1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx512f-rr1-lut16-p3-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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x80(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const __m512 vprescale = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.prescale));
29 const __m512 valpha = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.alpha));
30 const __m512 vbeta = _mm512_broadcast_f32x4(_mm_load_ps(params->sse.beta));
31
32 const __m512 vsat_cutoff = _mm512_set1_ps(-0x1.154246p+4f);
33 const __m512 vmagic_bias = _mm512_set1_ps(0x1.800000p19f);
34 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
35 const __m512i vtable = _mm512_set_epi32(
36 0x3F7D257D, 0x3F7AC0C7, 0x3F78CCDF, 0x3F7744FD, 0x3F76248C, 0x3F75672A, 0x3F7508A4, 0x3F7504F3,
37 0x3F75583F, 0x3F75FED7, 0x3F76F532, 0x3F7837F0, 0x3F79C3D3, 0x3F7B95C2, 0x3F7DAAC3, 0x3F800000);
38 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
39 const __m512 vc3 = _mm512_set1_ps(0x1.55561Cp-3f);
40 const __m512 vc2 = _mm512_set1_ps(0x1.0001ECp-1f);
41
42 for (; n >= 80 * sizeof(float); n -= 80 * sizeof(float)) {
43 __m512 vx0 = _mm512_loadu_ps(x);
44 __m512 vx1 = _mm512_loadu_ps(x + 16);
45 __m512 vx2 = _mm512_loadu_ps(x + 32);
46 __m512 vx3 = _mm512_loadu_ps(x + 48);
47 __m512 vx4 = _mm512_loadu_ps(x + 64);
48 x += 80;
49
50 const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale));
51 const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale));
52 const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale));
53 const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale));
54 const __m512 vz4 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx4, vprescale));
55
56 __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
57 __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
58 __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
59 __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
60 __m512 vn4 = _mm512_fmadd_ps(vz4, vlog2e, vmagic_bias);
61
62 const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19);
63 const __m512i vl0 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn0), vtable);
64 const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19);
65 const __m512i vl1 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn1), vtable);
66 const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19);
67 const __m512i vl2 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn2), vtable);
68 const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19);
69 const __m512i vl3 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn3), vtable);
70 const __m512i ven4 = _mm512_slli_epi32(_mm512_castps_si512(vn4), 19);
71 const __m512i vl4 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn4), vtable);
72
73 __m512 vs0 = _mm512_castsi512_ps(_mm512_add_epi32(vl0, ven0));
74 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
75 __m512 vs1 = _mm512_castsi512_ps(_mm512_add_epi32(vl1, ven1));
76 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
77 __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2));
78 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
79 __m512 vs3 = _mm512_castsi512_ps(_mm512_add_epi32(vl3, ven3));
80 vn3 = _mm512_sub_ps(vn3, vmagic_bias);
81 __m512 vs4 = _mm512_castsi512_ps(_mm512_add_epi32(vl4, ven4));
82 vn4 = _mm512_sub_ps(vn4, vmagic_bias);
83
84 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
85 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
86 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
87 __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
88 __m512 vt4 = _mm512_fmadd_ps(vn4, vminus_ln2, vz4);
89
90 __m512 vp0 = _mm512_fmadd_ps(vc3, vt0, vc2);
91 __m512 vp1 = _mm512_fmadd_ps(vc3, vt1, vc2);
92 __m512 vp2 = _mm512_fmadd_ps(vc3, vt2, vc2);
93 __m512 vp3 = _mm512_fmadd_ps(vc3, vt3, vc2);
94 __m512 vp4 = _mm512_fmadd_ps(vc3, vt4, vc2);
95
96 vp0 = _mm512_mul_ps(vp0, vt0);
97 vt0 = _mm512_mul_ps(vt0, vs0);
98 vp1 = _mm512_mul_ps(vp1, vt1);
99 vt1 = _mm512_mul_ps(vt1, vs1);
100 vp2 = _mm512_mul_ps(vp2, vt2);
101 vt2 = _mm512_mul_ps(vt2, vs2);
102 vp3 = _mm512_mul_ps(vp3, vt3);
103 vt3 = _mm512_mul_ps(vt3, vs3);
104 vp4 = _mm512_mul_ps(vp4, vt4);
105 vt4 = _mm512_mul_ps(vt4, vs4);
106
107 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
108 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
109 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
110 vs3 = _mm512_fmsub_ps(vs3, valpha, valpha);
111 vs4 = _mm512_fmsub_ps(vs4, valpha, valpha);
112
113 vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
114 vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
115 vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
116 vp3 = _mm512_fmadd_ps(vp3, vt3, vt3);
117 vp4 = _mm512_fmadd_ps(vp4, vt4, vt4);
118
119 const __m512 vzero = _mm512_setzero_ps();
120 __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
121 const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
122 __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
123 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
124 __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
125 const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
126 __m512 vy3 = _mm512_fmadd_ps(vp3, valpha, vs3);
127 const __mmask16 vsign3 = _mm512_cmp_ps_mask(vx3, vzero, _CMP_NLT_US);
128 __m512 vy4 = _mm512_fmadd_ps(vp4, valpha, vs4);
129 const __mmask16 vsign4 = _mm512_cmp_ps_mask(vx4, vzero, _CMP_NLT_US);
130
131 vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
132 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
133 vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
134 vy3 = _mm512_mask_mul_ps(vy3, vsign3, vx3, vbeta);
135 vy4 = _mm512_mask_mul_ps(vy4, vsign4, vx4, vbeta);
136
137 _mm512_storeu_ps(y, vy0);
138 _mm512_storeu_ps(y + 16, vy1);
139 _mm512_storeu_ps(y + 32, vy2);
140 _mm512_storeu_ps(y + 48, vy3);
141 _mm512_storeu_ps(y + 64, vy4);
142 y += 80;
143 }
144 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
145 __m512 vx = _mm512_loadu_ps(x);
146 x += 16;
147
148 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
149 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
150
151 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
152 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
153 const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
154 __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
155 vn = _mm512_sub_ps(vn, vmagic_bias);
156
157 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
158
159 __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
160 vp = _mm512_mul_ps(vp, vt);
161
162 vt = _mm512_mul_ps(vt, vs);
163 vs = _mm512_fmsub_ps(vs, valpha, valpha);
164 vp = _mm512_fmadd_ps(vp, vt, vt);
165 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
166
167 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
168
169 _mm512_storeu_ps(y, vy);
170 y += 16;
171 }
172 if XNN_UNLIKELY(n != 0) {
173 assert(n >= 1 * sizeof(float));
174 assert(n <= 15 * sizeof(float));
175 // Prepare mask for valid 32-bit elements (depends on n).
176 n >>= 2 /* log2(sizeof(float)) */;
177 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
178
179 __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
180
181 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
182 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
183
184 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
185 const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
186 const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
187 __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
188 vn = _mm512_sub_ps(vn, vmagic_bias);
189
190 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
191
192 __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
193 vp = _mm512_mul_ps(vp, vt);
194
195 vt = _mm512_mul_ps(vt, vs);
196 vs = _mm512_fmsub_ps(vs, valpha, valpha);
197 vp = _mm512_fmadd_ps(vp, vt, vt);
198 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
199
200 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
201
202 _mm512_mask_storeu_ps(y, vmask, vy);
203 }
204 }
205