• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2020 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 <math.h>
8 #include <stddef.h>
9 
10 #include <xnnpack/common.h>
11 #include <xnnpack/math-stubs.h>
12 
13 
xnn_math_f32_roundne__scalar_addsub(size_t n,const float * input,float * output)14 void xnn_math_f32_roundne__scalar_addsub(
15     size_t n,
16     const float* input,
17     float* output)
18 {
19   assert(n % sizeof(float) == 0);
20 
21   // Addition of this number to a floating-point number x cause rounding of the result to an integer. Then this magic
22   // number is subtracted back from the result to get original x rounded to integer. This trick works only for
23   // 0 <= x < 2**24, but all numbers in 2**23 <= x < 2**24 range are integers, so we can further restrict it to
24   // 0 <= x < 2**23. Then the upper bound of the validity interval is conveniently the same as the magic number.
25   const float vmagic_number = 0x1.000000p+23f;
26 
27   for (; n != 0; n -= sizeof(float)) {
28     const float vx = *input++;
29 
30     // The rounding trick works only for x >= 0, so we compute absolute value of x, round it, and restore the sign in
31     // the end. This method works for round-to-nearest-even because it is an odd function.
32     const float vabsx = fabsf(vx);
33     // Addition-subtraction trick with the magic number to cause rounding to integer for abs(x).
34     // Note: the result is valid only for 0 <= abs(x) < 2**23.
35     // Note: addition-subtraction implicitly converts SNaN inputs to QNaNs.
36     const float vrndabsx = (vabsx + vmagic_number) - vmagic_number;
37 
38     // Select between the abs(x) rounded using addition-subtraction trick and the abs(x) value.
39     // For abs(x) < 2**23, the result is abs(x) rounded via addition-subtraction trick.
40     // For abs(x) >= 2**23, the result is abs(x) itself (already an integer).
41     // For NaN inputs, the result is abs(x) converted to QNaN as a side-effect of addition-subtraction.
42     const float vabsy = XNN_UNPREDICTABLE(vabsx >= vmagic_number) ? vabsx : vrndabsx;
43     // Restore the sign of the rounded value.
44     const float vy = copysignf(vabsy, vx);
45 
46     *output++ = vy;
47   }
48 }
49