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_roundu__wasmsimd_cvt(size_t n,const float * input,float * output)15 void xnn_math_f32_roundu__wasmsimd_cvt(
16     size_t n,
17     const float* input,
18     float* output)
19 {
20   assert(n % (4 * sizeof(float)) == 0);
21 
22   // Threshold of non-integral values in single-precision floating-point representation.
23   // All inputs above this threshold (by absolute value) are integer numbers.
24   const v128_t vintegral_threshold = wasm_f32x4_splat(0x1.000000p+23f);
25   // Mask for the sign of a single-precision floating-point number.
26   const v128_t vsign_mask = wasm_f32x4_splat(-0.0f);
27   // Unit constant to increment results rounded "wrong way" (i.e. down) in the round-towards-zero operation.
28   const v128_t vone = wasm_f32x4_splat(1.0f);
29 
30   for (; n != 0; n -= 4 * sizeof(float)) {
31     const v128_t vx = wasm_v128_load(input);
32     input += 4;
33 
34     // Convert floating-point value x to integer, with rounding towards zero, and then back to floating-point.
35     // Note: the result is valid only for abs(x) < 2**31, but we further restrict its use to 2**23.
36     const v128_t vprerndx = wasm_f32x4_convert_i32x4(wasm_i32x4_trunc_saturate_f32x4(vx));
37 
38     // Compute bitmask for the bits we want to copy from the rounded x. Other bits will be copied from x.
39     // If abs(x) is below the integral threshold, use all but the sign bit from the rounded x and the sign bit from x.
40     // If x is guaranteed integral or NaN, use all bits from x.
41     const v128_t vrndmask = wasm_v128_andnot(wasm_f32x4_lt(wasm_f32x4_abs(vx), vintegral_threshold), vsign_mask);
42 
43     // Combine x rounded towardz zero via FP->INT->FP conversion and the input x value.
44     // For 0.0 <= x < 2**23, the result is x rounded via FP->INT->FP conversion.
45     // For -2**23 < x <= -0.0, the result is abs(x) rounded via FP->INT->FP conversion with the sign of x.
46     // For abs(x) >= 2**23 or NaN inputs, the result is x itself.
47     const v128_t vrndx = wasm_v128_bitselect(vprerndx, vx, vrndmask);
48 
49     // Compute bitmask for the bits to copy from the rounded x. Other bits will be copied from the adjusted rounded x.
50     // If rounded x >= x, we want all bits from rounded x.
51     // If rounded x < x or rounded x is NaN (implies x is NaN), we want all but the sign bit from the adjusted rounded
52     // x and the sign bit from rounded x (same as the sign bit of x).
53     const v128_t vadjmask = wasm_v128_or(wasm_f32x4_ge(vrndx, vx), vsign_mask);
54     // Adjust the rounded x value.
55     // The adjusted value is a unit above the rounded-towards-zero x value, but is used only if the rounded value is
56     // below x. In these cases, the adjusted value is x rounded up.
57     // Note: addition implicitly converts SNaN inputs to QNaNs.
58     const v128_t vadjrndx = wasm_f32x4_add(vrndx, vone);
59 
60     // Combine the adjusted rounded x and the original rounded towards zero x.
61     // For rounded x < x, the result is the absolute value of adjusted rounded-towards-zero x with the sign of
62     // rounded-towards x (same as sign of x). Propagating the sign of x is important to produce negative zero
63     // for -1.0 < x < -0.5 inputs, where otherwise we would get -1.0 (rounded x) + 1.0 (adjustment) = +0.0.
64     // For rounded x >= x, the result is the rounded-towards-zero x.
65     // For NaN inputs, the result is rounded x (same as x converted to QNaN as a side-effect of adjustment).
66     const v128_t vy = wasm_v128_bitselect(vrndx, vadjrndx, vadjmask);
67 
68     wasm_v128_store(output, vy);
69     output += 4;
70   }
71 }
72