• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Auto-generated file. Do not edit!
2 //   Template: src/f32-relu/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/vunary.h>
15 #include <xnnpack/common.h>
16 
17 
xnn_f32_relu_ukernel__wasmsimd_x16(size_t n,const float * x,float * y,const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_relu_ukernel__wasmsimd_x16(
19     size_t n,
20     const float* x,
21     float* y,
22     const union xnn_f32_relu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
23 {
24   assert(n != 0);
25   assert(n % sizeof(float) == 0);
26   assert(x != NULL);
27   assert(y != NULL);
28 
29   const v128_t vzero = wasm_f32x4_splat(0.0f);
30 
31   for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
32     v128_t vacc0123 = wasm_v128_load(x);
33     v128_t vacc4567 = wasm_v128_load(x + 4);
34     v128_t vacc89AB = wasm_v128_load(x + 8);
35     v128_t vaccCDEF = wasm_v128_load(x + 12);
36     x += 16;
37 
38     vacc0123 = wasm_i32x4_max(vacc0123, vzero);
39     vacc4567 = wasm_i32x4_max(vacc4567, vzero);
40     vacc89AB = wasm_i32x4_max(vacc89AB, vzero);
41     vaccCDEF = wasm_i32x4_max(vaccCDEF, vzero);
42 
43     wasm_v128_store(y, vacc0123);
44     wasm_v128_store(y + 4, vacc4567);
45     wasm_v128_store(y + 8, vacc89AB);
46     wasm_v128_store(y + 12, vaccCDEF);
47     y += 16;
48   }
49   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
50     v128_t vacc = wasm_v128_load(x);
51     x += 4;
52 
53     vacc = wasm_i32x4_max(vacc, vzero);
54 
55     wasm_v128_store(y, vacc);
56     y += 4;
57   }
58   if XNN_UNLIKELY(n != 0) {
59     v128_t vacc = wasm_v128_load(x);
60 
61     vacc = wasm_i32x4_max(vacc, vzero);
62 
63     if (n & (2 * sizeof(float))) {
64       *((double*) y) = wasm_f64x2_extract_lane(vacc, 0);
65       vacc = wasm_v32x4_shuffle(vacc, vacc, 2, 3, 2, 3);
66       y += 2;
67     }
68     if (n & (1 * sizeof(float))) {
69       *y = wasm_f32x4_extract_lane(vacc, 0);
70     }
71   }
72 }
73