1 // Auto-generated file. Do not edit!
2 // Template: src/f32-velu/avx512f-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/intrinsics-polyfill.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_velu_ukernel__avx512f_rr1_p6_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_p6_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(0x1.8000FEp23f);
34 const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f);
35 const __m512 vminus_ln2 = _mm512_set1_ps(-0x1.62E43p-1f);
36 const __m512 vc6 = _mm512_set1_ps(0x1.6b7338p-10f);
37 const __m512 vc5 = _mm512_set1_ps(0x1.12278Ep-7f);
38 const __m512 vc4 = _mm512_set1_ps(0x1.555716p-5f);
39 const __m512 vc3 = _mm512_set1_ps(0x1.5554B0p-3f);
40 const __m512 vc2 = _mm512_set1_ps(0x1.FFFFFEp-2f);
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 __m512 vs0 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn0), 23));
57 vn0 = _mm512_sub_ps(vn0, vmagic_bias);
58 __m512 vs1 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn1), 23));
59 vn1 = _mm512_sub_ps(vn1, vmagic_bias);
60 __m512 vs2 = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn2), 23));
61 vn2 = _mm512_sub_ps(vn2, vmagic_bias);
62
63 __m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2, vz0);
64 __m512 vt1 = _mm512_fmadd_ps(vn1, vminus_ln2, vz1);
65 __m512 vt2 = _mm512_fmadd_ps(vn2, vminus_ln2, vz2);
66
67 __m512 vp0 = _mm512_fmadd_ps(vc6, vt0, vc5);
68 __m512 vp1 = _mm512_fmadd_ps(vc6, vt1, vc5);
69 __m512 vp2 = _mm512_fmadd_ps(vc6, vt2, vc5);
70
71 vp0 = _mm512_fmadd_ps(vp0, vt0, vc4);
72 vp1 = _mm512_fmadd_ps(vp1, vt1, vc4);
73 vp2 = _mm512_fmadd_ps(vp2, vt2, vc4);
74
75 vp0 = _mm512_fmadd_ps(vp0, vt0, vc3);
76 vp1 = _mm512_fmadd_ps(vp1, vt1, vc3);
77 vp2 = _mm512_fmadd_ps(vp2, vt2, vc3);
78
79 vp0 = _mm512_fmadd_ps(vp0, vt0, vc2);
80 vp1 = _mm512_fmadd_ps(vp1, vt1, vc2);
81 vp2 = _mm512_fmadd_ps(vp2, vt2, vc2);
82
83 vp0 = _mm512_mul_ps(vp0, vt0);
84 vt0 = _mm512_mul_ps(vt0, vs0);
85 vp1 = _mm512_mul_ps(vp1, vt1);
86 vt1 = _mm512_mul_ps(vt1, vs1);
87 vp2 = _mm512_mul_ps(vp2, vt2);
88 vt2 = _mm512_mul_ps(vt2, vs2);
89
90 vs0 = _mm512_fmsub_ps(vs0, valpha, valpha);
91 vs1 = _mm512_fmsub_ps(vs1, valpha, valpha);
92 vs2 = _mm512_fmsub_ps(vs2, valpha, valpha);
93
94 vp0 = _mm512_fmadd_ps(vp0, vt0, vt0);
95 vp1 = _mm512_fmadd_ps(vp1, vt1, vt1);
96 vp2 = _mm512_fmadd_ps(vp2, vt2, vt2);
97
98 const __m512 vzero = _mm512_setzero_ps();
99 __m512 vy0 = _mm512_fmadd_ps(vp0, valpha, vs0);
100 const __mmask16 vsign0 = _mm512_cmp_ps_mask(vx0, vzero, _CMP_NLT_US);
101 __m512 vy1 = _mm512_fmadd_ps(vp1, valpha, vs1);
102 const __mmask16 vsign1 = _mm512_cmp_ps_mask(vx1, vzero, _CMP_NLT_US);
103 __m512 vy2 = _mm512_fmadd_ps(vp2, valpha, vs2);
104 const __mmask16 vsign2 = _mm512_cmp_ps_mask(vx2, vzero, _CMP_NLT_US);
105
106 vy0 = _mm512_mask_mul_ps(vy0, vsign0, vx0, vbeta);
107 vy1 = _mm512_mask_mul_ps(vy1, vsign1, vx1, vbeta);
108 vy2 = _mm512_mask_mul_ps(vy2, vsign2, vx2, vbeta);
109
110 _mm512_storeu_ps(y, vy0);
111 _mm512_storeu_ps(y + 16, vy1);
112 _mm512_storeu_ps(y + 32, vy2);
113 y += 48;
114 }
115 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
116 __m512 vx = _mm512_loadu_ps(x);
117 x += 16;
118
119 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
120 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
121
122 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
123 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
124 vn = _mm512_sub_ps(vn, vmagic_bias);
125
126 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
127
128 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
129 vp = _mm512_fmadd_ps(vp, vt, vc4);
130 vp = _mm512_fmadd_ps(vp, vt, vc3);
131 vp = _mm512_fmadd_ps(vp, vt, vc2);
132 vp = _mm512_mul_ps(vp, vt);
133
134 vt = _mm512_mul_ps(vt, vs);
135 vs = _mm512_fmsub_ps(vs, valpha, valpha);
136 vp = _mm512_fmadd_ps(vp, vt, vt);
137 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
138
139 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
140
141 _mm512_storeu_ps(y, vy);
142 y += 16;
143 }
144 if XNN_UNLIKELY(n != 0) {
145 assert(n >= 1 * sizeof(float));
146 assert(n <= 15 * sizeof(float));
147 // Prepare mask for valid 32-bit elements (depends on n).
148 n >>= 2 /* log2(sizeof(float)) */;
149 const __mmask16 vmask = _cvtu32_mask16((uint16_t) ((uint32_t) (UINT32_C(1) << n) - UINT32_C(1)));
150
151 __m512 vx = _mm512_maskz_loadu_ps(vmask, x);
152
153 const __m512 vz = _mm512_max_ps(vsat_cutoff, _mm512_mul_ps(vx, vprescale));
154 const __mmask16 vsign = _mm512_cmp_ps_mask(vx, _mm512_setzero_ps(), _CMP_NLT_US);
155
156 __m512 vn = _mm512_fmadd_ps(vz, vlog2e, vmagic_bias);
157 __m512 vs = _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_castps_si512(vn), 23));
158 vn = _mm512_sub_ps(vn, vmagic_bias);
159
160 __m512 vt = _mm512_fmadd_ps(vn, vminus_ln2, vz);
161
162 __m512 vp = _mm512_fmadd_ps(vc6, vt, vc5);
163 vp = _mm512_fmadd_ps(vp, vt, vc4);
164 vp = _mm512_fmadd_ps(vp, vt, vc3);
165 vp = _mm512_fmadd_ps(vp, vt, vc2);
166 vp = _mm512_mul_ps(vp, vt);
167
168 vt = _mm512_mul_ps(vt, vs);
169 vs = _mm512_fmsub_ps(vs, valpha, valpha);
170 vp = _mm512_fmadd_ps(vp, vt, vt);
171 __m512 vy = _mm512_fmadd_ps(vp, valpha, vs);
172
173 vy = _mm512_mask_mul_ps(vy, vsign, vx, vbeta);
174
175 _mm512_mask_storeu_ps(y, vmask, vy);
176 }
177 }
178