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_x32(size_t n,const float * x,float * y,const void * params)20 void xnn_f32_sigmoid_ukernel__avx2_rr1_p5_nr2fma_x32(
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 vsign_mask = _mm256_set1_ps(-0.0f);
29 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
30 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p0f);
31 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
32 const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f);
33 const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f);
34 const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f);
35 const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f);
36 const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f);
37 const __m256 vone = _mm256_set1_ps(1.0f);
38 const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep+6f);
39
40 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
41 const __m256 vx0 = _mm256_loadu_ps(x);
42 const __m256 vx1 = _mm256_loadu_ps(x + 8);
43 const __m256 vx2 = _mm256_loadu_ps(x + 16);
44 const __m256 vx3 = _mm256_loadu_ps(x + 24);
45 x += 32;
46
47 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
48 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
49 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
50 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
51
52 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
53 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
54 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
55 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
56
57 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
58 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
59 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
60 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
61
62 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
63 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
64 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
65 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
66
67 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
68 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
69 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
70 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
71
72 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
73 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
74 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
75 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
76
77 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
78 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
79 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
80 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
81
82 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
83 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
84 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
85 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
86
87 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
88 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
89 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
90 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
91
92 vt0 = _mm256_mul_ps(vt0, vs0);
93 vt1 = _mm256_mul_ps(vt1, vs1);
94 vt2 = _mm256_mul_ps(vt2, vs2);
95 vt3 = _mm256_mul_ps(vt3, vs3);
96
97 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
98 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
99 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
100 const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
101
102 const __m256 vd0 = _mm256_add_ps(ve0, vone);
103 const __m256 vd1 = _mm256_add_ps(ve1, vone);
104 const __m256 vd2 = _mm256_add_ps(ve2, vone);
105 const __m256 vd3 = _mm256_add_ps(ve3, vone);
106
107 __m256 vr0 = _mm256_rcp_ps(vd0);
108 __m256 vr1 = _mm256_rcp_ps(vd1);
109 __m256 vr2 = _mm256_rcp_ps(vd2);
110 __m256 vr3 = _mm256_rcp_ps(vd3);
111
112 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
113 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
114 vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
115 vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
116
117 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
118 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
119 vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
120 vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
121
122 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
123 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
124 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
125 __m256 vf3 = _mm256_mul_ps(ve3, vr3);
126
127 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
128 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
129 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
130 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
131
132 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
133 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
134 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
135 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
136
137 _mm256_storeu_ps(y, vf0);
138 _mm256_storeu_ps(y + 8, vf1);
139 _mm256_storeu_ps(y + 16, vf2);
140 _mm256_storeu_ps(y + 24, vf3);
141 y += 32;
142 }
143 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
144 const __m256 vx = _mm256_loadu_ps(x);
145 x += 8;
146
147 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
148
149 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
150 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
151 vn = _mm256_sub_ps(vn, vmagic_bias);
152
153 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
154
155 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
156 vp = _mm256_fmadd_ps(vp, vt, vc3);
157 vp = _mm256_fmadd_ps(vp, vt, vc2);
158 vp = _mm256_fmadd_ps(vp, vt, vc1);
159
160 vt = _mm256_mul_ps(vt, vs);
161 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
162
163 const __m256 vd = _mm256_add_ps(ve, vone);
164 __m256 vr = _mm256_rcp_ps(vd);
165 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
166 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
167 __m256 vf = _mm256_mul_ps(ve, vr);
168
169 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
170 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
171
172 _mm256_storeu_ps(y, vf);
173 y += 8;
174 }
175 if XNN_UNLIKELY(n != 0) {
176 assert(n >= 1 * sizeof(float));
177 assert(n <= 7 * sizeof(float));
178 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
179
180 const __m256 vx = _mm256_maskload_ps(x, vmask);
181
182 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
183
184 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
185 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
186 vn = _mm256_sub_ps(vn, vmagic_bias);
187
188 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
189
190 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
191 vp = _mm256_fmadd_ps(vp, vt, vc3);
192 vp = _mm256_fmadd_ps(vp, vt, vc2);
193 vp = _mm256_fmadd_ps(vp, vt, vc1);
194
195 vt = _mm256_mul_ps(vt, vs);
196 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
197
198 const __m256 vd = _mm256_add_ps(ve, vone);
199 __m256 vr = _mm256_rcp_ps(vd);
200 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
201 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
202 __m256 vf = _mm256_mul_ps(ve, vr);
203
204 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
205 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
206
207 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
208 __m128 vf_lo = _mm256_castps256_ps128(vf);
209 if (n & (4 * sizeof(float))) {
210 _mm_storeu_ps(y, vf_lo);
211 vf_lo = _mm256_extractf128_ps(vf, 1);
212 y += 4;
213 }
214 if (n & (2 * sizeof(float))) {
215 _mm_storel_pi((__m64*) y, vf_lo);
216 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
217 y += 2;
218 }
219 if (n & (1 * sizeof(float))) {
220 _mm_store_ss(y, vf_lo);
221 }
222 }
223 }
224