• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-sigmoid/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/common.h>
15 #include <xnnpack/vunary.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_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x80(
21     size_t n,
22     const float* x,
23     float* y,
24     const void* params)
25 {
26   assert(n % sizeof(float) == 0);
27 
28   const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
29   // The smallest x for which sigmoidf(x) is normalized.
30   // This number is also the smallest x for which expf(x) is normalized.
31   const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
32   const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
33   const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
34   const __m256 vone = _mm256_set1_ps(1.0f);
35   const __m256 vsign_mask = _mm256_set1_ps(-0.0f);
36 
37   const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
38   const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
39   const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
40   const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
41   const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
42 
43   for (; n >= 80 * sizeof(float); n -= 80 * sizeof(float)) {
44     const __m256 vx0 = _mm256_loadu_ps(x);
45     const __m256 vx1 = _mm256_loadu_ps(x + 8);
46     const __m256 vx2 = _mm256_loadu_ps(x + 16);
47     const __m256 vx3 = _mm256_loadu_ps(x + 24);
48     const __m256 vx4 = _mm256_loadu_ps(x + 32);
49     const __m256 vx5 = _mm256_loadu_ps(x + 40);
50     const __m256 vx6 = _mm256_loadu_ps(x + 48);
51     const __m256 vx7 = _mm256_loadu_ps(x + 56);
52     const __m256 vx8 = _mm256_loadu_ps(x + 64);
53     const __m256 vx9 = _mm256_loadu_ps(x + 72);
54     x += 80;
55 
56     // General structure of the algorithm:
57     //           / exp(x) / (1 + exp(x)) if x <= 0
58     //   f[x] :=
59     //           \ 1 - f[-x] if x >= 0
60     //
61     // First we compute f[z] := exp(z) / (1 + exp(z)) where z = -abs(x),
62     // then replace result with 1 - f[z] if x >= 0.
63     const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
64     const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
65     const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
66     const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
67     const __m256 vz4 = _mm256_or_ps(vx4, vsign_mask);
68     const __m256 vz5 = _mm256_or_ps(vx5, vsign_mask);
69     const __m256 vz6 = _mm256_or_ps(vx6, vsign_mask);
70     const __m256 vz7 = _mm256_or_ps(vx7, vsign_mask);
71     const __m256 vz8 = _mm256_or_ps(vx8, vsign_mask);
72     const __m256 vz9 = _mm256_or_ps(vx9, vsign_mask);
73 
74     // Compute reduced argument n := round(z / log(2)).
75     // We do it by adding a large number (magic bias) to the product z * (1/log(2)), which cause rounding of the result
76     // to an integer, then subtracing the large number back. The trick with adding large number is valid only within
77     // certain bounds (|x| <= 2**22), but thats ok, because inputs x outside of [-87.336544, 17.328678] (i.e. z outsize
78     // [0, 87.336544]) underflow or saturate sigmoidf(x) anyway. We fixup the result for such inputs at the very end of
79     // the algorithm.
80     __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
81     __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
82     __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
83     __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
84     __m256 vn4 = _mm256_fmadd_ps(vz4, vlog2e, vmagic_bias);
85     __m256 vn5 = _mm256_fmadd_ps(vz5, vlog2e, vmagic_bias);
86     __m256 vn6 = _mm256_fmadd_ps(vz6, vlog2e, vmagic_bias);
87     __m256 vn7 = _mm256_fmadd_ps(vz7, vlog2e, vmagic_bias);
88     __m256 vn8 = _mm256_fmadd_ps(vz8, vlog2e, vmagic_bias);
89     __m256 vn9 = _mm256_fmadd_ps(vz9, vlog2e, vmagic_bias);
90 
91     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
92     // -87.33642 <= z <= 0.0, and -126 <= n <= 0 accordingly.
93     const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
94     const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
95     const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
96     const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
97     const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23));
98     const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23));
99     const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23));
100     const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23));
101     const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23));
102     const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23));
103 
104     // Subtract the large number back to get final n := round(z / log(2)).
