xref: /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.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_x2(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_x2(
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 >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
42     float vx0 = x[0];
43     float vx1 = x[1];
44     x += 2;
45 
46     const float vz0 = vx0 * vprescale;
47     const float vz1 = vx1 * vprescale;
48 
49     float vn0 = vz0 * vlog2e + vmagic_bias;
50     float vn1 = vz1 * vlog2e + vmagic_bias;
51 
52     const uint32_t ven0 = float_as_uint32(vn0) << 19;
53     const uint32_t vidx0 = float_as_uint32(vn0) & vindex_mask;
54     vn0 -= vmagic_bias;
55     const uint32_t ven1 = float_as_uint32(vn1) << 19;
56     const uint32_t vidx1 = float_as_uint32(vn1) & vindex_mask;
57     vn1 -= vmagic_bias;
58 
59     float vt0 = vn0 * vminus_ln2_hi + vz0;
60     float vs0 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx0] + ven0);
61     float vt1 = vn1 * vminus_ln2_hi + vz1;
62     float vs1 = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx1] + ven1);
63 
64     vt0 = vn0 * vminus_ln2_lo + vt0;
65     if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) {
66       vs0 = 0.0f;
67       vt0 = 0.0f;
68     }
69     vt1 = vn1 * vminus_ln2_lo + vt1;
70     if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) {
71       vs1 = 0.0f;
72       vt1 = 0.0f;
73     }
74 
75     float vp0 = vc3 * vt0 + vc2;
76     float vp1 = vc3 * vt1 + vc2;
77 
78     vp0 *= vt0;
79     vp1 *= vt1;
80 
81     vt0 *= vs0;
82     vs0 -= vone;
83     vt1 *= vs1;
84     vs1 -= vone;
85 
86     vp0 = vp0 * vt0 + vt0;
87     vp1 = vp1 * vt1 + vt1;
88 
89     const float ve0 = (vp0 + vs0) * valpha;
90     float vy0 = vx0 * vbeta;
91     const float ve1 = (vp1 + vs1) * valpha;
92     float vy1 = vx1 * vbeta;
93 
94     if XNN_UNPREDICTABLE(vx0 < 0.0f) {
95       vy0 = ve0;
96     }
97     if XNN_UNPREDICTABLE(vx1 < 0.0f) {
98       vy1 = ve1;
99     }
100 
101     y[0] = vy0;
102     y[1] = vy1;
103     y += 2;
104   }
105   if XNN_UNLIKELY(n != 0) {
106     float vx = *x;
107 
108     const float vz = vx * vprescale;
109 
110     float vn = vz * vlog2e + vmagic_bias;
111     const uint32_t ven = float_as_uint32(vn) << 19;
112     const uint32_t vidx = float_as_uint32(vn) & vindex_mask;
113     vn -= vmagic_bias;
114 
115     float vt = vn * vminus_ln2_hi + vz;
116     float vs = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx] + ven);
117 
118     vt = vn * vminus_ln2_lo + vt;
119     if XNN_UNPREDICTABLE(vz <= vsat_cutoff) {
120       vs = 0.0f;
121       vt = 0.0f;
122     }
123 
124     float vp = vc3 * vt + vc2;
125     vp *= vt;
126 
127     vt *= vs;
128     vs -= vone;
129     vp = vp * vt + vt;
130     const float ve = (vp + vs) * valpha;
131 
132     float vy = vx * vbeta;
133     if XNN_UNPREDICTABLE(vx < 0.0f) {
134       vy = ve;
135     }
136 
137     *y = vy;
138   }
139 }
140