xref: /aosp_15_r20/external/XNNPACK/src/f32-velu/gen/velu-wasm-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__wasm_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__wasm_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 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx0 * vprescale, vsat_cutoff), 0.0f);
47     const float vz1 = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx1 * vprescale, vsat_cutoff), 0.0f);
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     vt1 = vn1 * vminus_ln2_lo + vt1;
66 
67     float vp0 = vc3 * vt0 + vc2;
68     float vp1 = vc3 * vt1 + vc2;
69 
70     vp0 *= vt0;
71     vp1 *= vt1;
72 
73     vt0 *= vs0;
74     vs0 -= vone;
75     vt1 *= vs1;
76     vs1 -= vone;
77 
78     vp0 = vp0 * vt0 + vt0;
79     vp1 = vp1 * vt1 + vt1;
80 
81     const float ve0 = (vp0 + vs0) * valpha;
82     float vy0 = __builtin_wasm_max_f32(vx0 * vbeta, 0.0f);
83     const float ve1 = (vp1 + vs1) * valpha;
84     float vy1 = __builtin_wasm_max_f32(vx1 * vbeta, 0.0f);
85 
86     vy0 += __builtin_wasm_min_f32(ve0, 0.0f);
87     vy1 += __builtin_wasm_min_f32(ve1, 0.0f);
88 
89     y[0] = vy0;
90     y[1] = vy1;
91     y += 2;
92   }
93   if XNN_UNLIKELY(n != 0) {
94     float vx = *x;
95 
96     const float vz = __builtin_wasm_min_f32(__builtin_wasm_max_f32(vx * vprescale, vsat_cutoff), 0.0f);
97 
98     float vn = vz * vlog2e + vmagic_bias;
99     const uint32_t ven = float_as_uint32(vn) << 19;
100     const uint32_t vidx = float_as_uint32(vn) & vindex_mask;
101     vn -= vmagic_bias;
102 
103     float vt = vn * vminus_ln2_hi + vz;
104     float vs = uint32_as_float(xnn_table_exp2minus_k_over_16[vidx] + ven);
105 
106     vt = vn * vminus_ln2_lo + vt;
107 
108     float vp = vc3 * vt + vc2;
109     vp *= vt;
110 
111     vt *= vs;
112     vs -= vone;
113     vp = vp * vt + vt;
114     const float ve = (vp + vs) * valpha;
115 
116     float vy = __builtin_wasm_max_f32(vx * vbeta, 0.0f);
117     vy += __builtin_wasm_min_f32(ve, 0.0f);
118 
119     *y = vy;
120   }
121 }
122