xref: /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/velu-scalar-rr2-p6-x2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Auto-generated file. Do not edit!
2*4bdc9457SAndroid Build Coastguard Worker //   Template: src/f32-velu/scalar-rr2-p6.c.in
3*4bdc9457SAndroid Build Coastguard Worker //   Generator: tools/xngen
4*4bdc9457SAndroid Build Coastguard Worker //
5*4bdc9457SAndroid Build Coastguard Worker // Copyright 2020 Google LLC
6*4bdc9457SAndroid Build Coastguard Worker //
7*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
8*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
11*4bdc9457SAndroid Build Coastguard Worker #include <math.h>
12*4bdc9457SAndroid Build Coastguard Worker 
13*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/common.h>
14*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/math.h>
15*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
16*4bdc9457SAndroid Build Coastguard Worker 
17*4bdc9457SAndroid Build Coastguard Worker 
xnn_f32_velu_ukernel__scalar_rr2_p6_x2(size_t n,const float * x,float * y,const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS (1)])18*4bdc9457SAndroid Build Coastguard Worker void xnn_f32_velu_ukernel__scalar_rr2_p6_x2(
19*4bdc9457SAndroid Build Coastguard Worker     size_t n,
20*4bdc9457SAndroid Build Coastguard Worker     const float* x,
21*4bdc9457SAndroid Build Coastguard Worker     float* y,
22*4bdc9457SAndroid Build Coastguard Worker     const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
23*4bdc9457SAndroid Build Coastguard Worker {
24*4bdc9457SAndroid Build Coastguard Worker   assert(n % sizeof(float) == 0);
25*4bdc9457SAndroid Build Coastguard Worker 
26*4bdc9457SAndroid Build Coastguard Worker   const float vprescale = params->scalar_rr2_p6.prescale;
27*4bdc9457SAndroid Build Coastguard Worker   const float valpha = params->scalar_rr2_p6.alpha;
28*4bdc9457SAndroid Build Coastguard Worker   const float vbeta = params->scalar_rr2_p6.beta;
29*4bdc9457SAndroid Build Coastguard Worker   const float vmagic_bias = params->scalar_rr2_p6.magic_bias;
30*4bdc9457SAndroid Build Coastguard Worker   const float vlog2e = params->scalar_rr2_p6.log2e;
31*4bdc9457SAndroid Build Coastguard Worker   const float vsat_cutoff = params->scalar_rr2_p6.sat_cutoff;
32*4bdc9457SAndroid Build Coastguard Worker   const float vminus_ln2_hi = params->scalar_rr2_p6.minus_ln2_hi;
33*4bdc9457SAndroid Build Coastguard Worker   const float vminus_ln2_lo = params->scalar_rr2_p6.minus_ln2_lo;
34*4bdc9457SAndroid Build Coastguard Worker   const float vc6 = params->scalar_rr2_p6.c6;
35*4bdc9457SAndroid Build Coastguard Worker   const float vc5 = params->scalar_rr2_p6.c5;
36*4bdc9457SAndroid Build Coastguard Worker   const float vc4 = params->scalar_rr2_p6.c4;
37*4bdc9457SAndroid Build Coastguard Worker   const float vc3 = params->scalar_rr2_p6.c3;
38*4bdc9457SAndroid Build Coastguard Worker   const float vc2 = params->scalar_rr2_p6.c2;
39*4bdc9457SAndroid Build Coastguard Worker   const float vone = params->scalar_rr2_p6.one;
40*4bdc9457SAndroid Build Coastguard Worker 
41*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 2 * sizeof(float); n -= 2 * sizeof(float)) {
42*4bdc9457SAndroid Build Coastguard Worker     float vx0 = x[0];
43*4bdc9457SAndroid Build Coastguard Worker     float vx1 = x[1];
