1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vsigmoid/avx2-rr1-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
xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x32(size_t n,const float * x,float * y,const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr2fma_x32(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p5.sign_mask);
27 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
28 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
29 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
30 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
31 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
32 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
33 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
34 const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
35 const __m256 vone = _mm256_load_ps(params->avx2_rr1_p5.one);
36 const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);
37
38 for (; n >= 32 * sizeof(float); n -= 32 * sizeof(float)) {
39 const __m256 vx0 = _mm256_loadu_ps(x);
40 const __m256 vx1 = _mm256_loadu_ps(x + 8);
41 const __m256 vx2 = _mm256_loadu_ps(x + 16);
42 const __m256 vx3 = _mm256_loadu_ps(x + 24);
43 x += 32;
44
45 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
46 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
47 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
48 const __m256 vz3 = _mm256_or_ps(vx3, vsign_mask);
49
50 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
51 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
52 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
53 __m256 vn3 = _mm256_fmadd_ps(vz3, vlog2e, vmagic_bias);
54
55 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
56 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
57 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
58 const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23));
59
60 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
61 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
62 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
63 vn3 = _mm256_sub_ps(vn3, vmagic_bias);
64
65 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
66 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
67 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
68 __m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2, vz3);
69
70 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
71 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
72 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
73 __m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4);
74
75 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
76 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
77 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
78 vp3 = _mm256_fmadd_ps(vp3, vt3, vc3);
79
80 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
81 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
82 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
83 vp3 = _mm256_fmadd_ps(vp3, vt3, vc2);
84
85 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
86 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
87 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
88 vp3 = _mm256_fmadd_ps(vp3, vt3, vc1);
89
90 vt0 = _mm256_mul_ps(vt0, vs0);
91 vt1 = _mm256_mul_ps(vt1, vs1);
92 vt2 = _mm256_mul_ps(vt2, vs2);
93 vt3 = _mm256_mul_ps(vt3, vs3);
94
95 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
96 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
97 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
98 const __m256 ve3 = _mm256_fmadd_ps(vt3, vp3, vs3);
99
100 const __m256 vd0 = _mm256_add_ps(ve0, vone);
101 const __m256 vd1 = _mm256_add_ps(ve1, vone);
102 const __m256 vd2 = _mm256_add_ps(ve2, vone);
103 const __m256 vd3 = _mm256_add_ps(ve3, vone);
104
105 __m256 vr0 = _mm256_rcp_ps(vd0);
106 __m256 vr1 = _mm256_rcp_ps(vd1);
107 __m256 vr2 = _mm256_rcp_ps(vd2);
108 __m256 vr3 = _mm256_rcp_ps(vd3);
109
110 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
111 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
112 vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
113 vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
114
115 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
116 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
117 vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
118 vr3 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr3, vd3, vone), vr3, vr3);
119
120 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
121 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
122 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
123 __m256 vf3 = _mm256_mul_ps(ve3, vr3);
124
125 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
126 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
127 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
128 vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vz3, vdenorm_cutoff, _CMP_LT_OS), vf3);
129
130 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
131 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
132 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
133 vf3 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf3), vf3, vx3);
134
135 _mm256_storeu_ps(y, vf0);
136 _mm256_storeu_ps(y + 8, vf1);
137 _mm256_storeu_ps(y + 16, vf2);
138 _mm256_storeu_ps(y + 24, vf3);
139 y += 32;
140 }
141 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
142 const __m256 vx = _mm256_loadu_ps(x);
143 x += 8;
144
145 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
146
147 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
148 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
149 vn = _mm256_sub_ps(vn, vmagic_bias);
150
151 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
152
153 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
154 vp = _mm256_fmadd_ps(vp, vt, vc3);
155 vp = _mm256_fmadd_ps(vp, vt, vc2);
156 vp = _mm256_fmadd_ps(vp, vt, vc1);
157
158 vt = _mm256_mul_ps(vt, vs);
159 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
160
161 const __m256 vd = _mm256_add_ps(ve, vone);
162 __m256 vr = _mm256_rcp_ps(vd);
163 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
164 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
165 __m256 vf = _mm256_mul_ps(ve, vr);
166
167 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
168 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
169
170 _mm256_storeu_ps(y, vf);
171 y += 8;
172 }
173 if XNN_UNLIKELY(n != 0) {
174 assert(n >= 1 * sizeof(float));
175 assert(n <= 7 * sizeof(float));
176 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - n));
177
178 const __m256 vx = _mm256_maskload_ps(x, vmask);
179
180 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
181
182 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
183 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
184 vn = _mm256_sub_ps(vn, vmagic_bias);
185
186 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
187
188 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
189 vp = _mm256_fmadd_ps(vp, vt, vc3);
190 vp = _mm256_fmadd_ps(vp, vt, vc2);
191 vp = _mm256_fmadd_ps(vp, vt, vc1);
192
193 vt = _mm256_mul_ps(vt, vs);
194 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
195
196 const __m256 vd = _mm256_add_ps(ve, vone);
197 __m256 vr = _mm256_rcp_ps(vd);
198 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
199 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
200 __m256 vf = _mm256_mul_ps(ve, vr);
201
202 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
203 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
204
205 __m128 vf_lo = _mm256_castps256_ps128(vf);
206 if (n & (4 * sizeof(float))) {
207 _mm_storeu_ps(y, vf_lo);
208 vf_lo = _mm256_extractf128_ps(vf, 1);
209 y += 4;
210 }
211 if (n & (2 * sizeof(float))) {
212 _mm_storel_pi((__m64*) y, vf_lo);
213 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
214 y += 2;
215 }
216 if (n & (1 * sizeof(float))) {
217 _mm_store_ss(y, vf_lo);
218 }
219 }
220 }
221