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_x20(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_x20(
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 >= 20 * sizeof(float); n -= 20 * sizeof(float)) {
45 __m128 vx0123 = _mm_loadu_ps(x);
46 __m128 vx4567 = _mm_loadu_ps(x + 4);
47 __m128 vx89AB = _mm_loadu_ps(x + 8);
48 __m128 vxCDEF = _mm_loadu_ps(x + 12);
49 __m128 vxGHIJ = _mm_loadu_ps(x + 16);
50 x += 20;
51
52 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
53 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
54 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
55 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
56 const __m128 vzGHIJ = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxGHIJ, vprescale));
57
58 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
59 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
60 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
61 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
62 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
63
64 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
65 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
66 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
67 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
68 __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
69
70 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
71 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
72 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
73 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
74 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
75
76 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
77 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
78 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
79 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
80 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
81
82 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
83 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
84 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
85 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
86 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
87
88 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
89 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
90 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
91 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
92 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5);
93
94 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
95 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
96 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
97 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
98 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4);
99
100 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
101 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
102 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
103 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
104 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
105
106 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
107 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
108 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
109 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
110 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
111
112 vp0123 = _mm_mul_ps(vp0123, vt0123);
113 vp4567 = _mm_mul_ps(vp4567, vt4567);
114 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
115 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
116 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ);
117
118 vt0123 = _mm_mul_ps(vt0123, vs0123);
119 vs0123 = _mm_sub_ps(vs0123, vone);
120 vt4567 = _mm_mul_ps(vt4567, vs4567);
121 vs4567 = _mm_sub_ps(vs4567, vone);
122 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
123 vs89AB = _mm_sub_ps(vs89AB, vone);
124 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
125 vsCDEF = _mm_sub_ps(vsCDEF, vone);
126 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
127 vsGHIJ = _mm_sub_ps(vsGHIJ, vone);
128
129 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
130 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
131 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
132 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
133 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ);
134
135 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
136 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
137 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
138 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
139 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha);
140
141 const __m128 vm0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx0123)));
142 vx0123 = _mm_mul_ps(vx0123, vbeta);
143 const __m128 vm4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx4567)));
144 vx4567 = _mm_mul_ps(vx4567, vbeta);
145 const __m128 vm89AB = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx89AB)));
146 vx89AB = _mm_mul_ps(vx89AB, vbeta);
147 const __m128 vmCDEF = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxCDEF)));
148 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
149 const __m128 vmGHIJ = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vxGHIJ)));
150 vxGHIJ = _mm_mul_ps(vxGHIJ, vbeta);
151
152 const __m128 vy0123 = _mm_or_ps(_mm_and_ps(ve0123, vm0123), _mm_andnot_ps(vm0123, vx0123));
153 const __m128 vy4567 = _mm_or_ps(_mm_and_ps(ve4567, vm4567), _mm_andnot_ps(vm4567, vx4567));
154 const __m128 vy89AB = _mm_or_ps(_mm_and_ps(ve89AB, vm89AB), _mm_andnot_ps(vm89AB, vx89AB));
155 const __m128 vyCDEF = _mm_or_ps(_mm_and_ps(veCDEF, vmCDEF), _mm_andnot_ps(vmCDEF, vxCDEF));
156 const __m128 vyGHIJ = _mm_or_ps(_mm_and_ps(veGHIJ, vmGHIJ), _mm_andnot_ps(vmGHIJ, vxGHIJ));
157
158 _mm_storeu_ps(y, vy0123);
159 _mm_storeu_ps(y + 4, vy4567);
160 _mm_storeu_ps(y + 8, vy89AB);
161 _mm_storeu_ps(y + 12, vyCDEF);
162 _mm_storeu_ps(y + 16, vyGHIJ);
163 y += 20;
164 }
165 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
166 __m128 vx = _mm_loadu_ps(x);
167 x += 4;
168
169 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
170
171 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
172 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
173 vn = _mm_sub_ps(vn, vmagic_bias);
174
175 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
176 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
177
178 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
179 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
180 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
181 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
182 vp = _mm_mul_ps(vp, vt);
183
184 vt = _mm_mul_ps(vt, vs);
185 vs = _mm_sub_ps(vs, vone);
186 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
187 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
188
189 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
190 vx = _mm_mul_ps(vx, vbeta);
191 const __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
192
193 _mm_storeu_ps(y, vy);
194 y += 4;
195 }
196 if XNN_UNLIKELY(n != 0) {
197 __m128 vx = _mm_loadu_ps(x);
198
199 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
200
201 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
202 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
203 vn = _mm_sub_ps(vn, vmagic_bias);
204
205 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
206 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
207
208 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
209 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
210 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
211 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
212 vp = _mm_mul_ps(vp, vt);
213
214 vt = _mm_mul_ps(vt, vs);
215 vs = _mm_sub_ps(vs, vone);
216 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
217 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
218
219 const __m128 vm = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(vx)));
220 vx = _mm_mul_ps(vx, vbeta);
221 __m128 vy = _mm_or_ps(_mm_and_ps(ve, vm), _mm_andnot_ps(vm, vx));
222
223 if (n & (2 * sizeof(float))) {
224 _mm_storel_pi((__m64*) y, vy);
225 vy = _mm_movehl_ps(vy, vy);
226 y += 2;
227 }
228 if (n & (1 * sizeof(float))) {
229 _mm_store_ss(y, vy);
230 }
231 }
232 }
233