105     vn0 = _mm256_sub_ps(vn0, vmagic_bias);
106     vn1 = _mm256_sub_ps(vn1, vmagic_bias);
107     vn2 = _mm256_sub_ps(vn2, vmagic_bias);
108     vn3 = _mm256_sub_ps(vn3, vmagic_bias);
109     vn4 = _mm256_sub_ps(vn4, vmagic_bias);
110     vn5 = _mm256_sub_ps(vn5, vmagic_bias);
111     vn6 = _mm256_sub_ps(vn6, vmagic_bias);
112     vn7 = _mm256_sub_ps(vn7, vmagic_bias);
113     vn8 = _mm256_sub_ps(vn8, vmagic_bias);
114     vn9 = _mm256_sub_ps(vn9, vmagic_bias);
115 
116     // Compute reduced argument t := z - n * log(2).
117     __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
118     __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
119     __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
120     __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
121     __m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2, vz4);
122     __m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2, vz5);
123     __m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2, vz6);
124     __m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2, vz7);
125     __m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2, vz8);
126     __m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2, vz9);
127 
128     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
129     __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
130     __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
131     __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
132     __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
133     __m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4);
134     __m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4);
135     __m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4);
136     __m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4);
137     __m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4);
138     __m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4);
139 
140     vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
141     vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
142     vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
143     vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
144     vp4 = _mm256_fmadd_ps(vp4, vt4, vc3);
145     vp5 = _mm256_fmadd_ps(vp5, vt5, vc3);
146     vp6 = _mm256_fmadd_ps(vp6, vt6, vc3);
147     vp7 = _mm256_fmadd_ps(vp7, vt7, vc3);
148     vp8 = _mm256_fmadd_ps(vp8, vt8, vc3);
149     vp9 = _mm256_fmadd_ps(vp9, vt9, vc3);
150 
151     vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
152     vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
153     vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
154     vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
155     vp4 = _mm256_fmadd_ps(vp4, vt4, vc2);
156     vp5 = _mm256_fmadd_ps(vp5, vt5, vc2);
157     vp6 = _mm256_fmadd_ps(vp6, vt6, vc2);
158     vp7 = _mm256_fmadd_ps(vp7, vt7, vc2);
159     vp8 = _mm256_fmadd_ps(vp8, vt8, vc2);
160     vp9 = _mm256_fmadd_ps(vp9, vt9, vc2);
161 
162     vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
163     vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
164     vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
165     vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
166     vp4 = _mm256_fmadd_ps(vp4, vt4, vc1);
167     vp5 = _mm256_fmadd_ps(vp5, vt5, vc1);
168     vp6 = _mm256_fmadd_ps(vp6, vt6, vc1);
169     vp7 = _mm256_fmadd_ps(vp7, vt7, vc1);
170     vp8 = _mm256_fmadd_ps(vp8, vt8, vc1);
171     vp9 = _mm256_fmadd_ps(vp9, vt9, vc1);
172 
173     // Reconstruct the exp(z) value:
174     //   e = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
175     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
176     //     = s + (t * s) * p
177     vt0 = _mm256_mul_ps(vt0, vs0);
178     vt1 = _mm256_mul_ps(vt1, vs1);
179     vt2 = _mm256_mul_ps(vt2, vs2);
180     vt3 = _mm256_mul_ps(vt3, vs3);
181     vt4 = _mm256_mul_ps(vt4, vs4);
182     vt5 = _mm256_mul_ps(vt5, vs5);
183     vt6 = _mm256_mul_ps(vt6, vs6);
184     vt7 = _mm256_mul_ps(vt7, vs7);