44*4bdc9457SAndroid Build Coastguard Worker     x += 2;
45*4bdc9457SAndroid Build Coastguard Worker 
46*4bdc9457SAndroid Build Coastguard Worker     const float vz0 = vx0 * vprescale;
47*4bdc9457SAndroid Build Coastguard Worker     const float vz1 = vx1 * vprescale;
48*4bdc9457SAndroid Build Coastguard Worker 
49*4bdc9457SAndroid Build Coastguard Worker     float vn0 = vz0 * vlog2e + vmagic_bias;
50*4bdc9457SAndroid Build Coastguard Worker     float vn1 = vz1 * vlog2e + vmagic_bias;
51*4bdc9457SAndroid Build Coastguard Worker 
52*4bdc9457SAndroid Build Coastguard Worker     float vs0 = uint32_as_float(float_as_uint32(vn0) << 23);
53*4bdc9457SAndroid Build Coastguard Worker     vn0 -= vmagic_bias;
54*4bdc9457SAndroid Build Coastguard Worker     float vs1 = uint32_as_float(float_as_uint32(vn1) << 23);
55*4bdc9457SAndroid Build Coastguard Worker     vn1 -= vmagic_bias;
56*4bdc9457SAndroid Build Coastguard Worker 
57*4bdc9457SAndroid Build Coastguard Worker     float vt0 = vn0 * vminus_ln2_hi + vz0;
58*4bdc9457SAndroid Build Coastguard Worker     float vt1 = vn1 * vminus_ln2_hi + vz1;
59*4bdc9457SAndroid Build Coastguard Worker 
60*4bdc9457SAndroid Build Coastguard Worker     vt0 = vn0 * vminus_ln2_lo + vt0;
61*4bdc9457SAndroid Build Coastguard Worker     vt1 = vn1 * vminus_ln2_lo + vt1;
62*4bdc9457SAndroid Build Coastguard Worker 
63*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vz0 <= vsat_cutoff) {
64*4bdc9457SAndroid Build Coastguard Worker       vs0 = 0.0f;
65*4bdc9457SAndroid Build Coastguard Worker       vt0 = 0.0f;
66*4bdc9457SAndroid Build Coastguard Worker     }
67*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vz1 <= vsat_cutoff) {
68*4bdc9457SAndroid Build Coastguard Worker       vs1 = 0.0f;
69*4bdc9457SAndroid Build Coastguard Worker       vt1 = 0.0f;
70*4bdc9457SAndroid Build Coastguard Worker     }
71*4bdc9457SAndroid Build Coastguard Worker 
72*4bdc9457SAndroid Build Coastguard Worker     float vp0 = vc6 * vt0 + vc5;
73*4bdc9457SAndroid Build Coastguard Worker     float vp1 = vc6 * vt1 + vc5;
74*4bdc9457SAndroid Build Coastguard Worker 
75*4bdc9457SAndroid Build Coastguard Worker     vp0 = vp0 * vt0 + vc4;
76*4bdc9457SAndroid Build Coastguard Worker     vp1 = vp1 * vt1 + vc4;
77*4bdc9457SAndroid Build Coastguard Worker 
78*4bdc9457SAndroid Build Coastguard Worker     vp0 = vp0 * vt0 + vc3;
79*4bdc9457SAndroid Build Coastguard Worker     vp1 = vp1 * vt1 + vc3;
80*4bdc9457SAndroid Build Coastguard Worker 
81*4bdc9457SAndroid Build Coastguard Worker     vp0 = vp0 * vt0 + vc2;
82*4bdc9457SAndroid Build Coastguard Worker     vp1 = vp1 * vt1 + vc2;
83*4bdc9457SAndroid Build Coastguard Worker 
84*4bdc9457SAndroid Build Coastguard Worker     vp0 *= vt0;
85*4bdc9457SAndroid Build Coastguard Worker     vp1 *= vt1;
86*4bdc9457SAndroid Build Coastguard Worker 
87*4bdc9457SAndroid Build Coastguard Worker     vt0 *= vs0;
88*4bdc9457SAndroid Build Coastguard Worker     vs0 -= vone;
89*4bdc9457SAndroid Build Coastguard Worker     vt1 *= vs1;
90*4bdc9457SAndroid Build Coastguard Worker     vs1 -= vone;
91*4bdc9457SAndroid Build Coastguard Worker 
92*4bdc9457SAndroid Build Coastguard Worker     vp0 = vp0 * vt0 + vt0;
93*4bdc9457SAndroid Build Coastguard Worker     vp1 = vp1 * vt1 + vt1;
94*4bdc9457SAndroid Build Coastguard Worker 
