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
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__avx2_rr1_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__avx2_rr1_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 = _mm256_set1_ps(-0x1.62E43p-1f);
36 const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
37 const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
38 const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
39 const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
41
42 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
43 __m256 vx = _mm256_loadu_ps(x);
44 x += 8;
45
46 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
47
48 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
49 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
50 vn = _mm256_sub_ps(vn, vmagic_bias);
51
52 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
53
54 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
55 vp = _mm256_fmadd_ps(vp, vt, vc4);
56 vp = _mm256_fmadd_ps(vp, vt, vc3);
57 vp = _mm256_fmadd_ps(vp, vt, vc2);
58 vp = _mm256_mul_ps(vp, vt);
59
60 vt = _mm256_mul_ps(vt, vs);
61 vs = _mm256_fmsub_ps(vs, valpha, valpha);
62 vp = _mm256_fmadd_ps(vp, vt, vt);
63 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
64
65 vx = _mm256_mul_ps(vx, vbeta);
66 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
67
68 _mm256_storeu_ps(y, vy);
69 y += 8;
70 }
71 if XNN_UNLIKELY(n != 0) {
72 assert(n >= 1 * sizeof(float));
73 assert(n <= 7 * sizeof(float));
74 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
75
76 __m256 vx = _mm256_maskload_ps(x, vmask);
77
78 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
79
80 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
81 __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
82 vn = _mm256_sub_ps(vn, vmagic_bias);
83
84 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
85
86 __m256 vp = _mm256_fmadd_ps(vc6, vt, vc5);
87 vp = _mm256_fmadd_ps(vp, vt, vc4);
88 vp = _mm256_fmadd_ps(vp, vt, vc3);
89 vp = _mm256_fmadd_ps(vp, vt, vc2);
90 vp = _mm256_mul_ps(vp, vt);
91
92 vt = _mm256_mul_ps(vt, vs);
93 vs = _mm256_fmsub_ps(vs, valpha, valpha);
94 vp = _mm256_fmadd_ps(vp, vt, vt);
95 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
96
97 vx = _mm256_mul_ps(vx, vbeta);
98 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
99
100 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
101 __m128 vy_lo = _mm256_castps256_ps128(vy);
102 if (n & (4 * sizeof(float))) {
103 _mm_storeu_ps(y, vy_lo);
104 vy_lo = _mm256_extractf128_ps(vy, 1);
105 y += 4;
106 }
107 if (n & (2 * sizeof(float))) {
108 _mm_storel_pi((__m64*) y, vy_lo);
109 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
110 y += 2;
111 }
112 if (n & (1 * sizeof(float))) {
113 _mm_store_ss(y, vy_lo);
114 }
115 }
116 }
117