185     vt8 = _mm256_mul_ps(vt8, vs8);
186     vt9 = _mm256_mul_ps(vt9, vs9);
187 
188     const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
189     const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
190     const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
191     const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
192     const __m256 ve4 = _mm256_fmadd_ps(vt4, vp4, vs4);
193     const __m256 ve5 = _mm256_fmadd_ps(vt5, vp5, vs5);
194     const __m256 ve6 = _mm256_fmadd_ps(vt6, vp6, vs6);
195     const __m256 ve7 = _mm256_fmadd_ps(vt7, vp7, vs7);
196     const __m256 ve8 = _mm256_fmadd_ps(vt8, vp8, vs8);
197     const __m256 ve9 = _mm256_fmadd_ps(vt9, vp9, vs9);
198 
199     // Denominator of the sigmoid fraction: 1.0 + exp(z)
200     const __m256 vd0 = _mm256_add_ps(ve0, vone);
201     const __m256 vd1 = _mm256_add_ps(ve1, vone);
202     const __m256 vd2 = _mm256_add_ps(ve2, vone);
203     const __m256 vd3 = _mm256_add_ps(ve3, vone);
204     const __m256 vd4 = _mm256_add_ps(ve4, vone);
205     const __m256 vd5 = _mm256_add_ps(ve5, vone);
206     const __m256 vd6 = _mm256_add_ps(ve6, vone);
207     const __m256 vd7 = _mm256_add_ps(ve7, vone);
208     const __m256 vd8 = _mm256_add_ps(ve8, vone);
209     const __m256 vd9 = _mm256_add_ps(ve9, vone);
210 
211     // Use Newton-Raphson method to compute reciprocal of denominator.
212     // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
213     // Thus the reciprocal of the denominator never overflows.
214     __m256 vr0 = _mm256_rcp_ps(vd0);
215     __m256 vr1 = _mm256_rcp_ps(vd1);
216     __m256 vr2 = _mm256_rcp_ps(vd2);
217     __m256 vr3 = _mm256_rcp_ps(vd3);
218     __m256 vr4 = _mm256_rcp_ps(vd4);
219     __m256 vr5 = _mm256_rcp_ps(vd5);
220     __m256 vr6 = _mm256_rcp_ps(vd6);
221     __m256 vr7 = _mm256_rcp_ps(vd7);
222     __m256 vr8 = _mm256_rcp_ps(vd8);
223     __m256 vr9 = _mm256_rcp_ps(vd9);
224 
225     vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
226     vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
227     vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
228     vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
229     vr4 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
230     vr5 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
231     vr6 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
232     vr7 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr7, vd7, vone), vr7, vr7);
233     vr8 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr8, vd8, vone), vr8, vr8);
234     vr9 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr9, vd9, vone), vr9, vr9);
235 
236     vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
237     vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
238     vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
239     vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
240     vr4 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr4, vd4, vone), vr4, vr4);
241     vr5 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr5, vd5, vone), vr5, vr5);
242     vr6 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr6, vd6, vone), vr6, vr6);
243     vr7 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr7, vd7, vone), vr7, vr7);
244     vr8 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr8, vd8, vone), vr8, vr8);
245     vr9 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr9, vd9, vone), vr9, vr9);
246 
247     // Reconstruct sigmoid(z) = exp(z) * recip(1.0 + exp(z))
248     __m256 vf0 = _mm256_mul_ps(ve0, vr0);
249     __m256 vf1 = _mm256_mul_ps(ve1, vr1);
250     __m256 vf2 = _mm256_mul_ps(ve2, vr2);
251     __m256 vf3 = _mm256_mul_ps(ve3, vr3);
252     __m256 vf4 = _mm256_mul_ps(ve4, vr4);
253     __m256 vf5 = _mm256_mul_ps(ve5, vr5);
254     __m256 vf6 = _mm256_mul_ps(ve6, vr6);
255     __m256 vf7 = _mm256_mul_ps(ve7, vr7);
256     __m256 vf8 = _mm256_mul_ps(ve8, vr8);
257     __m256 vf9 = _mm256_mul_ps(ve9, vr9);
258 
259     // For inputs below denormal cutoff, replace output with +0.0f.