95*4bdc9457SAndroid Build Coastguard Worker     const float ve0 = (vp0 + vs0) * valpha;
96*4bdc9457SAndroid Build Coastguard Worker     float vy0 = vx0 * vbeta;
97*4bdc9457SAndroid Build Coastguard Worker     const float ve1 = (vp1 + vs1) * valpha;
98*4bdc9457SAndroid Build Coastguard Worker     float vy1 = vx1 * vbeta;
99*4bdc9457SAndroid Build Coastguard Worker 
100*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vx0 < 0.0f) {
101*4bdc9457SAndroid Build Coastguard Worker       vy0 = ve0;
102*4bdc9457SAndroid Build Coastguard Worker     }
103*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vx1 < 0.0f) {
104*4bdc9457SAndroid Build Coastguard Worker       vy1 = ve1;
105*4bdc9457SAndroid Build Coastguard Worker     }
106*4bdc9457SAndroid Build Coastguard Worker 
107*4bdc9457SAndroid Build Coastguard Worker     y[0] = vy0;
108*4bdc9457SAndroid Build Coastguard Worker     y[1] = vy1;
109*4bdc9457SAndroid Build Coastguard Worker     y += 2;
110*4bdc9457SAndroid Build Coastguard Worker   }
111*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
112*4bdc9457SAndroid Build Coastguard Worker     float vx = *x;
113*4bdc9457SAndroid Build Coastguard Worker 
114*4bdc9457SAndroid Build Coastguard Worker     const float vz = vx * vprescale;
115*4bdc9457SAndroid Build Coastguard Worker 
116*4bdc9457SAndroid Build Coastguard Worker     float vn = vz * vlog2e + vmagic_bias;
117*4bdc9457SAndroid Build Coastguard Worker     float vs = uint32_as_float(float_as_uint32(vn) << 23);
118*4bdc9457SAndroid Build Coastguard Worker     vn -= vmagic_bias;
119*4bdc9457SAndroid Build Coastguard Worker 
120*4bdc9457SAndroid Build Coastguard Worker     float vt = vn * vminus_ln2_hi + vz;
121*4bdc9457SAndroid Build Coastguard Worker     vt = vn * vminus_ln2_lo + vt;
122*4bdc9457SAndroid Build Coastguard Worker 
123*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vz <= vsat_cutoff) {
124*4bdc9457SAndroid Build Coastguard Worker       vs = 0.0f;
125*4bdc9457SAndroid Build Coastguard Worker       vt = 0.0f;
126*4bdc9457SAndroid Build Coastguard Worker     }
127*4bdc9457SAndroid Build Coastguard Worker 
128*4bdc9457SAndroid Build Coastguard Worker     float vp = vc6 * vt + vc5;
129*4bdc9457SAndroid Build Coastguard Worker     vp = vp * vt + vc4;
130*4bdc9457SAndroid Build Coastguard Worker     vp = vp * vt + vc3;
131*4bdc9457SAndroid Build Coastguard Worker     vp = vp * vt + vc2;
132*4bdc9457SAndroid Build Coastguard Worker     vp *= vt;
133*4bdc9457SAndroid Build Coastguard Worker 
134*4bdc9457SAndroid Build Coastguard Worker     vt *= vs;
135*4bdc9457SAndroid Build Coastguard Worker     vs -= vone;
136*4bdc9457SAndroid Build Coastguard Worker     vp = vp * vt + vt;
137*4bdc9457SAndroid Build Coastguard Worker     const float ve = (vp + vs) * valpha;
138*4bdc9457SAndroid Build Coastguard Worker 
139*4bdc9457SAndroid Build Coastguard Worker     float vy = vx * vbeta;
140*4bdc9457SAndroid Build Coastguard Worker     if XNN_UNPREDICTABLE(vx < 0.0f) {
141*4bdc9457SAndroid Build Coastguard Worker       vy = ve;
142*4bdc9457SAndroid Build Coastguard Worker     }
143*4bdc9457SAndroid Build Coastguard Worker 
144*4bdc9457SAndroid Build Coastguard Worker     *y = vy;
145*4bdc9457SAndroid Build Coastguard Worker   }
146*4bdc9457SAndroid Build Coastguard Worker }
147