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_x48(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_x48(
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 >= 48 * sizeof(float); n -= 48 * 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     x += 48;
47 
48     const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale));
49     const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale));
50     const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale));
51 
52     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
53     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
54     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
55 
56     const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19);
57     const __m512i vl0 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn0), vtable);
58     const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19);
59     const __m512i vl1 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn1), vtable);
60     const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19);
61     const __m512i vl2 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn2), vtable);
62 
63     __m512 vs0 = _mm512_castsi512_ps(_mm512_add_epi32(vl0, ven0));
64     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
65     __m512 vs1 = _mm512_castsi512_ps(_mm512_add_epi32(vl1, ven1));
66     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
67     __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2));
68     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
69 
70     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
71     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
72     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
73 
74     __m512 vp0 = _mm512_fmadd_ps(vc3, vt0, vc2);
75     __m512 vp1 = _mm512_fmadd_ps(vc3, vt1, vc2);
76     __m512 vp2 = _mm512_fmadd_ps(vc3, vt2, vc2);
77 
78     vp0 = _mm512_mul_ps(vp0, vt0);
79     vt0 = _mm512_mul_ps(vt0, vs0);
80     vp1 = _mm512_mul_ps(vp1, vt1);
81     vt1 = _mm512_mul_ps(vt1, vs1);
82     vp2 = _mm512_mul_ps(vp2, vt2);
83     vt2 = _mm512_mul_ps(vt2, vs2);
84 
85     vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
86     vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
87     vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
88 
89     vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
90     vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
91     vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
92 
93     const __m512 vzero = _mm512_setzero_ps();
94     __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
95     const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
96     __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
97     const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
98     __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
99     const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
100 
101     vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
102     vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
103     vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
104 
105     _mm512_storeu_ps(y, vy0);
106     _mm512_storeu_ps(y + 16, vy1);
107     _mm512_storeu_ps(y + 32, vy2);
108     y += 48;
109   }
110   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
111     __m512 vx = _mm512_loadu_ps(x);
112     x += 16;
113 
114     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
115     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
116 
117     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
118     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
119     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
120     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
121     vn = _mm512_sub_ps(vn, vmagic_bias);
122 
123     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
124 
125     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
126     vp = _mm512_mul_ps(vp, vt);
127 
128     vt = _mm512_mul_ps(vt, vs);
129     vs = _mm512_fmsub_ps(vs, valpha, valpha);
130     vp = _mm512_fmadd_ps(vp, vt, vt);
131     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
132 
133     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
134 
135     _mm512_storeu_ps(y, vy);
136     y += 16;
137   }
138   if XNN_UNLIKELY(n != 0) {
139     assert(n >= 1 * sizeof(float));
140     assert(n <= 15 * sizeof(float));
141     // Prepare mask for valid 32-bit elements (depends on n).
142     n >>= 2 /* log2(sizeof(float)) */;
143     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
144 
145     __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
146 
147     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
148     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
149 
150     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
151     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
152     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
153     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
154     vn = _mm512_sub_ps(vn, vmagic_bias);
155 
156     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
157 
158     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
159     vp = _mm512_mul_ps(vp, vt);
160 
161     vt = _mm512_mul_ps(vt, vs);
162     vs = _mm512_fmsub_ps(vs, valpha, valpha);
163     vp = _mm512_fmadd_ps(vp, vt, vt);
164     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
165 
166     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
167 
168     _mm512_mask_storeu_ps(y, vmask, vy);
169   }
170 }
171