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_vdiv_minmax_ukernel__wasmsimd_x86_x16(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_vdiv_minmax_ukernel__wasmsimd_x86_x16(
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_OOB_READS
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_v128_load64_splat(params->wasmsimd.min);
32 const v128_t vy_max = wasm_v128_load64_splat(params->wasmsimd.max);
33
34 for (; n >= 16 * sizeof(float); n -= 16 * sizeof(float)) {
35 const v128_t va0123 = wasm_v128_load(a);
36 const v128_t va4567 = wasm_v128_load(a + 4);
37 const v128_t va89AB = wasm_v128_load(a + 8);
38 const v128_t vaCDEF = wasm_v128_load(a + 12);
39 a += 16;
40
41 const v128_t vb0123 = wasm_v128_load(b);
42 const v128_t vb4567 = wasm_v128_load(b + 4);
43 const v128_t vb89AB = wasm_v128_load(b + 8);
44 const v128_t vbCDEF = wasm_v128_load(b + 12);
45 b += 16;
46
47 v128_t vy0123 = wasm_f32x4_div(va0123, vb0123);
48 v128_t vy4567 = wasm_f32x4_div(va4567, vb4567);
49 v128_t vy89AB = wasm_f32x4_div(va89AB, vb89AB);
50 v128_t vyCDEF = wasm_f32x4_div(vaCDEF, vbCDEF);
51
52
53 vy0123 = wasm_f32x4_pmax(vy_min, vy0123);
54 vy4567 = wasm_f32x4_pmax(vy_min, vy4567);
55 vy89AB = wasm_f32x4_pmax(vy_min, vy89AB);
56 vyCDEF = wasm_f32x4_pmax(vy_min, vyCDEF);
57
58 vy0123 = wasm_f32x4_pmin(vy_max, vy0123);
59 vy4567 = wasm_f32x4_pmin(vy_max, vy4567);
60 vy89AB = wasm_f32x4_pmin(vy_max, vy89AB);
61 vyCDEF = wasm_f32x4_pmin(vy_max, vyCDEF);
62
63 wasm_v128_store(y, vy0123);
64 wasm_v128_store(y + 4, vy4567);
65 wasm_v128_store(y + 8, vy89AB);
66 wasm_v128_store(y + 12, vyCDEF);
67 y += 16;
68 }
69 for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
70 const v128_t va = wasm_v128_load(a);
71 a += 4;
72
73 const v128_t vb = wasm_v128_load(b);
74 b += 4;
75
76 v128_t vy = wasm_f32x4_div(va, vb);
77
78 vy = wasm_f32x4_pmax(vy_min, vy);
79 vy = wasm_f32x4_pmin(vy_max, vy);
80
81 wasm_v128_store(y, vy);
82 y += 4;
83 }
84 if XNN_UNLIKELY(n != 0) {
85 const v128_t va = wasm_v128_load(a);
86 const v128_t vb = wasm_v128_load(b);
87
88 v128_t vy = wasm_f32x4_div(va, vb);
89
90 vy = wasm_f32x4_pmax(vy_min, vy);
91 vy = wasm_f32x4_pmin(vy_max, vy);
92
93 if (n & (2 * sizeof(float))) {
94 *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
95 vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
96 y += 2;
97 }
98 if (n & (1 * sizeof(float))) {
99 *y = wasm_f32x4_extract_lane(vy, 0);
100 }
101 }
102 }
103