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 <stddef.h>
8 #include <stdint.h>
9 
10 #include <wasm_simd128.h>
11 
12 #include <xnnpack/math-stubs.h>
13 
14 
xnn_math_f32_roundz__wasmsimd_addsub(size_t n,const float * input,float * output)15 void xnn_math_f32_roundz__wasmsimd_addsub(
16     size_t n,
17     const float* input,
18     float* output)
19 {
20   assert(n % (4 * sizeof(float)) == 0);
21 
22   // Mask for the sign bit of a floating-point number.
23   const v128_t vsign_mask = wasm_i32x4_splat(INT32_C(0x80000000));
24   // Addition of this number to a floating-point number x cause rounding of the result to an integer. Then this magic
25   // number is subtracted back from the result to get original x rounded to integer. This trick works only for
26   // 0 <= x < 2**24, but all numbers in 2**23 <= x < 2**24 range are integers, so we can further restrict it to
27   // 0 <= x < 2**23. Then the upper bound of the validity interval is conveniently the same as the magic number.
28   const v128_t vmagic_number = wasm_f32x4_splat(0x1.000000p+23f);
29   // Unit constant to decrement absolute values rounded "wrong way" (i.e. away from zero) in the round-to-nearest-even
30   // operation.
31   const v128_t vone = wasm_f32x4_splat(1.0f);
32 
33   for (; n != 0; n -= 4 * sizeof(float)) {
34     const v128_t vx = wasm_v128_load(input);
35     input += 4;
36 
37     // The rounding trick works only for x >= 0, so we compute absolute value of x, round it, and restore the sign in
38     // the end. This method works for round-toward-zero because it is an odd function.
39     const v128_t vabsx = wasm_v128_andnot(vx, vsign_mask);
40 
41     // Compute bitmask for the bits we want to copy from x. Other bits will be copied from the rounded abs(x).
42     // If abs(x) < 2**23 or x is NaN, we want the sign bit from x and the rest from the rounded abs(x).
43     // Otherwise (abs(x) >= 2**23), we want all bits from x.
44     const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_ge(vabsx, vmagic_number));
45     // Addition-subtraction trick with the magic number to cause rounding to integer for abs(x).
46     // Note: the result is valid only for 0 <= abs(x) < 2**23.
47     // Note: addition-subtraction implicitly converts SNaN inputs to QNaNs.
48     const v128_t vrndabsx = wasm_f32x4_sub(wasm_f32x4_add(vabsx, vmagic_number), vmagic_number);
49 
50     // Compute adjustment to be subtracted from the rounded-to-nearest-even abs(x) value.
51     // Adjustment is one if the rounded value is greater than the abs(x) value and zero otherwise (including NaN input).
52     const v128_t vadjustment = wasm_v128_and(wasm_f32x4_gt(vrndabsx, vabsx), vone);
53     // Adjust abs(x) rounded to nearest-even via the addition-subtraction trick to get abs(x) rounded down.
54     // Note: subtraction implicitly converts SNaN inputs to QNaNs.
55     const v128_t vflrabsx = wasm_f32x4_sub(vrndabsx, vadjustment);
56 
57     // Combine abs(x) rounded down via addition-subtraction trick with adjustment and the input x value.
58     // For abs(x) < 2**23, the result is abs(x) rounded via addition-subtraction trick with the sign of x.
59     // For NaN inputs, the result is x converted to QNaN as a side-effect of addition-subtraction and adjustment.
60     // For abs(x) >= 2**23, the result is x itself.
61     const v128_t vy = wasm_v128_bitselect(vx, vflrabsx, vrndmask);
62 
63     wasm_v128_store(output, vy);
64     output += 4;
65   }
66 }
67