1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx-rr2-lut4-p4-perm.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_lut4_p4_perm_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_lut4_p4_perm_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(0x1.8003F8p21f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256 vindex_mask = _mm256_castsi256_ps(_mm256_set1_epi32(0x3));
36 const __m256 vtable = _mm256_set_ps(
37 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f,
38 0x1.AE89FAp+0f, 0x1.6A09E6p+0f, 0x1.306FE0p+0f, 0x1.000000p+0f);
39 const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E400p-1f);
40 const __m256 vminus_ln2_lo = _mm256_set1_ps(-0x1.7F7D1Cp-20f);
41 const __m256 vc4 = _mm256_set1_ps(0x1.554F9Ap-5f);
42 const __m256 vc3 = _mm256_set1_ps(0x1.557082p-3f);
43 const __m256 vc2 = _mm256_set1_ps(0x1.000002p-1f);
44 const __m256 vone = _mm256_set1_ps(1.0f);
45
46 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
47 __m256 vx = _mm256_loadu_ps(x);
48 x += 8;
49
50 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
51
52 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
53 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
54 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
55 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
56 vn = _mm256_sub_ps(vn, vmagic_bias);
57 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
58
59 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
60 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
61 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
62 __m256 vs = _mm256_mul_ps(vl, ven);
63
64 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
65 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
66 vp = _mm256_mul_ps(vp, vt);
67
68 vt = _mm256_mul_ps(vt, vs);
69 vs = _mm256_sub_ps(vs, vone);
70 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
71
72 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
73 vx = _mm256_mul_ps(vx, vbeta);
74 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
75
76 _mm256_storeu_ps(y, vy);
77 y += 8;
78 }
79 if XNN_UNLIKELY(n != 0) {
80 assert(n >= 1 * sizeof(float));
81 assert(n <= 7 * sizeof(float));
82 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
83
84 __m256 vx = _mm256_maskload_ps(x, vmask);
85
86 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
87
88 __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
89 __m256 ven = _mm256_andnot_ps(vindex_mask, vn);
90 const __m256 vl = _mm256_permutevar_ps(vtable, _mm256_castps_si256(vn));
91 const __m128 ven_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(ven)), 21));
92 vn = _mm256_sub_ps(vn, vmagic_bias);
93 const __m128 ven_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(ven, 1)), 21));
94
95 __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
96 ven = _mm256_insertf128_ps(_mm256_castps128_ps256(ven_lo), ven_hi, 1);
97 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);
98 __m256 vs = _mm256_mul_ps(vl, ven);
99
100 __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc4, vt), vc3);
101 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
102 vp = _mm256_mul_ps(vp, vt);
103
104 vt = _mm256_mul_ps(vt, vs);
105 vs = _mm256_sub_ps(vs, vone);
106 vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);
107
108 const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
109 vx = _mm256_mul_ps(vx, vbeta);
110 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
111
112 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
113 __m128 vy_lo = _mm256_castps256_ps128(vy);
114 if (n & (4 * sizeof(float))) {
115 _mm_storeu_ps(y, vy_lo);
116 vy_lo = _mm256_extractf128_ps(vy, 1);
117 y += 4;
118 }
119 if (n & (2 * sizeof(float))) {
120 _mm_storel_pi((__m64*) y, vy_lo);
121 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
122 y += 2;
123 }
124 if (n & (1 * sizeof(float))) {
125 _mm_store_ss(y, vy_lo);
126 }
127 }
128 }
129