1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_velu_ukernel__avx2_rr1_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__avx2_rr1_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)])
23 {
24 assert(n % sizeof(float) == 0);
25
26 const __m256 vprescale = _mm256_load_ps(params->avx2_rr1_p6.prescale);
27 const __m256 valpha = _mm256_load_ps(params->avx2_rr1_p6.alpha);
28 const __m256 vbeta = _mm256_load_ps(params->avx2_rr1_p6.beta);
29 const __m256 vsat_cutoff = _mm256_load_ps(params->avx2_rr1_p6.sat_cutoff);
30 const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p6.magic_bias);
31 const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p6.log2e);
32 const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p6.minus_ln2);
33 const __m256 vc6 = _mm256_load_ps(params->avx2_rr1_p6.c6);
34 const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p6.c5);
35 const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p6.c4);
36 const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p6.c3);
37 const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p6.c2);
38
39 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
40 __m256 vx0 = _mm256_loadu_ps(x);
41 __m256 vx1 = _mm256_loadu_ps(x + 8);
42 x += 16;
43
44 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
45 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
46
47 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
48 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
49
50 __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23));
51 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
52 __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23));
53 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
54
55 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
56 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
57
58 __m256 vp0 = _mm256_fmadd_ps(vc6, vt0, vc5);
59 __m256 vp1 = _mm256_fmadd_ps(vc6, vt1, vc5);
60
61 vp0 = _mm256_fmadd_ps(vp0, vt0, vc4);
62 vp1 = _mm256_fmadd_ps(vp1, vt1, vc4);
63
64 vp0 = _mm256_fmadd_ps(vp0, vt0, vc3);
65 vp1 = _mm256_fmadd_ps(vp1, vt1, vc3);
66
67 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
68 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
69
70 vp0 = _mm256_mul_ps(vp0, vt0);
71 vt0 = _mm256_mul_ps(vt0, vs0);
72 vp1 = _mm256_mul_ps(vp1, vt1);
73 vt1 = _mm256_mul_ps(vt1, vs1);
74
75 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
76 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
77 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
78 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
79
80 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
81 vx0 = _mm256_mul_ps(vx0, vbeta);
82 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
83 vx1 = _mm256_mul_ps(vx1, vbeta);
84
85 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
86 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
87
88 _mm256_storeu_ps(y, vy0);
89 _mm256_storeu_ps(y + 8, vy1);
90 y += 16;
91 }
92 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
93 __m256 vx = _mm256_loadu_ps(x);
94 x += 8;
95
96 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
97
98 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
99 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
100 vn = _mm256_sub_ps(vn, vmagic_bias);
101
102 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
103
104 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
105 vp = _mm256_fmadd_ps(vp, vt, vc4);
106 vp = _mm256_fmadd_ps(vp, vt, vc3);
107 vp = _mm256_fmadd_ps(vp, vt, vc2);
108 vp = _mm256_mul_ps(vp, vt);
109
110 vt = _mm256_mul_ps(vt, vs);
111 vs = _mm256_fmsub_ps(vs, valpha, valpha);
112 vp = _mm256_fmadd_ps(vp, vt, vt);
113 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
114
115 vx = _mm256_mul_ps(vx, vbeta);
116 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
117
118 _mm256_storeu_ps(y, vy);
119 y += 8;
120 }
121 if XNN_UNLIKELY(n != 0) {
122 assert(n >= 1 * sizeof(float));
123 assert(n <= 7 * sizeof(float));
124 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2_rr1_p6.mask_table[7] - n));
125
126 __m256 vx = _mm256_maskload_ps(x, vmask);
127
128 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
129
130 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
131 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
132 vn = _mm256_sub_ps(vn, vmagic_bias);
133
134 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
135
136 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
137 vp = _mm256_fmadd_ps(vp, vt, vc4);
138 vp = _mm256_fmadd_ps(vp, vt, vc3);
139 vp = _mm256_fmadd_ps(vp, vt, vc2);
140 vp = _mm256_mul_ps(vp, vt);
141
142 vt = _mm256_mul_ps(vt, vs);
143 vs = _mm256_fmsub_ps(vs, valpha, valpha);
144 vp = _mm256_fmadd_ps(vp, vt, vt);
145 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
146
147 vx = _mm256_mul_ps(vx, vbeta);
148 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
149
150 __m128 vy_lo = _mm256_castps256_ps128(vy);
151 if (n & (4 * sizeof(float))) {
152 _mm_storeu_ps(y, vy_lo);
153 vy_lo = _mm256_extractf128_ps(vy, 1);
154 y += 4;
155 }
156 if (n & (2 * sizeof(float))) {
157 _mm_storel_pi((__m64*) y, vy_lo);
158 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
159 y += 2;
160 }
161 if (n & (1 * sizeof(float))) {
162 _mm_store_ss(y, vy_lo);
163 }
164 }
165 }
166