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_x16(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_x16(
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 >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
39 const __m256 vx0 = _mm256_loadu_ps(x);
40 const __m256 vx1 = _mm256_loadu_ps(x + 8);
41 x += 16;
42
43 const __m256 vz0 = _mm256_or_ps(vx0, vsign_mask);
44 const __m256 vz1 = _mm256_or_ps(vx1, vsign_mask);
45
46 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
47 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
48
49 const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
50 const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
51
52 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
53 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
54
55 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
56 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
57
58 __m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4);
59 __m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4);
60
61 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
62 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
63
64 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
65 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
66
67 vp0 = _mm256_fmadd_ps(vp0, vt0, vc1);
68 vp1 = _mm256_fmadd_ps(vp1, vt1, vc1);
69
70 vt0 = _mm256_mul_ps(vt0, vs0);
71 vt1 = _mm256_mul_ps(vt1, vs1);
72
73 const __m256 ve0 = _mm256_fmadd_ps(vt0, vp0, vs0);
74 const __m256 ve1 = _mm256_fmadd_ps(vt1, vp1, vs1);
75
76 const __m256 vd0 = _mm256_add_ps(ve0, vone);
77 const __m256 vd1 = _mm256_add_ps(ve1, vone);
78
79 __m256 vr0 = _mm256_rcp_ps(vd0);
80 __m256 vr1 = _mm256_rcp_ps(vd1);
81
82 vr0 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr0, vd0, vone), vr0, vr0);
83 vr1 = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr1, vd1, vone), vr1, vr1);
84
85
86 __m256 vf0 = _mm256_mul_ps(ve0, vr0);
87 __m256 vf1 = _mm256_mul_ps(ve1, vr1);
88
89 vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vz0, vdenorm_cutoff, _CMP_LT_OS), vf0);
90 vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vz1, vdenorm_cutoff, _CMP_LT_OS), vf1);
91
92 vf0 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf0), vf0, vx0);
93 vf1 = _mm256_blendv_ps(_mm256_sub_ps(vone, vf1), vf1, vx1);
94
95 _mm256_storeu_ps(y, vf0);
96 _mm256_storeu_ps(y + 8, vf1);
97 y += 16;
98 }
99 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
100 const __m256 vx = _mm256_loadu_ps(x);
101 x += 8;
102
103 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
104
105 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
106 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
107 vn = _mm256_sub_ps(vn, vmagic_bias);
108
109 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
110
111 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
112 vp = _mm256_fmadd_ps(vp, vt, vc3);
113 vp = _mm256_fmadd_ps(vp, vt, vc2);
114 vp = _mm256_fmadd_ps(vp, vt, vc1);
115
116 vt = _mm256_mul_ps(vt, vs);
117 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
118
119 const __m256 vd = _mm256_add_ps(ve, vone);
120 __m256 vr = _mm256_rcp_ps(vd);
121 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
122 __m256 vf = _mm256_mul_ps(ve, vr);
123
124 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
125 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
126
127 _mm256_storeu_ps(y, vf);
128 y += 8;
129 }
130 if XNN_UNLIKELY(n != 0) {
131 assert(n >= 1 * sizeof(float));
132 assert(n <= 7 * sizeof(float));
133 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p5.mask_table[7] - n));
134
135 const __m256 vx = _mm256_maskload_ps(x, vmask);
136
137 const __m256 vz = _mm256_or_ps(vx, vsign_mask);
138
139 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
140 const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
141 vn = _mm256_sub_ps(vn, vmagic_bias);
142
143 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
144
145 __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
146 vp = _mm256_fmadd_ps(vp, vt, vc3);
147 vp = _mm256_fmadd_ps(vp, vt, vc2);
148 vp = _mm256_fmadd_ps(vp, vt, vc1);
149
150 vt = _mm256_mul_ps(vt, vs);
151 const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);
152
153 const __m256 vd = _mm256_add_ps(ve, vone);
154 __m256 vr = _mm256_rcp_ps(vd);
155 vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
156 __m256 vf = _mm256_mul_ps(ve, vr);
157
158 vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
159 vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);
160
161 __m128 vf_lo = _mm256_castps256_ps128(vf);
162 if (n & (4 * sizeof(float))) {
163 _mm_storeu_ps(y, vf_lo);
164 vf_lo = _mm256_extractf128_ps(vf, 1);
165 y += 4;
166 }
167 if (n & (2 * sizeof(float))) {
168 _mm_storel_pi((__m64*) y, vf_lo);
169 vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
170 y += 2;
171 }
172 if (n & (1 * sizeof(float))) {
173 _mm_store_ss(y, vf_lo);
174 }
175 }
176 }
177