1 // Auto-generated file. Do not edit!
2 // Template: src/f32-raddexpminusmax/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/raddexpminusmax.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_raddexpminusmax_ukernel__avx2_p5_x96_acc6(size_t elements,const float * input,float * sum,float max)19 void xnn_f32_raddexpminusmax_ukernel__avx2_p5_x96_acc6(
20 size_t elements,
21 const float* input,
22 float* sum,
23 float max)
24 {
25 assert(elements % sizeof(float) == 0);
26
27 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
28 // The smallest x for which expf(x) is normalized.
29 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f);
30 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
31 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f);
33
34 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
35 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
36 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
37 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
38 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
39
40 const __m256 vi_max = _mm256_set1_ps(max);
41
42 __m256 vacc0 = _mm256_setzero_ps();
43 __m256 vacc1 = _mm256_setzero_ps();
44 __m256 vacc2 = _mm256_setzero_ps();
45 __m256 vacc3 = _mm256_setzero_ps();
46 __m256 vacc4 = _mm256_setzero_ps();
47 __m256 vacc5 = _mm256_setzero_ps();
48 for (; elements >= 96 * sizeof(float); elements -= 96 * sizeof(float)) {
49 // Load 96 (12x8) inputs at a time.
50 const __m256 vi0 = _mm256_loadu_ps(input);
51 const __m256 vi1 = _mm256_loadu_ps(input + 8);
52 const __m256 vi2 = _mm256_loadu_ps(input + 16);
53 const __m256 vi3 = _mm256_loadu_ps(input + 24);
54 const __m256 vi4 = _mm256_loadu_ps(input + 32);
55 const __m256 vi5 = _mm256_loadu_ps(input + 40);
56 const __m256 vi6 = _mm256_loadu_ps(input + 48);
57 const __m256 vi7 = _mm256_loadu_ps(input + 56);
58 const __m256 vi8 = _mm256_loadu_ps(input + 64);
59 const __m256 vi9 = _mm256_loadu_ps(input + 72);
60 const __m256 vi10 = _mm256_loadu_ps(input + 80);
61 const __m256 vi11 = _mm256_loadu_ps(input + 88);
62 input += 96;
63
64 // Subtract maximum input x := i - i_max. This implies x <= 0.
65 const __m256 vx0 = _mm256_sub_ps(vi0, vi_max);
66 const __m256 vx1 = _mm256_sub_ps(vi1, vi_max);
67 const __m256 vx2 = _mm256_sub_ps(vi2, vi_max);
68 const __m256 vx3 = _mm256_sub_ps(vi3, vi_max);
69 const __m256 vx4 = _mm256_sub_ps(vi4, vi_max);
70 const __m256 vx5 = _mm256_sub_ps(vi5, vi_max);
71 const __m256 vx6 = _mm256_sub_ps(vi6, vi_max);
72 const __m256 vx7 = _mm256_sub_ps(vi7, vi_max);
73 const __m256 vx8 = _mm256_sub_ps(vi8, vi_max);
74 const __m256 vx9 = _mm256_sub_ps(vi9, vi_max);
75 const __m256 vx10 = _mm256_sub_ps(vi10, vi_max);
76 const __m256 vx11 = _mm256_sub_ps(vi11, vi_max);
77
78 // Compute reduced argument elements := round(x / log(2)).
79 __m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias);
80 __m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias);
81 __m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias);
82 __m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias);
83 __m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias);
84 __m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias);
85 __m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias);
86 __m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias);
87 __m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias);
88 __m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias);
89 __m256 vn10 = _mm256_fmadd_ps(vx10, vlog2e, vmagic_bias);
90 __m256 vn11 = _mm256_fmadd_ps(vx11, vlog2e, vmagic_bias);
91
92 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
93 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
94 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
95 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
96 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
97 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
98 const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
99 const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
100 const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
101 const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
102 const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
103 const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
104 const __m256 vs10 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn10), 23));
105 const __m256 vs11 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn11), 23));
106
107 // Subtract the large number back to get final elements := round(x / log(2)).
108 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
109 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
110 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
111 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
112 vn4 = _mm256_sub_ps(vn4, vmagic_bias);
113 vn5 = _mm256_sub_ps(vn5, vmagic_bias);
114 vn6 = _mm256_sub_ps(vn6, vmagic_bias);
115 vn7 = _mm256_sub_ps(vn7, vmagic_bias);
116 vn8 = _mm256_sub_ps(vn8, vmagic_bias);
117 vn9 = _mm256_sub_ps(vn9, vmagic_bias);
118 vn10 = _mm256_sub_ps(vn10, vmagic_bias);
119 vn11 = _mm256_sub_ps(vn11, vmagic_bias);
120
121 // Compute reduced argument t := x - elements * log(2).