260     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
261     vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
262     vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
263     vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
264     vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
265     vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vz4, vdenorm_cutoff, _CMP_LT_OS), vf4);
266     vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vz5, vdenorm_cutoff, _CMP_LT_OS), vf5);
267     vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vz6, vdenorm_cutoff, _CMP_LT_OS), vf6);
268     vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vz7, vdenorm_cutoff, _CMP_LT_OS), vf7);
269     vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vz8, vdenorm_cutoff, _CMP_LT_OS), vf8);
270     vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vz9, vdenorm_cutoff, _CMP_LT_OS), vf9);
271 
272     // Reconstruct sigmoid(x) = x < 0 ? sigmoid(z) : 1.0 - sigmoid(z)
273     vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
274     vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
275     vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
276     vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
277     vf4 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf4), vf4, vx4);
278     vf5 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf5), vf5, vx5);
279     vf6 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf6), vf6, vx6);
280     vf7 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf7), vf7, vx7);
281     vf8 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf8), vf8, vx8);
282     vf9 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf9), vf9, vx9);
283 
284     _mm256_storeu_ps(y, vf0);
285     _mm256_storeu_ps(y + 8, vf1);
286     _mm256_storeu_ps(y + 16, vf2);
287     _mm256_storeu_ps(y + 24, vf3);
288     _mm256_storeu_ps(y + 32, vf4);
289     _mm256_storeu_ps(y + 40, vf5);
290     _mm256_storeu_ps(y + 48, vf6);
291     _mm256_storeu_ps(y + 56, vf7);
292     _mm256_storeu_ps(y + 64, vf8);
293     _mm256_storeu_ps(y + 72, vf9);
294     y += 80;
295   }
296   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
297     const __m256 vx = _mm256_loadu_ps(x);
298     x += 8;
299 
300     // General structure of the algorithm:
301     //           / exp(x) / (1 + exp(x)) if x <= 0
302     //   f[x] :=
303     //           \ 1 - f[-x] if x >= 0
304     //
305     // First we compute f[z] := exp(z) / (1 + exp(z)) where z = -abs(x),
306     // then replace result with 1 - f[z] if x >= 0.
307     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
308 
309     // Compute reduced argument n := round(z / log(2)).
310     // We do it by adding a large number (magic bias) to the product z * (1/log(2)), which cause rounding of the result
311     // to an integer, then subtracing the large number back. The trick with adding large number is valid only within
312     // certain bounds (|x| <= 2**22), but thats ok, because inputs x outside of [-87.336544, 17.328678] (i.e. z outsize
313     // [0, 87.336544]) underflow or saturate sigmoidf(x) anyway. We fixup the result for such inputs at the very end of
314     // the algorithm.
315     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
316 
317     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
318     // -87.33642 <= z <= 0.0, and -126 <= n <= 0 accordingly.
319     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
320 
321     // Subtract the large number back to get final n := round(z / log(2)).
322     vn = _mm256_sub_ps(vn, vmagic_bias);
323 
324     // Compute reduced argument t := z - n * log(2).
325     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
326 
327     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
328     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
329     vp = _mm256_fmadd_ps(vp, vt, vc3);
330     vp = _mm256_fmadd_ps(vp, vt, vc2);
331     vp = _mm256_fmadd_ps(vp, vt, vc1);
332 
333     // Reconstruct the exp(z) value:
334     //   e = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
335     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
336     //     = s + (t * s) * p
337     vt = _mm256_mul_ps(vt, vs);
338     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
339 
340     // Denominator of the sigmoid fraction: 1.0 + exp(z)
341     const __m256 vd = _mm256_add_ps(ve, vone);
342 
343     // Use Newton-Raphson method to compute reciprocal of denominator.
344     // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
345     // Thus the reciprocal of the denominator never overflows.
346     __m256 vr = _mm256_rcp_ps(vd);
347     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
348     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
349 
350     // Reconstruct sigmoid(z) = exp(z) * recip(1.0 + exp(z))
351     __m256 vf = _mm256_mul_ps(ve, vr);
352 
353     // For inputs below denormal cutoff, replace output with +0.0f.
