1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx2-rr1-lut8-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__avx2_rr1_lut8_p4_perm_x16(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_lut8_p4_perm_x16(
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(0x1.800000p20f);
34 const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
35 const __m256i vtable = _mm256_set_epi32(
36 0x3F7AC0C7, 0x3F7744FD, 0x3F75672A, 0x3F7504F3, 0x3F75FED7, 0x3F7837F0, 0x3F7B95C2, 0x3F800000);
37 const __m256 vminus_ln2 = _mm256_set1_ps(-0x1.62E43p-1f);
38 const __m256 vc4 = _mm256_set1_ps(0x1.5558ECp-5f);
39 const __m256 vc3 = _mm256_set1_ps(0x1.555C20p-3f);
40 const __m256 vc2 = _mm256_set1_ps(0x1.000000p-1f);
41
42 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
43 __m256 vx0 = _mm256_loadu_ps(x);
44 __m256 vx1 = _mm256_loadu_ps(x + 8);
45 x += 16;
46
47 const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
48 const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
49
50 __m256 vn0 = _mm256_fmadd_ps(vz0, vlog2e, vmagic_bias);
51 __m256 vn1 = _mm256_fmadd_ps(vz1, vlog2e, vmagic_bias);
52
53 const __m256i ven0 = _mm256_slli_epi32(_mm256_castps_si256(vn0), 20);
54 const __m256i vl0 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn0));
55 vn0 = _mm256_sub_ps(vn0, vmagic_bias);
56 const __m256i ven1 = _mm256_slli_epi32(_mm256_castps_si256(vn1), 20);
57 const __m256i vl1 = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn1));
58 vn1 = _mm256_sub_ps(vn1, vmagic_bias);
59
60 __m256 vs0 = _mm256_castsi256_ps(_mm256_add_epi32(vl0, ven0));
61 __m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2, vz0);
62 __m256 vs1 = _mm256_castsi256_ps(_mm256_add_epi32(vl1, ven1));
63 __m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2, vz1);
64
65 __m256 vp0 = _mm256_fmadd_ps(vc4, vt0, vc3);
66 __m256 vp1 = _mm256_fmadd_ps(vc4, vt1, vc3);
67
68 vp0 = _mm256_fmadd_ps(vp0, vt0, vc2);
69 vp1 = _mm256_fmadd_ps(vp1, vt1, vc2);
70
71 vp0 = _mm256_mul_ps(vp0, vt0);
72 vt0 = _mm256_mul_ps(vt0, vs0);
73 vp1 = _mm256_mul_ps(vp1, vt1);
74 vt1 = _mm256_mul_ps(vt1, vs1);
75
76 vs0 = _mm256_fmsub_ps(vs0, valpha, valpha);
77 vp0 = _mm256_fmadd_ps(vp0, vt0, vt0);
78 vs1 = _mm256_fmsub_ps(vs1, valpha, valpha);
79 vp1 = _mm256_fmadd_ps(vp1, vt1, vt1);
80
81 const __m256 ve0 = _mm256_fmadd_ps(vp0, valpha, vs0);
82 vx0 = _mm256_mul_ps(vx0, vbeta);
83 const __m256 ve1 = _mm256_fmadd_ps(vp1, valpha, vs1);
84 vx1 = _mm256_mul_ps(vx1, vbeta);
85
86 const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
87 const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
88
89 _mm256_storeu_ps(y, vy0);
90 _mm256_storeu_ps(y + 8, vy1);
91 y += 16;
92 }
93 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
94 __m256 vx = _mm256_loadu_ps(x);
95 x += 8;
96
97 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
98
99 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
100 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
101 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
102 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
103 vn = _mm256_sub_ps(vn, vmagic_bias);
104
105 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
106
107 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
108 vp = _mm256_fmadd_ps(vp, vt, vc2);
109 vp = _mm256_mul_ps(vp, vt);
110
111 vt = _mm256_mul_ps(vt, vs);
112 vs = _mm256_fmsub_ps(vs, valpha, valpha);
113 vp = _mm256_fmadd_ps(vp, vt, vt);
114 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
115
116 vx = _mm256_mul_ps(vx, vbeta);
117 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
118
119 _mm256_storeu_ps(y, vy);
120 y += 8;
121 }
122 if XNN_UNLIKELY(n != 0) {
123 assert(n >= 1 * sizeof(float));
124 assert(n <= 7 * sizeof(float));
125 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n));
126
127 __m256 vx = _mm256_maskload_ps(x, vmask);
128
129 const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));
130
131 __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
132 const __m256i ven = _mm256_slli_epi32(_mm256_castps_si256(vn), 20);
133 const __m256i vl = _mm256_permutevar8x32_epi32(vtable, _mm256_castps_si256(vn));
134 __m256 vs = _mm256_castsi256_ps(_mm256_add_epi32(vl, ven));
135 vn = _mm256_sub_ps(vn, vmagic_bias);
136
137 __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);
138
139 __m256 vp = _mm256_fmadd_ps(vc4, vt, vc3);
140 vp = _mm256_fmadd_ps(vp, vt, vc2);
141 vp = _mm256_mul_ps(vp, vt);
142
143 vt = _mm256_mul_ps(vt, vs);
144 vs = _mm256_fmsub_ps(vs, valpha, valpha);
145 vp = _mm256_fmadd_ps(vp, vt, vt);
146 const __m256 ve = _mm256_fmadd_ps(vp, valpha, vs);
147
148 vx = _mm256_mul_ps(vx, vbeta);
149 const __m256 vy = _mm256_blendv_ps(vx, ve, vx);
150
151 // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
152 __m128 vy_lo = _mm256_castps256_ps128(vy);
153 if (n & (4 * sizeof(float))) {
154 _mm_storeu_ps(y, vy_lo);
155 vy_lo = _mm256_extractf128_ps(vy, 1);
156 y += 4;
157 }
158 if (n & (2 * sizeof(float))) {
159 _mm_storel_pi((__m64*) y, vy_lo);
160 vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
161 y += 2;
162 }
163 if (n & (1 * sizeof(float))) {
164 _mm_store_ss(y, vy_lo);
165 }
166 }
167 }
168