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_x24(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_x24(
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 >= 24 * sizeof(float); n -= 24 * 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 __m128 vxKLMN = _mm_loadu_ps(x + 20);
51 x += 24;
52
53 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
54 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
55 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
56 const __m128 vzCDEF = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxCDEF, vprescale));
57 const __m128 vzGHIJ = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxGHIJ, vprescale));
58 const __m128 vzKLMN = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vxKLMN, vprescale));
59
60 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
61 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
62 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
63 __m128 vnCDEF = _mm_add_ps(_mm_mul_ps(vzCDEF, vlog2e), vmagic_bias);
64 __m128 vnGHIJ = _mm_add_ps(_mm_mul_ps(vzGHIJ, vlog2e), vmagic_bias);
65 __m128 vnKLMN = _mm_add_ps(_mm_mul_ps(vzKLMN, vlog2e), vmagic_bias);
66
67 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
68 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
69 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
70 __m128 vsCDEF = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnCDEF), 23));
71 __m128 vsGHIJ = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnGHIJ), 23));
72 __m128 vsKLMN = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vnKLMN), 23));
73
74 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
75 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
76 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
77 vnCDEF = _mm_sub_ps(vnCDEF, vmagic_bias);
78 vnGHIJ = _mm_sub_ps(vnGHIJ, vmagic_bias);
79 vnKLMN = _mm_sub_ps(vnKLMN, vmagic_bias);
80
81 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
82 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
83 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
84 __m128 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_hi), vzCDEF);
85 __m128 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_hi), vzGHIJ);
86 __m128 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_hi), vzKLMN);
87
88 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
89 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
90 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
91 vtCDEF = _mm_add_ps(_mm_mul_ps(vnCDEF, vminus_ln2_lo), vtCDEF);
92 vtGHIJ = _mm_add_ps(_mm_mul_ps(vnGHIJ, vminus_ln2_lo), vtGHIJ);
93 vtKLMN = _mm_add_ps(_mm_mul_ps(vnKLMN, vminus_ln2_lo), vtKLMN);
94
95 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
96 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
97 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
98 __m128 vpCDEF = _mm_add_ps(_mm_mul_ps(vc6, vtCDEF), vc5);
99 __m128 vpGHIJ = _mm_add_ps(_mm_mul_ps(vc6, vtGHIJ), vc5);
100 __m128 vpKLMN = _mm_add_ps(_mm_mul_ps(vc6, vtKLMN), vc5);
101
102 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
103 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
104 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
105 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc4);
106 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc4);
107 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc4);
108
109 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
110 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
111 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
112 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc3);
113 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc3);
114 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc3);
115
116 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
117 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
118 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
119 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vc2);
120 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vc2);
121 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vc2);
122
123 vp0123 = _mm_mul_ps(vp0123, vt0123);
124 vp4567 = _mm_mul_ps(vp4567, vt4567);
125 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
126 vpCDEF = _mm_mul_ps(vpCDEF, vtCDEF);
127 vpGHIJ = _mm_mul_ps(vpGHIJ, vtGHIJ);
128 vpKLMN = _mm_mul_ps(vpKLMN, vtKLMN);
129
130 vt0123 = _mm_mul_ps(vt0123, vs0123);
131 vs0123 = _mm_sub_ps(vs0123, vone);
132 vt4567 = _mm_mul_ps(vt4567, vs4567);
133 vs4567 = _mm_sub_ps(vs4567, vone);
134 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
135 vs89AB = _mm_sub_ps(vs89AB, vone);
136 vtCDEF = _mm_mul_ps(vtCDEF, vsCDEF);
137 vsCDEF = _mm_sub_ps(vsCDEF, vone);
138 vtGHIJ = _mm_mul_ps(vtGHIJ, vsGHIJ);
139 vsGHIJ = _mm_sub_ps(vsGHIJ, vone);
140 vtKLMN = _mm_mul_ps(vtKLMN, vsKLMN);
141 vsKLMN = _mm_sub_ps(vsKLMN, vone);
142
143 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
144 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
145 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
146 vpCDEF = _mm_add_ps(_mm_mul_ps(vpCDEF, vtCDEF), vtCDEF);
147 vpGHIJ = _mm_add_ps(_mm_mul_ps(vpGHIJ, vtGHIJ), vtGHIJ);
148 vpKLMN = _mm_add_ps(_mm_mul_ps(vpKLMN, vtKLMN), vtKLMN);
149
150 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
151 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
152 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
153 const __m128 veCDEF = _mm_mul_ps(_mm_add_ps(vpCDEF, vsCDEF), valpha);
154 const __m128 veGHIJ = _mm_mul_ps(_mm_add_ps(vpGHIJ, vsGHIJ), valpha);
155 const __m128 veKLMN = _mm_mul_ps(_mm_add_ps(vpKLMN, vsKLMN), valpha);
156
157 vx0123 = _mm_mul_ps(vx0123, vbeta);
158 vx4567 = _mm_mul_ps(vx4567, vbeta);
159 vx89AB = _mm_mul_ps(vx89AB, vbeta);
160 vxCDEF = _mm_mul_ps(vxCDEF, vbeta);
161 vxGHIJ = _mm_mul_ps(vxGHIJ, vbeta);
162 vxKLMN = _mm_mul_ps(vxKLMN, vbeta);
163
164 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
165 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
166 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
167 const __m128 vyCDEF = _mm_blendv_ps(vxCDEF, veCDEF, vxCDEF);
168 const __m128 vyGHIJ = _mm_blendv_ps(vxGHIJ, veGHIJ, vxGHIJ);
169 const __m128 vyKLMN = _mm_blendv_ps(vxKLMN, veKLMN, vxKLMN);
170
171 _mm_storeu_ps(y, vy0123);
172 _mm_storeu_ps(y + 4, vy4567);
173 _mm_storeu_ps(y + 8, vy89AB);
174 _mm_storeu_ps(y + 12, vyCDEF);
175 _mm_storeu_ps(y + 16, vyGHIJ);
176 _mm_storeu_ps(y + 20, vyKLMN);
177 y += 24;
178 }
179 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
180 __m128 vx = _mm_loadu_ps(x);
181 x += 4;
182
183 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
184
185 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
186 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
187 vn = _mm_sub_ps(vn, vmagic_bias);
188
189 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
190 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
191
192 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
193 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
194 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
195 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
196 vp = _mm_mul_ps(vp, vt);
197
198 vt = _mm_mul_ps(vt, vs);
199 vs = _mm_sub_ps(vs, vone);
200 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
201 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
202
203 vx = _mm_mul_ps(vx, vbeta);
204 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
205
206 _mm_storeu_ps(y, vy);
207 y += 4;
208 }
209 if XNN_UNLIKELY(n != 0) {
210 __m128 vx = _mm_loadu_ps(x);
211
212 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
213
214 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
215 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
216 vn = _mm_sub_ps(vn, vmagic_bias);
217
218 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
219 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
220
221 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
222 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
223 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
224 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
225 vp = _mm_mul_ps(vp, vt);
226
227 vt = _mm_mul_ps(vt, vs);
228 vs = _mm_sub_ps(vs, vone);
229 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
230 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
231
232 vx = _mm_mul_ps(vx, vbeta);
233 __m128 vy = _mm_blendv_ps(vx, ve, vx);
234
235 if (n & (2 * sizeof(float))) {
236 _mm_storel_pi((__m64*) y, vy);
237 vy = _mm_movehl_ps(vy, vy);
238 y += 2;
239 }
240 if (n & (1 * sizeof(float))) {
241 _mm_store_ss(y, vy);
242 }
243 }
244 }
245