• 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_roundz__scalar_addsub(size_t n,const float * input,float * output)14 void xnn_math_f32_roundz__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   // Unit constant to decrement absolute values rounded "wrong way" (i.e. away from zero) in the round-to-nearest-even
27   // operation.
28   const float vone = 1.0f;
29 
30   for (; n != 0; n -= sizeof(float)) {
31     const float vx = *input++;
32 
33     // The rounding trick works only for x >= 0, so we compute absolute value of x, round it, and restore the sign in
34     // the end. This method works for round-towards-zero because it is an odd function.
35     const float vabsx = fabsf(vx);
36     // Addition-subtraction trick with the magic number to cause rounding to the nearest-even integer for abs(x).
37     // Note: the result is valid only for 0 <= abs(x) < 2**23.
38     // Note: addition-subtraction implicitly converts SNaN inputs to QNaNs.
39     const float vrndabsx = (vabsx + vmagic_number) - vmagic_number;
40 
41     // Adjust abs(x) rounded to nearest-even via the addition-subtraction trick to get abs(x) rounded down.
42     // Note: subtraction implicitly converts SNaN inputs to QNaNs.
43     const float vflrabsx = XNN_UNPREDICTABLE(vrndabsx <= vabsx) ? vrndabsx : vrndabsx - vone;
44 
45     // Select between the abs(x) rounded down using addition-subtraction trick with adjustment and the abs(x) value.
46     // For abs(x) < 2**23, the result is abs(x) rounded via addition-subtraction trick.
47     // For abs(x) >= 2**23, the result is abs(x) itself (already an integer).
48     // For NaN inputs, the result is abs(x) converted to QNaN as a side-effect of addition-subtraction.
49     const float vabsy = XNN_UNPREDICTABLE(vabsx >= vmagic_number) ? vabsx : vflrabsx;
50     // Restore the sign of the rounded value.
51     const float vy = copysignf(vabsy, vx);
52 
53     *output++ = vy;
54   }
55 }
56