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_cvt(size_t n,const float * input,float * output)15 void xnn_math_f32_roundz__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
28 for (; n != 0; n -= 4 * sizeof(float)) {
29 const v128_t vx = wasm_v128_load(input);
30 input += 4;
31
32 // Convert floating-point value x to integer, with rounding towards zero, and then back to floating-point.
33 // Note: the result is valid only for abs(x) < 2**31, but we further restrict its use to 2**23.
34 const v128_t vrndx = wasm_f32x4_convert_i32x4(wasm_i32x4_trunc_saturate_f32x4(vx));
35
36 // Compute bitmask for the bits we want to copy from the rounded x. Other bits will be copied from x.
37 // If abs(x) is below the integral threshold, use all but the sign bit from the rounded x and the sign bit from x.
38 // If x is guaranteed integral or NaN, use all bits from x.
39 const v128_t vrndmask = wasm_v128_andnot(wasm_f32x4_lt(wasm_f32x4_abs(vx), vintegral_threshold), vsign_mask);
40
41 // Combine x rounded towardz zero via FP->INT->FP conversion and the input x value.
42 // For 0.0 <= x < 2**23, the result is x rounded via FP->INT->FP conversion.
43 // For -2**23 < x <= -0.0, the result is abs(x) rounded via FP->INT->FP conversion with the sign of x.
44 // For abs(x) >= 2**23 or NaN inputs, the result is x itself.
45 const v128_t vy = wasm_v128_bitselect(vrndx, vx, vrndmask);
46
47 wasm_v128_store(output, vy);
48 output += 4;
49 }
50 }
51