm256_add_ps(vacc0, vf6);
271     vacc1 = _mm256_add_ps(vacc1, vf7);
272     vacc2 = _mm256_add_ps(vacc2, vf8);
273     vacc3 = _mm256_add_ps(vacc3, vf9);
274     vacc4 = _mm256_add_ps(vacc4, vf10);
275     vacc5 = _mm256_add_ps(vacc5, vf11);
276   }
277   // Add up all accumulators to vacc0
278   vacc0 = _mm256_add_ps(vacc0, vacc1);
279   vacc2 = _mm256_add_ps(vacc2, vacc3);
280   vacc4 = _mm256_add_ps(vacc4, vacc5);
281   vacc0 = _mm256_add_ps(vacc0, vacc2);
282   vacc0 = _mm256_add_ps(vacc0, vacc4);
283 
284   __m256 vacc = vacc0;
285   for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
286     // Load 8 inputs at a time.
287     const __m256 vi = _mm256_loadu_ps(input);
288     input += 8;
289 
290     // Subtract maximum input x := i - i_max. This implies x <= 0.
291     const __m256 vx = _mm256_sub_ps(vi, vi_max);
292 
293     // Compute reduced argument elements := round(x / log(2)).
294     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
295 
296     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
297     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
298     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
299 
300     // Subtract the large number back to get final elements := round(x / log(2)).
301     vn = _mm256_sub_ps(vn, vmagic_bias);
302 
303     // Compute reduced argument t := x - elements * log(2).
304     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
305     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
306     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
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 
314     // Reconstruct the final f value:
315     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
316     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
317     //     = s + (t * s) * p
318     vt = _mm256_mul_ps(vt, vs);
319     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
320 
321     // For inputs below zero cutoff, replace output with +0.0f.
322     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
323     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
324 
325     // Store 8 outputs at a time.
326     _mm256_storeu_ps(output, vf);
327     output += 8;
328 
329     // Accumulate computed exponents.
330     vacc = _mm256_add_ps(vacc, vf);
331   }
332   if (elements != 0) {
333     assert(elements >= 1 * sizeof(float));
334     assert(elements <= 7 * sizeof(float));
335     const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
336 
337     // Load up to 7 inputs at a time.
338     const __m256 vi = _mm256_maskload_ps(input, vmask);
339 
340     // Subtract maximum input x := i - i_max. This implies x <= 0.
341     const __m256 vx = _mm256_sub_ps(vi, vi_max);
342 
343     // Compute reduced argument elements := round(x / log(2)).
344     __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
345 
346     // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
347     // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
348     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
349 
350     // Subtract the large number back to get final elements := round(x / log(2)).
351     vn = _mm256_sub_ps(vn, vmagic_bias);
352 
353     // Compute reduced argument t := x - elements * log(2).
354     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
355     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
356     vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
357 
358     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
359     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
360     vp = _mm256_fmadd_ps(vp, vt, vc3);
361     vp = _mm256_fmadd_ps(vp, vt, vc2);
362     vp = _mm256_fmadd_ps(vp, vt, vc1);
363 
364     // Reconstruct the final f value:
365     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
366     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
367     //     = s + (t * s) * p
368     vt = _mm256_mul_ps(vt, vs);
369     __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
370 
371     // For inputs below zero cutoff, replace output with +0.0f.
372     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
373     vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
374 
375     // Store up to 7 outputs at a time.
376     _mm256_maskstore_ps(output, vmask, vf);
377 
378     // Accumulate computed exponents. And addend with mask to leave unmasked 32-bit lanes unchanged.
379     vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
380   }
381   // Reduce 8 elements in the SIMD register
382   __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
383   vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
384   vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
385   _mm_store_ss(sum, vacc_lo);
386   _mm256_zeroupper();
387 }
388