122 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
123 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0);
124 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1);
125 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2);
126 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3);
127 __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4);
128 __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5);
129 __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6);
130 __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7);
131 __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8);
132 __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9);
133 __m256 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_hi, vx10);
134 __m256 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_hi, vx11);
135
136 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0);
137 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1);
138 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2);
139 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3);
140 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4);
141 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5);
142 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6);
143 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7);
144 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8);
145 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9);
146 vt10 = _mm256_fmadd_ps(vn10, vminus_ln2_lo, vt10);
147 vt11 = _mm256_fmadd_ps(vn11, vminus_ln2_lo, vt11);
148
149 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
150 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
151 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
152 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
153 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
154 __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
155 __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
156 __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
157 __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
158 __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
159 __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
160 __m256 vp10 = _mm256_fmadd_ps(vc5, vt10, vc4);
161 __m256 vp11 = _mm256_fmadd_ps(vc5, vt11, vc4);
162
163 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
164 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
165 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
166 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
167 vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
168 vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
169 vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
170 vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
171 vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
172 vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
173 vp10 = _mm256_fmadd_ps(vp10, vt10, vc3);
174 vp11 = _mm256_fmadd_ps(vp11, vt11, vc3);
175
176 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
177 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
178 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
179 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
180 vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
181 vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
182 vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
183 vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
184 vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
185 vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
186 vp10 = _mm256_fmadd_ps(vp10, vt10, vc2);
187 vp11 = _mm256_fmadd_ps(vp11, vt11, vc2);
188
189 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
190 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
191 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
192 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
193 vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
194 vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
195 vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
196 vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
197 vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
198 vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
199 vp10 = _mm256_fmadd_ps(vp10, vt10, vc1);
200 vp11 = _mm256_fmadd_ps(vp11, vt11, vc1);
201
202 // Reconstruct the final f value:
203 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
204 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
205 // = s + (t * s) * p
206 vt0 = _mm256_mul_ps(vt0, vs0);
207 vt1 = _mm256_mul_ps(vt1, vs1);
208 vt2 = _mm256_mul_ps(vt2, vs2);
209 vt3 = _mm256_mul_ps(vt3, vs3);
210 vt4 = _mm256_mul_ps(vt4, vs4);
211 vt5 = _mm256_mul_ps(vt5, vs5);
212 vt6 = _mm256_mul_ps(vt6, vs6);
213 vt7 = _mm256_mul_ps(vt7, vs7);
214 vt8 = _mm256_mul_ps(vt8, vs8);
215 vt9 = _mm256_mul_ps(vt9, vs9);
216 vt10 = _mm256_mul_ps(vt10, vs10);
217 vt11 = _mm256_mul_ps(vt11, vs11);
218
219 __m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0);
220 __m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1);
221 __m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2);
222 __m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3);
223 __m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4);
224 __m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5);
225 __m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6);
226 __m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7);
227 __m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8);
228 __m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9);
229 __m256 vf10 = _mm256_fmadd_ps(vt10, vp10, vs10);
230 __m256 vf11 = _mm256_fmadd_ps(vt11, vp11, vs11);
231
232 // For inputs below zero cutoff, replace output with +0.0f.
233 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
234 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0);
235 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1);
236 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2);
237 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3);
238 vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4);
239 vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5);
240 vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6);
241 vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7);
242 vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8);
243 vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9);
244 vf10 = _mm256_andnot_ps(_mm256_cmp_ps(vx10, vdenorm_cutoff, _CMP_LT_OS), vf10);
245 vf11 = _mm256_andnot_ps(_mm256_cmp_ps(vx11, vdenorm_cutoff, _CMP_LT_OS), vf11);
246
247 // Accumulate computed exponents.