354     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
355     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
356 
357     // Reconstruct sigmoid(x) = x < 0 ? sigmoid(z) : 1.0 - sigmoid(z)
358     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
359 
360     _mm256_storeu_ps(y, vf);
361     y += 8;
362   }
363   if XNN_UNLIKELY(n != 0) {
364     assert(n >= 1 * sizeof(float));
365     assert(n <= 7 * sizeof(float));
366     __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
367 
368     const __m256 vx = _mm256_maskload_ps(x, vmask);
369 
370     // General structure of the algorithm:
371     //           / exp(x) / (1 + exp(x)) if x <= 0
372     //   f[x] :=
373     //           \ 1 - f[-x] if x >= 0
374     //
375     // First we compute f[z] := exp(z) / (1 + exp(z)) where z = -abs(x),
376     // then replace result with 1 - f[z] if x >= 0.
377     const __m256 vz = _mm256_or_ps(vx, vsign_mask);
378 
379     // Compute reduced argument n := round(z / log(2)).
380     // We do it by adding a large number (magic bias) to the product z * (1/log(2)), which cause rounding of the result
381     // to an integer, then subtracing the large number back. The trick with adding large number is valid only within
382     // certain bounds (|x| <= 2**22), but thats ok, because inputs x outside of [-87.336544, 17.328678] (i.e. z outsize
383     // [0, 87.336544]) underflow or saturate sigmoidf(x) anyway. We fixup the result for such inputs at the very end of
384     // the algorithm.
385     __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
386 
387     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
388     // -87.33642 <= z <= 0.0, and -126 <= n <= 0 accordingly.
389     const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
390 
391     // Subtract the large number back to get final n := round(z / log(2)).
392     vn = _mm256_sub_ps(vn, vmagic_bias);
393 
394     // Compute reduced argument t := z - n * log(2).
395     __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
396 
397     // Compute degree-5 polynomial approxiatmion for exp(t) on [-log(2)/2, log(2)/2].
398     __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
399     vp = _mm256_fmadd_ps(vp, vt, vc3);
400     vp = _mm256_fmadd_ps(vp, vt, vc2);
401     vp = _mm256_fmadd_ps(vp, vt, vc1);
402 
403     // Reconstruct the exp(z) value:
404     //   e = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
405     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
406     //     = s + (t * s) * p
407     vt = _mm256_mul_ps(vt, vs);
408     const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
409 
410     // Denominator of the sigmoid fraction: 1.0 + exp(z)
411     const __m256 vd = _mm256_add_ps(ve, vone);
412 
413     // Use Newton-Raphson method to compute reciprocal of denominator.
414     // Note: 1 < d <= 2, because z >= 0.0 and 0 < exp(-z) <= 1.0.
415     // Thus the reciprocal of the denominator never overflows.
416     __m256 vr = _mm256_rcp_ps(vd);
417     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
418     vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
419 
420     // Reconstruct sigmoid(z) = exp(z) * recip(1.0 + exp(z))
421     __m256 vf = _mm256_mul_ps(ve, vr);
422 
423     // For inputs below denormal cutoff, replace output with +0.0f.
424     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
425     vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
426 
427     // Reconstruct sigmoid(x) = x < 0 ? sigmoid(z) : 1.0 - sigmoid(z)
428     vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
429 
430     // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
431     __m128 vf_lo = _mm256_castps256_ps128(vf);
432     if (n & (4 * sizeof(float))) {
433       _mm_storeu_ps(y, vf_lo);
434       vf_lo = _mm256_extractf128_ps(vf, 1);
435       y += 4;
436     }
437     if (n & (2 * sizeof(float))) {
438       _mm_storel_pi((__m64*) y, vf_lo);
439       vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
440       y += 2;
441     }
442     if (n & (1 * sizeof(float))) {
443       _mm_store_ss(y, vf_lo);
444     }
445   }
446 }
447