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_x64(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_x64(
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 >= 64 * sizeof(float); n -= 64 * 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     x += 64;
48 
49     const __m512 vz0 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx0, vprescale));
50     const __m512 vz1 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx1, vprescale));
51     const __m512 vz2 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx2, vprescale));
52     const __m512 vz3 = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx3, vprescale));
53 
54     __m512 vn0 = _mm512_fmadd_ps(vz0, vlog2e, vmagic_bias);
55     __m512 vn1 = _mm512_fmadd_ps(vz1, vlog2e, vmagic_bias);
56     __m512 vn2 = _mm512_fmadd_ps(vz2, vlog2e, vmagic_bias);
57     __m512 vn3 = _mm512_fmadd_ps(vz3, vlog2e, vmagic_bias);
58 
59     const __m512i ven0 = _mm512_slli_epi32(_mm512_castps_si512(vn0), 19);
60     const __m512i vl0 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn0), vtable);
61     const __m512i ven1 = _mm512_slli_epi32(_mm512_castps_si512(vn1), 19);
62     const __m512i vl1 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn1), vtable);
63     const __m512i ven2 = _mm512_slli_epi32(_mm512_castps_si512(vn2), 19);
64     const __m512i vl2 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn2), vtable);
65     const __m512i ven3 = _mm512_slli_epi32(_mm512_castps_si512(vn3), 19);
66     const __m512i vl3 = _mm512_permutexvar_epi32(_mm512_castps_si512(vn3), vtable);
67 
68     __m512 vs0 = _mm512_castsi512_ps(_mm512_add_epi32(vl0, ven0));
69     vn0 = _mm512_sub_ps(vn0, vmagic_bias);
70     __m512 vs1 = _mm512_castsi512_ps(_mm512_add_epi32(vl1, ven1));
71     vn1 = _mm512_sub_ps(vn1, vmagic_bias);
72     __m512 vs2 = _mm512_castsi512_ps(_mm512_add_epi32(vl2, ven2));
73     vn2 = _mm512_sub_ps(vn2, vmagic_bias);
74     __m512 vs3 = _mm512_castsi512_ps(_mm512_add_epi32(vl3, ven3));
75     vn3 = _mm512_sub_ps(vn3, vmagic_bias);
76 
77     __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
78     __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
79     __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
80     __m512 vt3 = _mm512_fmadd_ps(vn3, vminus_ln2, vz3);
81 
82     __m512 vp0 = _mm512_fmadd_ps(vc3, vt0, vc2);
83     __m512 vp1 = _mm512_fmadd_ps(vc3, vt1, vc2);
84     __m512 vp2 = _mm512_fmadd_ps(vc3, vt2, vc2);
85     __m512 vp3 = _mm512_fmadd_ps(vc3, vt3, vc2);
86 
87     vp0 = _mm512_mul_ps(vp0, vt0);
88     vt0 = _mm512_mul_ps(vt0, vs0);
89     vp1 = _mm512_mul_ps(vp1, vt1);
90     vt1 = _mm512_mul_ps(vt1, vs1);
91     vp2 = _mm512_mul_ps(vp2, vt2);
92     vt2 = _mm512_mul_ps(vt2, vs2);
93     vp3 = _mm512_mul_ps(vp3, vt3);
94     vt3 = _mm512_mul_ps(vt3, vs3);
95 
96     vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
97     vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
98     vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
99     vs3 = _mm512_fmsub_ps(vs3, valpha, valpha);
100 
101     vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
102     vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
103     vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
104     vp3 = _mm512_fmadd_ps(vp3, vt3, vt3);
105 
106     const __m512 vzero = _mm512_setzero_ps();
107     __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
108     const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
109     __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
110     const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
111     __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
112     const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
113     __m512 vy3 = _mm512_fmadd_ps(vp3, valpha, vs3);
114     const __mmask16 vsign3 = _mm512_cmp_ps_mask(vx3, vzero, _CMP_NLT_US);
115 
116     vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
117     vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
118     vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
119     vy3 = _mm512_mask_mul_ps(vy3, vsign3, vx3, vbeta);
120 
121     _mm512_storeu_ps(y, vy0);
122     _mm512_storeu_ps(y + 16, vy1);
123     _mm512_storeu_ps(y + 32, vy2);
124     _mm512_storeu_ps(y + 48, vy3);
125     y += 64;
126   }
127   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
128     __m512 vx = _mm512_loadu_ps(x);
129     x += 16;
130 
131     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
132     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
133 
134     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
135     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
136     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
137     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
138     vn = _mm512_sub_ps(vn, vmagic_bias);
139 
140     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
141 
142     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
143     vp = _mm512_mul_ps(vp, vt);
144 
145     vt = _mm512_mul_ps(vt, vs);
146     vs = _mm512_fmsub_ps(vs, valpha, valpha);
147     vp = _mm512_fmadd_ps(vp, vt, vt);
148     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
149 
150     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
151 
152     _mm512_storeu_ps(y, vy);
153     y += 16;
154   }
155   if XNN_UNLIKELY(n != 0) {
156     assert(n >= 1 * sizeof(float));
157     assert(n <= 15 * sizeof(float));
158     // Prepare mask for valid 32-bit elements (depends on n).
159     n >>= 2 /* log2(sizeof(float)) */;
160     const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
161 
162     __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
163 
164     const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
165     const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
166 
167     __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
168     const __m512i ven = _mm512_slli_epi32(_mm512_castps_si512(vn), 19);
169     const __m512i vl = _mm512_permutexvar_epi32(_mm512_castps_si512(vn), vtable);
170     __m512 vs = _mm512_castsi512_ps(_mm512_add_epi32(vl, ven));
171     vn = _mm512_sub_ps(vn, vmagic_bias);
172 
173     __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
174 
175     __m512 vp = _mm512_fmadd_ps(vc3, vt, vc2);
176     vp = _mm512_mul_ps(vp, vt);
177 
178     vt = _mm512_mul_ps(vt, vs);
179     vs = _mm512_fmsub_ps(vs, valpha, valpha);
180     vp = _mm512_fmadd_ps(vp, vt, vt);
181     __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
182 
183     vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
184 
185     _mm512_mask_storeu_ps(y, vmask, vy);
186   }
187 }
188