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_vadd_minmax_ukernel__wasmsimd_x86_x4(size_t n,const float * a,const float * b,float * y,const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x4(
19 size_t n,
20 const float* a,
21 const float* b,
22 float* y,
23 const union xnn_f32_minmax_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 vy_min = wasm_v32x4_load_splat(¶ms->scalar.min);
32 const v128_t vy_max = wasm_v32x4_load_splat(¶ms->scalar.max);
33
34 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
35 const v128_t va0123 = wasm_v128_load(a);
36 a += 4;
37
38 const v128_t vb0123 = wasm_v128_load(b);
39 b += 4;
40
41 v128_t vy0123 = wasm_f32x4_add(va0123, vb0123);
42
43
44 const v128_t vltmask0123 = wasm_f32x4_lt(vy0123, vy_min);
45
46 const v128_t vngtmask0123 = wasm_f32x4_le(vy0123, vy_max);
47 vy0123 = wasm_v128_bitselect(vy_min, vy0123, vltmask0123);
48
49 vy0123 = wasm_v128_bitselect(vy0123, vy_max, vngtmask0123);
50
51 wasm_v128_store(y, vy0123);
52 y += 4;
53 }
54 if XNN_UNLIKELY(n != 0) {
55 const v128_t va = wasm_v128_load(a);
56 const v128_t vb = wasm_v128_load(b);
57
58 v128_t vy = wasm_f32x4_add(va, vb);
59
60 const v128_t vltmask = wasm_f32x4_lt(vy, vy_min);
61 const v128_t vngtmask = wasm_f32x4_le(vy, vy_max);
62 vy = wasm_v128_bitselect(vy_min, vy, vltmask);
63 vy = wasm_v128_bitselect(vy, vy_max, vngtmask);
64
65 if (n & (2 * sizeof(float))) {
66 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
67 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
68 y += 2;
69 }
70 if (n & (1 * sizeof(float))) {
71 *y = wasm_f32x4_extract_lane(vy, 0);
72 }
73 }
74 }
75