1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-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 <immintrin.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/vunary.h>
16
17
xnn_f32_velu_ukernel__avx_rr2_p6_x8(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__avx_rr2_p6_x8(
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->avx_rr2_p6.prescale);
27 const __m256 valpha = _mm256_load_ps(params->avx_rr2_p6.alpha);
28 const __m256 vbeta = _mm256_load_ps(params->avx_rr2_p6.beta);
29 const __m256 vsat_cutoff = _mm256_load_ps(params->avx_rr2_p6.sat_cutoff);
30 const __m256 vmagic_bias = _mm256_load_ps(params->avx_rr2_p6.magic_bias);
31 const __m256 vlog2e = _mm256_load_ps(params->avx_rr2_p6.log2e);
32 const __m256 vminus_ln2_hi = _mm256_load_ps(params->avx_rr2_p6.minus_ln2_hi);
33 const __m256 vminus_ln2_lo = _mm256_load_ps(params->avx_rr2_p6.minus_ln2_lo);
34 const __m256 vc6 = _mm256_load_ps(params->avx_rr2_p6.c6);
35 const __m256 vc5 = _mm256_load_ps(params->avx_rr2_p6.c5);
36 const __m256 vc4 = _mm256_load_ps(params->avx_rr2_p6.c4);
37 const __m256 vc3 = _mm256_load_ps(params->avx_rr2_p6.c3);
38 const __m256 vc2 = _mm256_load_ps(params->avx_rr2_p6.c2);
39 const __m256 vone = _mm256_load_ps(params->avx_rr2_p6.one);
40
41 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
42 __m256 vx = _mm256_loadu_ps(x);
43 x += 8;
44
45 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
46
47 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
48 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
49 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
50 vn = _mm256_sub_ps(vn, vmagic_bias);
51
52 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
53 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
54 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
55
56 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
57 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
58 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
59 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
60 vp = _mm256_mul_ps(vp, vt);
61
62 vt = _mm256_mul_ps(vt, vs);
63 vs = _mm256_sub_ps(vs, vone);
64 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
65
66 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
67 vx = _mm256_mul_ps(vx, vbeta);
68 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
69
70 _mm256_storeu_ps(y, vy);
71 y += 8;
72 }
73 if XNN_UNLIKELY(n != 0) {
74 assert(n >= 1 * sizeof(float));
75 assert(n <= 7 * sizeof(float));
76 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx_rr2_p6.mask_table[7] - n));
77
78 __m256 vx = _mm256_maskload_ps(x, vmask);
79
80 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
81
82 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
83 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
84 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
85 vn = _mm256_sub_ps(vn, vmagic_bias);
86
87 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
88 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
89 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
90
91 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
92 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
93 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
94 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
95 vp = _mm256_mul_ps(vp, vt);
96
97 vt = _mm256_mul_ps(vt, vs);
98 vs = _mm256_sub_ps(vs, vone);
99 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
100
101 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
102 vx = _mm256_mul_ps(vx, vbeta);
103 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
104
105 __m128 vy_lo = _mm256_castps256_ps128(vy);
106 if (n & (4 * sizeof(float))) {
107 _mm_storeu_ps(y, vy_lo);
108 vy_lo = _mm256_extractf128_ps(vy, 1);
109 y += 4;
110 }
111 if (n & (2 * sizeof(float))) {
112 _mm_storel_pi((__m64*) y, vy_lo);
113 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
114 y += 2;
115 }
116 if (n & (1 * sizeof(float))) {
117 _mm_store_ss(y, vy_lo);
118 }
119 }
120 }
121