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_x12(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_x12(
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_DISABLE_TSAN
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->sse.prescale);
30 const __m128 valpha = _mm_load_ps(params->sse.alpha);
31 const __m128 vbeta = _mm_load_ps(params->sse.beta);
32
33 const __m128 vsat_cutoff = _mm_set1_ps(-0x1.154246p+4f);
34 const __m128 vmagic_bias = _mm_set1_ps(0x1.8000FEp23f);
35 const __m128 vlog2e = _mm_set1_ps(0x1.715476p+0f);
36 const __m128 vminus_ln2_hi = _mm_set1_ps(-0x1.62E440p-1f);
37 const __m128 vminus_ln2_lo = _mm_set1_ps(0x1.0105C6p-21f);
38 const __m128 vc6 = _mm_set1_ps(0x1.6b7338p-10f);
39 const __m128 vc5 = _mm_set1_ps(0x1.12278Ep-7f);
40 const __m128 vc4 = _mm_set1_ps(0x1.555716p-5f);
41 const __m128 vc3 = _mm_set1_ps(0x1.5554B0p-3f);
42 const __m128 vc2 = _mm_set1_ps(0x1.FFFFFEp-2f);
43 const __m128 vone = _mm_set1_ps(1.0f);
44
45 for (; n >= 12 * sizeof(float); n -= 12 * sizeof(float)) {
46 __m128 vx0123 = _mm_loadu_ps(x);
47 __m128 vx4567 = _mm_loadu_ps(x + 4);
48 __m128 vx89AB = _mm_loadu_ps(x + 8);
49 x += 12;
50
51 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
52 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
53 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
54
55 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
56 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
57 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
58
59 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
60 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
61 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
62
63 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
64 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
65 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
66
67 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
68 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
69 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
70
71 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
72 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
73 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
74
75 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
76 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
77 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
78
79 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
80 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
81 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
82
83 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
84 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
85 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
86
87 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
88 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
89 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
90
91 vp0123 = _mm_mul_ps(vp0123, vt0123);
92 vp4567 = _mm_mul_ps(vp4567, vt4567);
93 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
94
95 vt0123 = _mm_mul_ps(vt0123, vs0123);
96 vs0123 = _mm_sub_ps(vs0123, vone);
97 vt4567 = _mm_mul_ps(vt4567, vs4567);
98 vs4567 = _mm_sub_ps(vs4567, vone);
99 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
100 vs89AB = _mm_sub_ps(vs89AB, vone);
101
102 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
103 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
104 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
105
106 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
107 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
108 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
109
110 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
111 vx0123 = _mm_mul_ps(vx0123, vbeta);
112 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
113 vx4567 = _mm_mul_ps(vx4567, vbeta);
114 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
115 vx89AB = _mm_mul_ps(vx89AB, vbeta);
116
117 const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
118 const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
119 const __m128 vy89AB = _mm_or_ps(_mm_and_ps(ve89AB, vm89AB), _mm_andnot_ps(vm89AB, vx89AB));
120
121 _mm_storeu_ps(y, vy0123);
122 _mm_storeu_ps(y + 4, vy4567);
123 _mm_storeu_ps(y + 8, vy89AB);
124 y += 12;
125 }
126 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
127 __m128 vx = _mm_loadu_ps(x);
128 x += 4;
129
130 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
131
132 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
133 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
134 vn = _mm_sub_ps(vn, vmagic_bias);
135
136 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
137 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
138
139 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
140 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
141 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
142 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
143 vp = _mm_mul_ps(vp, vt);
144
145 vt = _mm_mul_ps(vt, vs);
146 vs = _mm_sub_ps(vs, vone);
147 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
148 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
149
150 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
151 vx = _mm_mul_ps(vx, vbeta);
152 const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
153
154 _mm_storeu_ps(y, vy);
155 y += 4;
156 }
157 if XNN_UNLIKELY(n != 0) {
158 __m128 vx = _mm_loadu_ps(x);
159
160 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
161
162 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
163 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
164 vn = _mm_sub_ps(vn, vmagic_bias);
165
166 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
167 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
168
169 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
170 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
171 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
172 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
173 vp = _mm_mul_ps(vp, vt);
174
175 vt = _mm_mul_ps(vt, vs);
176 vs = _mm_sub_ps(vs, vone);
177 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
178 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
179
180 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
181 vx = _mm_mul_ps(vx, vbeta);
182 __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
183
184 if (n & (2 * sizeof(float))) {
185 _mm_storel_pi((__m64*) y, vy);
186 vy = _mm_movehl_ps(vy, vy);
187 y += 2;
188 }
189 if (n & (1 * sizeof(float))) {
190 _mm_store_ss(y, vy);
191 }
192 }
193 }
194