248 vacc0 = _mm256_add_ps(vacc0, vf0);
249 vacc1 = _mm256_add_ps(vacc1, vf1);
250 vacc2 = _mm256_add_ps(vacc2, vf2);
251 vacc3 = _mm256_add_ps(vacc3, vf3);
252 vacc4 = _mm256_add_ps(vacc4, vf4);
253 vacc5 = _mm256_add_ps(vacc5, vf5);
254 vacc0 = _mm256_add_ps(vacc0, vf6);
255 vacc1 = _mm256_add_ps(vacc1, vf7);
256 vacc2 = _mm256_add_ps(vacc2, vf8);
257 vacc3 = _mm256_add_ps(vacc3, vf9);
258 vacc4 = _mm256_add_ps(vacc4, vf10);
259 vacc5 = _mm256_add_ps(vacc5, vf11);
260 }
261 // Add up all accumulators to vacc0
262 vacc0 = _mm256_add_ps(vacc0, vacc1);
263 vacc2 = _mm256_add_ps(vacc2, vacc3);
264 vacc4 = _mm256_add_ps(vacc4, vacc5);
265 vacc0 = _mm256_add_ps(vacc0, vacc2);
266 vacc0 = _mm256_add_ps(vacc0, vacc4);
267
268 __m256 vacc = vacc0;
269 for (; elements >= 8 * sizeof(float); elements -= 8 * sizeof(float)) {
270 // Load 8 inputs at a time.
271 const __m256 vi = _mm256_loadu_ps(input);
272 input += 8;
273
274 // Subtract maximum input x := i - i_max. This implies x <= 0.
275 const __m256 vx = _mm256_sub_ps(vi, vi_max);
276
277 // Compute reduced argument elements := round(x / log(2)).
278 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
279
280 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
281 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
282 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
283
284 // Subtract the large number back to get final elements := round(x / log(2)).
285 vn = _mm256_sub_ps(vn, vmagic_bias);
286
287 // Compute reduced argument t := x - elements * log(2).
288 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
289 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
290 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
291
292 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
293 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
294 vp = _mm256_fmadd_ps(vp, vt, vc3);
295 vp = _mm256_fmadd_ps(vp, vt, vc2);
296 vp = _mm256_fmadd_ps(vp, vt, vc1);
297
298 // Reconstruct the final f value:
299 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
300 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
301 // = s + (t * s) * p
302 vt = _mm256_mul_ps(vt, vs);
303 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
304
305 // For inputs below zero cutoff, replace output with +0.0f.
306 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
307 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
308
309 // Accumulate computed exponents.
310 vacc = _mm256_add_ps(vacc, vf);
311 }
312 if (elements != 0) {
313 assert(elements >= 1 * sizeof(float));
314 assert(elements <= 7 * sizeof(float));
315 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - elements));
316
317 // Load up to 7 inputs at a time.
318 const __m256 vi = _mm256_maskload_ps(input, vmask);
319
320 // Subtract maximum input x := i - i_max. This implies x <= 0.
321 const __m256 vx = _mm256_sub_ps(vi, vi_max);
322
323 // Compute reduced argument elements := round(x / log(2)).
324 __m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias);
325
326 // Create a floating-point number s (scale) such that s == 2**elements for inputs which don't cause underflow, i.e.
327 // -87.33642 <= x <= 0.0, and -126 <= elements <= 0 accordingly.
328 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
329
330 // Subtract the large number back to get final elements := round(x / log(2)).
331 vn = _mm256_sub_ps(vn, vmagic_bias);
332
333 // Compute reduced argument t := x - elements * log(2).
334 // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
335 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx);
336 vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt);
337
338 // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
339 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
340 vp = _mm256_fmadd_ps(vp, vt, vc3);
341 vp = _mm256_fmadd_ps(vp, vt, vc2);
342 vp = _mm256_fmadd_ps(vp, vt, vc1);
343
344 // Reconstruct the final f value:
345 // f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
346 // = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
347 // = s + (t * s) * p
348 vt = _mm256_mul_ps(vt, vs);
349 __m256 vf = _mm256_fmadd_ps(vt, vp, vs);
350
351 // For inputs below zero cutoff, replace output with +0.0f.
352 // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
353 vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf);
354
355 // Accumulate computed exponents. And addend with mask to leave unmasked 32-bit lanes unchanged.
356 vacc = _mm256_add_ps(vacc, _mm256_and_ps(vf, _mm256_castsi256_ps(vmask)));
357 }
358 // Reduce 8 elements in the SIMD register
359 __m128 vacc_lo = _mm_add_ps(_mm256_castps256_ps128(vacc), _mm256_extractf128_ps(vacc, 1));
360 vacc_lo = _mm_add_ps(vacc_lo, _mm_movehl_ps(vacc_lo, vacc_lo));
361 vacc_lo = _mm_add_ss(vacc_lo, _mm_movehdup_ps(vacc_lo));
362 _mm_store_ss(sum, vacc_lo);
363 _mm256_zeroupper();
364 }
365