1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/sse-rr2-p6.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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 <emmintrin.h>
13
14 #include <xnnpack/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__sse2_rr2_p6_x8(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_velu_ukernel__sse2_rr2_p6_x8(
19 size_t n,
20 const float* x,
21 float* y,
22 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
23 {
24 assert(n != 0);
25 assert(n % sizeof(float) == 0);
26 assert(x != NULL);
27 assert(y != NULL);
28
29 const __m128 vprescale = _mm_load_ps(params->sse2_rr2_p6.prescale);
30 const __m128 valpha = _mm_load_ps(params->sse2_rr2_p6.alpha);
31 const __m128 vbeta = _mm_load_ps(params->sse2_rr2_p6.beta);
32 const __m128 vsat_cutoff = _mm_load_ps(params->sse2_rr2_p6.sat_cutoff);
33 const __m128 vmagic_bias = _mm_load_ps(params->sse2_rr2_p6.magic_bias);
34 const __m128 vlog2e = _mm_load_ps(params->sse2_rr2_p6.log2e);
35 const __m128 vminus_ln2_hi = _mm_load_ps(params->sse2_rr2_p6.minus_ln2_hi);
36 const __m128 vminus_ln2_lo = _mm_load_ps(params->sse2_rr2_p6.minus_ln2_lo);
37 const __m128 vc6 = _mm_load_ps(params->sse2_rr2_p6.c6);
38 const __m128 vc5 = _mm_load_ps(params->sse2_rr2_p6.c5);
39 const __m128 vc4 = _mm_load_ps(params->sse2_rr2_p6.c4);
40 const __m128 vc3 = _mm_load_ps(params->sse2_rr2_p6.c3);
41 const __m128 vc2 = _mm_load_ps(params->sse2_rr2_p6.c2);
42 const __m128 vone = _mm_load_ps(params->sse2_rr2_p6.one);
43
44 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
45 __m128 vx0123 = _mm_loadu_ps(x);
46 __m128 vx4567 = _mm_loadu_ps(x + 4);
47 x += 8;
48
49 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
50 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
51
52 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
53 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
54
55 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
56 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
57
58 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
59 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
60
61 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
62 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
63
64 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
65 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
66
67 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
68 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
69
70 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
71 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
72
73 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
74 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
75
76 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
77 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
78
79 vp0123 = _mm_mul_ps(vp0123, vt0123);
80 vp4567 = _mm_mul_ps(vp4567, vt4567);
81
82 vt0123 = _mm_mul_ps(vt0123, vs0123);
83 vs0123 = _mm_sub_ps(vs0123, vone);
84 vt4567 = _mm_mul_ps(vt4567, vs4567);
85 vs4567 = _mm_sub_ps(vs4567, vone);
86
87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
89
90 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
91 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
92
93 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
94 vx0123 = _mm_mul_ps(vx0123, vbeta);
95 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
96 vx4567 = _mm_mul_ps(vx4567, vbeta);
97
98 const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
99 const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
100
101 _mm_storeu_ps(y, vy0123);
102 _mm_storeu_ps(y + 4, vy4567);
103 y += 8;
104 }
105 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
106 __m128 vx = _mm_loadu_ps(x);
107 x += 4;
108
109 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
110
111 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
112 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
113 vn = _mm_sub_ps(vn, vmagic_bias);
114
115 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
116 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
117
118 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
119 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
120 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
121 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
122 vp = _mm_mul_ps(vp, vt);
123
124 vt = _mm_mul_ps(vt, vs);
125 vs = _mm_sub_ps(vs, vone);
126 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
127 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
128
129 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
130 vx = _mm_mul_ps(vx, vbeta);
131 const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
132
133 _mm_storeu_ps(y, vy);
134 y += 4;
135 }
136 if XNN_UNLIKELY(n != 0) {
137 __m128 vx = _mm_loadu_ps(x);
138
139 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
140
141 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
142 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
143 vn = _mm_sub_ps(vn, vmagic_bias);
144
145 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
146 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
147
148 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
149 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
150 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
151 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
152 vp = _mm_mul_ps(vp, vt);
153
154 vt = _mm_mul_ps(vt, vs);
155 vs = _mm_sub_ps(vs, vone);
156 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
157 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
158
159 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
160 vx = _mm_mul_ps(vx, vbeta);
161 __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
162
163 if (n & (2 * sizeof(float))) {
164 _mm_storel_pi((__m64*) y, vy);
165 vy = _mm_movehl_ps(vy, vy);
166 y += 2;
167 }
168 if (n & (1 * sizeof(float))) {
169 _mm_store_ss(y, vy);
170 }
171 }
172 }
173