xref: /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-velu/scalar-rr2-lut16-p3.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 #include <math.h>
12 
13 #include <xnnpack/common.h>
14 #include <xnnpack/math.h>
15 #include <xnnpack/vunary.h>
16 
17 
18 extern XNN_INTERNAL const uint32_t xnn_table_exp2minus_k_over_16[16];
19 
xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x3(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__scalar_rr2_lut16_p3_x3(
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 float vprescale = params->scalar_rr2_lut16_p3.prescale;
29   const float valpha = params->scalar_rr2_lut16_p3.alpha;
30   const float vbeta = params->scalar_rr2_lut16_p3.beta;
31   const float vmagic_bias = params->scalar_rr2_lut16_p3.magic_bias;
32   const float vlog2e = params->scalar_rr2_lut16_p3.log2e;
33   const uint32_t vindex_mask = UINT32_C(0xF);
34   const float vsat_cutoff = params->scalar_rr2_lut16_p3.sat_cutoff;
35   const float vminus_ln2_hi = params->scalar_rr2_lut16_p3.minus_ln2_hi;
36   const float vminus_ln2_lo = params->scalar_rr2_lut16_p3.minus_ln2_lo;
37   const float vc3 = params->scalar_rr2_lut16_p3.c3;
38   const float vc2 = params->scalar_rr2_lut16_p3.c2;
39   const float vone = params->scalar_rr2_lut16_p3.one;
40 
41   for (; n >= 3 * sizeof(float); n -= 3 * sizeof(float)) {
42     float vx0 = x[0];
43     float vx1 = x[1];
44     float vx2 = x[2];
45     x += 3;
46 
47     const float vz0 = vx0 * vprescale;
48     const float vz1 = vx1 * vprescale;
49     const float vz2 = vx2 * vprescale;
50 
51     float vn0 = vz0 * vlog2e + vmagic_bias;
52     float vn1 = vz1 * vlog2e + vmagic_bias;
53     float vn2 = vz2 * vlog2e + vmagic_bias;
54 
55     const uint32_t ven0 = float_as_uint32(vn0) << 19;
56     const uint32_t vidx0 = float_as_uint32(vn0) & vindex_mask;
57     vn0 -= vmagic_bias;
58     const uint32_t ven1 = float_as_uint32(vn1) << 19;
59     const uint32_t vidx1 = float_as_uint32(vn1) & vindex_mask;
60     vn1 -= vmagic_bias;
61     const uint32_t ven2 = float_as_uint32(vn2) << 19;
62     const uint32_t vidx2 = float_as_uint32(vn2) & vindex_mask;
63     vn2 -= vmagic_bias;
64 
65     float vt0 = vn0 * vminus_ln2_hi + vz0;
66     float vs0 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx0] + ven0);
67     float vt1 = vn1 * vminus_ln2_hi + vz1;
68     float vs1 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx1] + ven1);
69     float vt2 = vn2 * vminus_ln2_hi + vz2;
70     float vs2 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx2] + ven2);
71 
72     vt0 = vn0 * vminus_ln2_lo + vt0;
73     if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) {
74       vs0 = 0.0f;
75       vt0 = 0.0f;
76     }
77     vt1 = vn1 * vminus_ln2_lo + vt1;
78     if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) {
79       vs1 = 0.0f;
80       vt1 = 0.0f;
81     }
82     vt2 = vn2 * vminus_ln2_lo + vt2;
83     if XNN_UNPREDICTABLE(vz2 <= vsat_cutoff) {
84       vs2 = 0.0f;
85       vt2 = 0.0f;
86     }
87 
88     float vp0 = vc3 * vt0 + vc2;
89     float vp1 = vc3 * vt1 + vc2;
90     float vp2 = vc3 * vt2 + vc2;
91 
92     vp0 *= vt0;
93     vp1 *= vt1;
94     vp2 *= vt2;
95 
96     vt0 *= vs0;
97     vs0 -= vone;
98     vt1 *= vs1;
99     vs1 -= vone;
100     vt2 *= vs2;
101     vs2 -= vone;
102 
103     vp0 = vp0 * vt0 + vt0;
104     vp1 = vp1 * vt1 + vt1;
105     vp2 = vp2 * vt2 + vt2;
106 
107     const float ve0 = (vp0 + vs0) * valpha;
108     float vy0 = vx0 * vbeta;
109     const float ve1 = (vp1 + vs1) * valpha;
110     float vy1 = vx1 * vbeta;
111     const float ve2 = (vp2 + vs2) * valpha;
112     float vy2 = vx2 * vbeta;
113 
114     if XNN_UNPREDICTABLE(vx0 < 0.0f) {
115       vy0 = ve0;
116     }
117     if XNN_UNPREDICTABLE(vx1 < 0.0f) {
118       vy1 = ve1;
119     }
120     if XNN_UNPREDICTABLE(vx2 < 0.0f) {
121       vy2 = ve2;
122     }
123 
124     y[0] = vy0;
125     y[1] = vy1;
126     y[2] = vy2;
127     y += 3;
128   }
129   if XNN_UNLIKELY(n != 0) {
130     do {
131       float vx = *x++;
132 
133       const float vz = vx * vprescale;
134 
135       float vn = vz * vlog2e + vmagic_bias;
136       const uint32_t ven = float_as_uint32(vn) << 19;
137       const uint32_t vidx = float_as_uint32(vn) & vindex_mask;
138       vn -= vmagic_bias;
139 
140       float vt = vn * vminus_ln2_hi + vz;
141       float vs = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx] + ven);
142 
143       vt = vn * vminus_ln2_lo + vt;
144       if XNN_UNPREDICTABLE(vz <= vsat_cutoff) {
145         vs = 0.0f;
146         vt = 0.0f;
147       }
148 
149       float vp = vc3 * vt + vc2;
150       vp *= vt;
151 
152       vt *= vs;
153       vs -= vone;
154       vp = vp * vt + vt;
155       const float ve = (vp + vs) * valpha;
156 
157       float vy = vx * vbeta;
158       if XNN_UNPREDICTABLE(vx < 0.0f) {
159         vy = ve;
160       }
161 
162       *y++ = vy;
163 
164       n -= sizeof(float);
165     } while (n != 0);
166   }
167 }
168