blob: a83c82bb769804e9a9a8c5271d904a01a7559834 [file] [log] [blame]
Frank Barchard9c7308f2020-08-31 17:03:01 -07001// Auto-generated file. Do not edit!
2// Template: src/f32-vbinary/vopc-wasmsimd.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/vbinary.h>
16
17
18void xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
23 const union xnn_f32_default_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
24{
25 assert(n != 0);
26 assert(n % sizeof(float) == 0);
27 assert(a != NULL);
28 assert(b != NULL);
29 assert(y != NULL);
30
Marat Dukhanee029b22021-06-30 12:47:02 -070031 const v128_t vb = wasm_v128_load32_splat(b);
Frank Barchard9c7308f2020-08-31 17:03:01 -070032 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
33 const v128_t va0123 = wasm_v128_load(a);
34 const v128_t va4567 = wasm_v128_load(a + 4);
35 const v128_t va89AB = wasm_v128_load(a + 8);
36 const v128_t vaCDEF = wasm_v128_load(a + 12);
37 a += 16;
38
39 v128_t vy0123 = wasm_f32x4_max(va0123, vb);
40 v128_t vy4567 = wasm_f32x4_max(va4567, vb);
41 v128_t vy89AB = wasm_f32x4_max(va89AB, vb);
42 v128_t vyCDEF = wasm_f32x4_max(vaCDEF, vb);
43
44
45
46 wasm_v128_store(y, vy0123);
47 wasm_v128_store(y + 4, vy4567);
48 wasm_v128_store(y + 8, vy89AB);
49 wasm_v128_store(y + 12, vyCDEF);
50 y += 16;
51 }
52 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
53 const v128_t va = wasm_v128_load(a);
54 a += 4;
55
56 v128_t vy = wasm_f32x4_max(va, vb);
57
58
59 wasm_v128_store(y, vy);
60 y += 4;
61 }
62 if XNN_UNLIKELY(n != 0) {
63 const v128_t va = wasm_v128_load(a);
64
65 v128_t vy = wasm_f32x4_max(va, vb);
66
67
68 if (n & (2 * sizeof(float))) {
69 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
70 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
71 y += 2;
72 }
73 if (n & (1 * sizeof(float))) {
74 *y = wasm_f32x4_extract_lane(vy, 0);
75 }
76 }
77}