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_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__sse41_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_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 >= 12 * sizeof(float); n -= 12 * 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 x += 12;
49
50 const __m128 vz0123 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx0123, vprescale));
51 const __m128 vz4567 = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx4567, vprescale));
52 const __m128 vz89AB = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx89AB, vprescale));
53
54 __m128 vn0123 = _mm_add_ps(_mm_mul_ps(vz0123, vlog2e), vmagic_bias);
55 __m128 vn4567 = _mm_add_ps(_mm_mul_ps(vz4567, vlog2e), vmagic_bias);
56 __m128 vn89AB = _mm_add_ps(_mm_mul_ps(vz89AB, vlog2e), vmagic_bias);
57
58 __m128 vs0123 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn0123), 23));
59 __m128 vs4567 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn4567), 23));
60 __m128 vs89AB = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn89AB), 23));
61
62 vn0123 = _mm_sub_ps(vn0123, vmagic_bias);
63 vn4567 = _mm_sub_ps(vn4567, vmagic_bias);
64 vn89AB = _mm_sub_ps(vn89AB, vmagic_bias);
65
66 __m128 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_hi), vz0123);
67 __m128 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_hi), vz4567);
68 __m128 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_hi), vz89AB);
69
70 vt0123 = _mm_add_ps(_mm_mul_ps(vn0123, vminus_ln2_lo), vt0123);
71 vt4567 = _mm_add_ps(_mm_mul_ps(vn4567, vminus_ln2_lo), vt4567);
72 vt89AB = _mm_add_ps(_mm_mul_ps(vn89AB, vminus_ln2_lo), vt89AB);
73
74 __m128 vp0123 = _mm_add_ps(_mm_mul_ps(vc6, vt0123), vc5);
75 __m128 vp4567 = _mm_add_ps(_mm_mul_ps(vc6, vt4567), vc5);
76 __m128 vp89AB = _mm_add_ps(_mm_mul_ps(vc6, vt89AB), vc5);
77
78 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc4);
79 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc4);
80 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc4);
81
82 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc3);
83 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc3);
84 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc3);
85
86 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vc2);
87 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vc2);
88 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vc2);
89
90 vp0123 = _mm_mul_ps(vp0123, vt0123);
91 vp4567 = _mm_mul_ps(vp4567, vt4567);
92 vp89AB = _mm_mul_ps(vp89AB, vt89AB);
93
94 vt0123 = _mm_mul_ps(vt0123, vs0123);
95 vs0123 = _mm_sub_ps(vs0123, vone);
96 vt4567 = _mm_mul_ps(vt4567, vs4567);
97 vs4567 = _mm_sub_ps(vs4567, vone);
98 vt89AB = _mm_mul_ps(vt89AB, vs89AB);
99 vs89AB = _mm_sub_ps(vs89AB, vone);
100
101 vp0123 = _mm_add_ps(_mm_mul_ps(vp0123, vt0123), vt0123);
102 vp4567 = _mm_add_ps(_mm_mul_ps(vp4567, vt4567), vt4567);
103 vp89AB = _mm_add_ps(_mm_mul_ps(vp89AB, vt89AB), vt89AB);
104
105 const __m128 ve0123 = _mm_mul_ps(_mm_add_ps(vp0123, vs0123), valpha);
106 const __m128 ve4567 = _mm_mul_ps(_mm_add_ps(vp4567, vs4567), valpha);
107 const __m128 ve89AB = _mm_mul_ps(_mm_add_ps(vp89AB, vs89AB), valpha);
108
109 vx0123 = _mm_mul_ps(vx0123, vbeta);
110 vx4567 = _mm_mul_ps(vx4567, vbeta);
111 vx89AB = _mm_mul_ps(vx89AB, vbeta);
112
113 const __m128 vy0123 = _mm_blendv_ps(vx0123, ve0123, vx0123);
114 const __m128 vy4567 = _mm_blendv_ps(vx4567, ve4567, vx4567);
115 const __m128 vy89AB = _mm_blendv_ps(vx89AB, ve89AB, vx89AB);
116
117 _mm_storeu_ps(y, vy0123);
118 _mm_storeu_ps(y + 4, vy4567);
119 _mm_storeu_ps(y + 8, vy89AB);
120 y += 12;
121 }
122 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
123 __m128 vx = _mm_loadu_ps(x);
124 x += 4;
125
126 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
127
128 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
129 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
130 vn = _mm_sub_ps(vn, vmagic_bias);
131
132 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
133 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
134
135 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
136 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
137 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
138 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
139 vp = _mm_mul_ps(vp, vt);
140
141 vt = _mm_mul_ps(vt, vs);
142 vs = _mm_sub_ps(vs, vone);
143 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
144 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
145
146 vx = _mm_mul_ps(vx, vbeta);
147 const __m128 vy = _mm_blendv_ps(vx, ve, vx);
148
149 _mm_storeu_ps(y, vy);
150 y += 4;
151 }
152 if XNN_UNLIKELY(n != 0) {
153 __m128 vx = _mm_loadu_ps(x);
154
155 const __m128 vz = _mm_max_ps(vsat_cutoff, _mm_mul_ps(vx, vprescale));
156
157 __m128 vn = _mm_add_ps(_mm_mul_ps(vz, vlog2e), vmagic_bias);
158 __m128 vs = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(vn), 23));
159 vn = _mm_sub_ps(vn, vmagic_bias);
160
161 __m128 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_hi), vz);
162 vt = _mm_add_ps(_mm_mul_ps(vn, vminus_ln2_lo), vt);
163
164 __m128 vp = _mm_add_ps(_mm_mul_ps(vc6, vt), vc5);
165 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc4);
166 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc3);
167 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vc2);
168 vp = _mm_mul_ps(vp, vt);
169
170 vt = _mm_mul_ps(vt, vs);
171 vs = _mm_sub_ps(vs, vone);
172 vp = _mm_add_ps(_mm_mul_ps(vp, vt), vt);
173 const __m128 ve = _mm_mul_ps(_mm_add_ps(vp, vs), valpha);
174
175 vx = _mm_mul_ps(vx, vbeta);
176 __m128 vy = _mm_blendv_ps(vx, ve, vx);
177
178 if (n & (2 * sizeof(float))) {
179 _mm_storel_pi((__m64*) y, vy);
180 vy = _mm_movehl_ps(vy, vy);
181 y += 2;
182 }
183 if (n & (1 * sizeof(float))) {
184 _mm_store_ss(y, vy);
185 }
186 }
187 }
188