1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vrnd/vrndu-wasmsimd-addsub.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #include <wasm_simd128.h>
13
14 #include <xnnpack/common.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/vunary.h>
17
18
xnn_f32_vrndu_ukernel__wasmsimd_addsub_x4(size_t n,const float * x,float * y,const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_f32_vrndu_ukernel__wasmsimd_addsub_x4(
20 size_t n,
21 const float* x,
22 float* y,
23 const union xnn_f32_rnd_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24 {
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27
28 const v128_t vsign_mask = wasm_f32x4_splat(-0.0f);
29 const v128_t vmagic_number = wasm_f32x4_splat(0x1.000000p+23f);
30 const v128_t vone = wasm_f32x4_splat(1.0f);
31 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
32 const v128_t vx = wasm_v128_load(x);
33 x += 4;
34
35 const v128_t vabsx = wasm_v128_andnot(vx, vsign_mask);
36 const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx));
37 const v128_t vrndabsx = wasm_f32x4_sub(wasm_f32x4_add(vabsx, vmagic_number), vmagic_number);
38 const v128_t vrndx = wasm_v128_bitselect(vx, vrndabsx, vrndmask);
39 const v128_t vadjmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx, vrndx));
40 const v128_t vadjrndx = wasm_f32x4_add(vrndx, vone);
41 const v128_t vy = wasm_v128_bitselect(vrndx, vadjrndx, vadjmask);
42
43 wasm_v128_store(y, vy);
44 y += 4;
45 }
46 if XNN_UNLIKELY(n != 0) {
47 const v128_t vx = wasm_v128_load(x);
48
49 const v128_t vabsx = wasm_v128_andnot(vx, vsign_mask);
50 const v128_t vrndmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vmagic_number, vabsx));
51 const v128_t vrndabsx = wasm_f32x4_sub(wasm_f32x4_add(vabsx, vmagic_number), vmagic_number);
52 const v128_t vrndx = wasm_v128_bitselect(vx, vrndabsx, vrndmask);
53 const v128_t vadjmask = wasm_v128_or(vsign_mask, wasm_f32x4_le(vx, vrndx));
54 const v128_t vadjrndx = wasm_f32x4_add(vrndx, vone);
55 v128_t vy = wasm_v128_bitselect(vrndx, vadjrndx, vadjmask);
56
57 if (n & (2 * sizeof(float))) {
58 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
59 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
60 y += 2;
61 }
62 if (n & (1 * sizeof(float))) {
63 *y = wasm_f32x4_extract_lane(vy, 0);
64 }
65 }
66 }
67