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_nr1fma_x24(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_nr1fma_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 x += 24;
43
44 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
45 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
46 const __m256 vz2 = _mm256_or_ps(vx2, vsign_mask);
47
48 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
49 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
50 __m256 vn2 = _mm256_fmadd_ps(vz2, vlog2e, vmagic_bias);
51
52 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
53 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
54 const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23));
55
56 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
57 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
58 vn2 = _mm256_sub_ps(vn2, vmagic_bias);
59
60 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
61 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
62 __m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2, vz2);
63
64 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
65 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
66 __m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4);
67
68 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
69 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
70 vp2 = _mm256_fmadd_ps(vp2, vt2, vc3);
71
72 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
73 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
74 vp2 = _mm256_fmadd_ps(vp2, vt2, vc2);
75
76 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
77 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
78 vp2 = _mm256_fmadd_ps(vp2, vt2, vc1);
79
80 vt0 = _mm256_mul_ps(vt0, vs0);
81 vt1 = _mm256_mul_ps(vt1, vs1);
82 vt2 = _mm256_mul_ps(vt2, vs2);
83
84 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
85 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
86 const __m256 ve2 = _mm256_fmadd_ps(vt2, vp2, vs2);
87
88 const __m256 vd0 = _mm256_add_ps(ve0, vone);
89 const __m256 vd1 = _mm256_add_ps(ve1, vone);
90 const __m256 vd2 = _mm256_add_ps(ve2, vone);
91
92 __m256 vr0 = _mm256_rcp_ps(vd0);
93 __m256 vr1 = _mm256_rcp_ps(vd1);
94 __m256 vr2 = _mm256_rcp_ps(vd2);
95
96 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
97 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
98 vr2 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr2, vd2, vone), vr2, vr2);
99
100
101 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
102 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
103 __m256 vf2 = _mm256_mul_ps(ve2, vr2);
104
105 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
106 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
107 vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vz2, vdenorm_cutoff, _CMP_LT_OS), vf2);
108
109 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
110 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
111 vf2 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf2), vf2, vx2);
112
113 _mm256_storeu_ps(y, vf0);
114 _mm256_storeu_ps(y + 8, vf1);
115 _mm256_storeu_ps(y + 16, vf2);
116 y += 24;
117 }
118 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
119 const __m256 vx = _mm256_loadu_ps(x);
120 x += 8;
121
122 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
123
124 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
125 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
126 vn = _mm256_sub_ps(vn, vmagic_bias);
127
128 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
129
130 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
131 vp = _mm256_fmadd_ps(vp, vt, vc3);
132 vp = _mm256_fmadd_ps(vp, vt, vc2);
133 vp = _mm256_fmadd_ps(vp, vt, vc1);
134
135 vt = _mm256_mul_ps(vt, vs);
136 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
137
138 const __m256 vd = _mm256_add_ps(ve, vone);
139 __m256 vr = _mm256_rcp_ps(vd);
140 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
141 __m256 vf = _mm256_mul_ps(ve, vr);
142
143 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
144 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
145
146 _mm256_storeu_ps(y, vf);
147 y += 8;
148 }
149 if XNN_UNLIKELY(n != 0) {
150 assert(n >= 1 * sizeof(float));
151 assert(n <= 7 * sizeof(float));
152 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - n));
153
154 const __m256 vx = _mm256_maskload_ps(x, vmask);
155
156 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
157
158 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
159 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
160 vn = _mm256_sub_ps(vn, vmagic_bias);
161
162 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
163
164 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
165 vp = _mm256_fmadd_ps(vp, vt, vc3);
166 vp = _mm256_fmadd_ps(vp, vt, vc2);
167 vp = _mm256_fmadd_ps(vp, vt, vc1);
168
169 vt = _mm256_mul_ps(vt, vs);
170 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
171
172 const __m256 vd = _mm256_add_ps(ve, vone);
173 __m256 vr = _mm256_rcp_ps(vd);
174 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
175 __m256 vf = _mm256_mul_ps(ve, vr);
176
177 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
178 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
179
180 __m128 vf_lo = _mm256_castps256_ps128(vf);
181 if (n & (4 * sizeof(float))) {
182 _mm_storeu_ps(y, vf_lo);
183 vf_lo = _mm256_extractf128_ps(vf, 1);
184 y += 4;
185 }
186 if (n & (2 * sizeof(float))) {
187 _mm_storel_pi((__m64*) y, vf_lo);
188 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
189 y += 2;
190 }
191 if (n & (1 * sizeof(float))) {
192 _mm_store_ss(y, vf_lo);
193 }
194 }
195 }
196