xref: /aosp_15_r20/external/XNNPACK/src/math/expminus-f32-scalar-rr2-p5.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #include <assert.h>
7 #include <stddef.h>
8 
9 #include <xnnpack/common.h>
10 #include <xnnpack/math.h>
11 #include <xnnpack/math-stubs.h>
12 
13 
xnn_math_f32_expminus__scalar_rr2_p5(size_t n,const float * input,float * output)14 void xnn_math_f32_expminus__scalar_rr2_p5(
15     size_t n,
16     const float* input,
17     float* output)
18 {
19   assert(n % sizeof(float) == 0);
20 
21   // Large number such that ulp(magic bias) == 1 and magic bias === 127 mod 2**22.
22   const float vmagic_bias = 0x1.8000FEp23f;
23   const float vlog2e = 0x1.715476p+0f;
24   // Last 7 bits are zeroes
25   const float vminus_ln2_hi = -0x1.62E400p-1f;
26   const float vminus_ln2_lo = -0x1.7F7D1Cp-20f;
27   // Coefficient of polynomial approximation
28   //   exp(t) ~ 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
29   // on [-log(2)/2, log(2)/2]
30   const float vc5 = 0x1.0F9F9Cp-7f;
31   const float vc4 = 0x1.573A1Ap-5f;
32   const float vc3 = 0x1.555A80p-3f;
33   const float vc2 = 0x1.FFFDC6p-2f;
34   const float vc1 = 0x1.FFFFF6p-1f;
35   // The smallest x for which expf(x) is normalized.
36   const float vdenorm_cutoff = -0x1.5D589Ep6f;
37 
38   for (; n != 0; n -= sizeof(float)) {
39     const float vx = *input++;
40 
41     // Compute reduced argument n := round(x / log(2)).
42     // We do it by adding a large number (magic bias) to the product x * (1/log(2)), which cause rounding of the result
43     // to an integer, then subtracing the large number back. The trick with adding large number is valid only within
44     // certain bounds (|x / log(2)| <= 2**22, i.e. |x| <= 0x1.62E43p+21 = 2907270.0), but that is acceptable, because
45     // inputs outside of [-87.336540, 0.0] underflow expf(x) anyway. We fixup the result for such inputs at the very
46     // end of the algorithm.
47     float vn = vx * vlog2e + vmagic_bias;
48 
49     // Create a floating-point number s (scale) such that s == 2**n for inputs which don't cause underflow, i.e.
50     // -87.33642 <= x <= 0.0, and -126 <= n <= 0 accordingly.
51     const float vs = uint32_as_float(float_as_uint32(vn) << 23);
52 
53     // Subtract the large number back to get final n := round(x / log(2)) as a floating-point number.
54     vn -= vmagic_bias;
55 
56     // Compute reduced argument t := x - n * log(2).
57     // Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy.
58     float vt = vn * vminus_ln2_hi + vx;
59     vt = vn * vminus_ln2_lo + vt;
60 
61     // Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]:
62     //   P(t) = 1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) = 1 + t * p
63     float vp = vc5 * vt + vc4;
64     vp = vp * vt + vc3;
65     vp = vp * vt + vc2;
66     vp = vp * vt + vc1;
67 
68     // Reconstruct the exp(x) value:
69     //   exp(x) = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))))
70     //          = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))
71     //          = s + (t * s) * p
72     vt *= vs;
73     float vf = vt * vp + vs;
74 
75     // For inputs below denormal cutoff, replace output with +0.0f.
76     // Note that for NaN inputs, comparison result is false, and outputs are left unchanged.
77     if XNN_UNPREDICTABLE(vx < vdenorm_cutoff) {
78       vf = 0.0f;
79     }
80 
81     *output++ = vf;
82   }
83 }
84