1 // Auto-generated file. Do not edit!
2 // Template: src/f32-vbinary/vop-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
xnn_f32_vsub_relu_ukernel__wasmsimd_x16(size_t n,const float * a,const float * b,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vsub_relu_ukernel__wasmsimd_x16(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
23 const union xnn_f32_relu_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
31 const v128_t vzero = wasm_f32x4_splat(0.0f);
32
33 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
34 const v128_t va0123 = wasm_v128_load(a);
35 const v128_t va4567 = wasm_v128_load(a + 4);
36 const v128_t va89AB = wasm_v128_load(a + 8);
37 const v128_t vaCDEF = wasm_v128_load(a + 12);
38 a += 16;
39
40 const v128_t vb0123 = wasm_v128_load(b);
41 const v128_t vb4567 = wasm_v128_load(b + 4);
42 const v128_t vb89AB = wasm_v128_load(b + 8);
43 const v128_t vbCDEF = wasm_v128_load(b + 12);
44 b += 16;
45
46 v128_t vy0123 = wasm_f32x4_sub(va0123, vb0123);
47 v128_t vy4567 = wasm_f32x4_sub(va4567, vb4567);
48 v128_t vy89AB = wasm_f32x4_sub(va89AB, vb89AB);
49 v128_t vyCDEF = wasm_f32x4_sub(vaCDEF, vbCDEF);
50
51
52 vy0123 = wasm_i32x4_max(vy0123, vzero);
53 vy4567 = wasm_i32x4_max(vy4567, vzero);
54 vy89AB = wasm_i32x4_max(vy89AB, vzero);
55 vyCDEF = wasm_i32x4_max(vyCDEF, vzero);
56
57 wasm_v128_store(y, vy0123);
58 wasm_v128_store(y + 4, vy4567);
59 wasm_v128_store(y + 8, vy89AB);
60 wasm_v128_store(y + 12, vyCDEF);
61 y += 16;
62 }
63 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
64 const v128_t va = wasm_v128_load(a);
65 a += 4;
66
67 const v128_t vb = wasm_v128_load(b);
68 b += 4;
69
70 v128_t vy = wasm_f32x4_sub(va, vb);
71
72 vy = wasm_i32x4_max(vy, vzero);
73
74 wasm_v128_store(y, vy);
75 y += 4;
76 }
77 if XNN_UNLIKELY(n != 0) {
78 const v128_t va = wasm_v128_load(a);
79 const v128_t vb = wasm_v128_load(b);
80
81 v128_t vy = wasm_f32x4_sub(va, vb);
82
83 vy = wasm_i32x4_max(vy, vzero);
84
85 if (n & (2 * sizeof(float))) {
86 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
87 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
88 y += 2;
89 }
90 if (n & (1 * sizeof(float))) {
91 *y = wasm_f32x4_extract_lane(vy, 0);
92 }
93 }
94 }
95