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 <smmintrin.h>
13
14 #include <xnnpack/vunary.h>
15 #include <xnnpack/common.h>
16
17
xnn_f32_velu_ukernel__sse41_rr2_p6_x16(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__sse41_rr2_p6_x16(
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 >= 16 * sizeof(float); n -= 16 * 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 x += 16;
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 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
55
56 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
57 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
58 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
59 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
60
61 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
62 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
63 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
64 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
65
66 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
67 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
68 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
69 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
70
71 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
72 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
73 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
74 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
75
76 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
77 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
78 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
79 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
80
81 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
82 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
83 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
84 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
85
86 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
88 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
89 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
90
91 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
92 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
93 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
94 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
95
96 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
97 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
98 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
99 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
100
101 vp0123 = _mm_mul_ps(vp0123, vt0123);
102 vp4567 = _mm_mul_ps(vp4567, vt4567);
103 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
104 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
105
106 vt0123 = _mm_mul_ps(vt0123, vs0123);
107 vs0123 = _mm_sub_ps(vs0123, vone);
108 vt4567 = _mm_mul_ps(vt4567, vs4567);
109 vs4567 = _mm_sub_ps(vs4567, vone);
110 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
111 vs89AB = _mm_sub_ps(vs89AB, vone);
112 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
113 vsCDEF = _mm_sub_ps(vsCDEF, vone);
114
115 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
116 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
117 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
118 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
119
120 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
121 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
122 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
123 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
124
125 vx0123 = _mm_mul_ps(vx0123, vbeta);
126 vx4567 = _mm_mul_ps(vx4567, vbeta);
127 vx89AB = _mm_mul_ps(vx89AB, vbeta);
128 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
129
130 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
131 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
132 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
133 const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
134
135 _mm_storeu_ps(y, vy0123);
136 _mm_storeu_ps(y + 4, vy4567);
137 _mm_storeu_ps(y + 8, vy89AB);
138 _mm_storeu_ps(y + 12, vyCDEF);
139 y += 16;
140 }
141 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
142 __m128 vx = _mm_loadu_ps(x);
143 x += 4;
144
145 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
146
147 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
148 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
149 vn = _mm_sub_ps(vn, vmagic_bias);
150
151 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
152 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
153
154 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
155 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
156 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
157 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
158 vp = _mm_mul_ps(vp, vt);
159
160 vt = _mm_mul_ps(vt, vs);
161 vs = _mm_sub_ps(vs, vone);
162 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
163 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
164
165 vx = _mm_mul_ps(vx, vbeta);
166 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
167
168 _mm_storeu_ps(y, vy);
169 y += 4;
170 }
171 if XNN_UNLIKELY(n != 0) {
172 __m128 vx = _mm_loadu_ps(x);
173
174 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
175
176 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
177 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
178 vn = _mm_sub_ps(vn, vmagic_bias);
179
180 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
181 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
182
183 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
184 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
185 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
186 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
187 vp = _mm_mul_ps(vp, vt);
188
189 vt = _mm_mul_ps(vt, vs);
190 vs = _mm_sub_ps(vs, vone);
191 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
192 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
193
194 vx = _mm_mul_ps(vx, vbeta);
195 __m128 vy = _mm_blendv_ps(vx, ve, vx);
196
197 if (n & (2 * sizeof(float))) {
198 _mm_storel_pi((__m64*) y, vy);
199 vy = _mm_movehl_ps(vy, vy);
200 y += 2;
201 }
202 if (n & (1 * sizeof(float))) {
203 _mm_store_ss(y, vy);
204 }
205 }
206 }
207