xref: /aosp_15_r20/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/scalar-rr2-p5-x1.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-raddstoreexpminusmax/scalar-rr2-p5.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 <xnnpack/common.h>
13 #include <xnnpack/math.h>
14 #include <xnnpack/raddstoreexpminusmax.h>
15 
16 
xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_p5_x1(size_t elements,const float * input,const float * max,float * output,float * sum,const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_p5_x1(
18     size_t elements,
19     const float* input,
20     const float* max,
21     float* output,
22     float* sum,
23     const union xnn_f32_expminus_params params[restrict XNN_MIN_ELEMENTS(1)])
24 {
25   assert(elements % sizeof(float) == 0);
26 
27   const float vi_max = *max;
28   const float vlog2e = params->scalar_rr2_p5.log2e;
29   const float vmagic_bias = params->scalar_rr2_p5.magic_bias;
30   const float vminus_ln2_hi = params->scalar_rr2_p5.minus_ln2_hi;
31   const float vminus_ln2_lo = params->scalar_rr2_p5.minus_ln2_lo;
32   const float vc5 = params->scalar_rr2_p5.c5;
33   const float vc4 = params->scalar_rr2_p5.c4;
34   const float vc3 = params->scalar_rr2_p5.c3;
35   const float vc2 = params->scalar_rr2_p5.c2;
36   const float vc1 = params->scalar_rr2_p5.c1;
37   const float vdenorm_cutoff = params->scalar_rr2_p5.denorm_cutoff;
38 
39   float vacc = 0.0f;
40   for (; elements >= sizeof(float); elements -= sizeof(float)) {
41     // Load 1 input at a time.
42     const float vi = *input++;
43 
44     // Subtract maximum input x := i - i_max. This implies x <= 0.
45     const float vx = vi - vi_max;
46 
47     // Compute reduced argument n := round(x / log(2)).
48     // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the result
49     // to an integer, then subtracing the large number back. The trick with adding large number is valid only within
50     // certain bounds (|x| <= 2**22), but that's ok, because inputs outside of [-87.336540, 0.0] underflow expf(x)
51     // anyway. We fixup the result for such inputs at the very end of the algorithm.
52     float vn = vx * vlog2e + vmagic_bias;
53 
54     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
55     // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
56     const float vs = uint32_as_float(float_as_uint32(vn) << 23);
57 
58     // Subtract the large number back to get final n := round(x / log(2)).
59     vn -= vmagic_bias;
60 
61     // Compute reduced argument t := x - n * log(2).
62     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
63     float vt = vn * vminus_ln2_hi + vx;
64     vt = vn * vminus_ln2_lo + vt;
65 
66     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2].
67     float vp = vc5 * vt + vc4;
68     vp = vp * vt + vc3;
69     vp = vp * vt + vc2;
70     vp = vp * vt + vc1;
71 
72     // Reconstruct the final f value:
73     //   f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
74     //     = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
75     //     = s + (t * s) * p
76     vt *= vs;
77     float vf = vt * vp + vs;
78 
79     // For inputs below denormal cutoff, replace output with +0.0f.
80     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
81     if XNN_UNPREDICTABLE(vx < vdenorm_cutoff) {
82       vf = 0.0f;
83     }
84 
85     // Store 1 output at a time.
86     *output++ = vf;
87 
88     // Accumulate computed exponents.
89     vacc += vf;
90   }
91   *sum = vacc;
92 }
93