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
18 static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};
19
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)])20 void xnn_f32_velu_ukernel__avx_rr2_p6_x8(
21 size_t n,
22 const float* x,
23 float* y,
24 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
25 {
26 assert(n % sizeof(float) == 0);
27
28 const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
29 const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
30 const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);
31
32 const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
33 const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E440p-1f);
36 const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.0105C6p-21f);
37 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
38 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
39 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
40 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
41 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
42 const __m256 vone = _mm256_set1_ps(1.0f);
43
44 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
45 __m256 vx = _mm256_loadu_ps(x);
46 x += 8;
47
48 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
49
50 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
51 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
52 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
53 vn = _mm256_sub_ps(vn, vmagic_bias);
54
55 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
56 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
57 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
58
59 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
60 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
61 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
62 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
63 vp = _mm256_mul_ps(vp, vt);
64
65 vt = _mm256_mul_ps(vt, vs);
66 vs = _mm256_sub_ps(vs, vone);
67 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
68
69 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
70 vx = _mm256_mul_ps(vx, vbeta);
71 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
72
73 _mm256_storeu_ps(y, vy);
74 y += 8;
75 }
76 if XNN_UNLIKELY(n != 0) {
77 assert(n >= 1 * sizeof(float));
78 assert(n <= 7 * sizeof(float));
79 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
80
81 __m256 vx = _mm256_maskload_ps(x, vmask);
82
83 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
84
85 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
86 const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
87 const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
88 vn = _mm256_sub_ps(vn, vmagic_bias);
89
90 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
91 __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
92 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
93
94 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
95 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
96 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
97 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
98 vp = _mm256_mul_ps(vp, vt);
99
100 vt = _mm256_mul_ps(vt, vs);
101 vs = _mm256_sub_ps(vs, vone);
102 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
103
104 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
105 vx = _mm256_mul_ps(vx, vbeta);
106 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
107
108 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
109 __m128 vy_lo = _mm256_castps256_ps128(vy);
110 if (n & (4 * sizeof(float))) {
111 _mm_storeu_ps(y, vy_lo);
112 vy_lo = _mm256_extractf128_ps(vy, 1);
113 y += 4;
114 }
115 if (n & (2 * sizeof(float))) {
116 _mm_storel_pi((__m64*) y, vy_lo);
117 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
118 y += 2;
119 }
120 if (n & (1 * sizeof(float))) {
121 _mm_store_ss(y, vy_lo);
122 }
123 }
124 }
125