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_minmax_ukernel__wasmsimd_arm_x8(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_vsub_minmax_ukernel__wasmsimd_arm_x8(
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(&params->scalar.min);
32   const v128_t vy_max = wasm_v32x4_load_splat(&params->scalar.max);
33 
34   for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
35     const v128_t va0123 = wasm_v128_load(a);
36     const v128_t va4567 = wasm_v128_load(a + 4);
37     a += 8;
38 
39     const v128_t vb0123 = wasm_v128_load(b);
40     const v128_t vb4567 = wasm_v128_load(b + 4);
41     b += 8;
42 
43     v128_t vy0123 = wasm_f32x4_sub(va0123, vb0123);
44     v128_t vy4567 = wasm_f32x4_sub(va4567, vb4567);
45 
46 
47     vy0123 = wasm_f32x4_max(vy0123, vy_min);
48     vy4567 = wasm_f32x4_max(vy4567, vy_min);
49 
50     vy0123 = wasm_f32x4_min(vy0123, vy_max);
51     vy4567 = wasm_f32x4_min(vy4567, vy_max);
52 
53     wasm_v128_store(y, vy0123);
54     wasm_v128_store(y + 4, vy4567);
55     y += 8;
56   }
57   for (; n >= 4 * sizeof(float); n -= 4 * sizeof(float)) {
58     const v128_t va = wasm_v128_load(a);
59     a += 4;
60 
61     const v128_t vb = wasm_v128_load(b);
62     b += 4;
63 
64     v128_t vy = wasm_f32x4_sub(va, vb);
65 
66     vy = wasm_f32x4_max(vy, vy_min);
67     vy = wasm_f32x4_min(vy, vy_max);
68 
69     wasm_v128_store(y, vy);
70     y += 4;
71   }
72   if XNN_UNLIKELY(n != 0) {
73     const v128_t va = wasm_v128_load(a);
74     const v128_t vb = wasm_v128_load(b);
75 
76     v128_t vy = wasm_f32x4_sub(va, vb);
77 
78     vy = wasm_f32x4_max(vy, vy_min);
79     vy = wasm_f32x4_min(vy, vy_max);
80 
81     if (n & (2 * sizeof(float))) {
82       *((double*) y) = wasm_f64x2_extract_lane(vy, 0);
83       vy = wasm_v32x4_shuffle(vy, vy, 2, 3, 2, 3);
84       y += 2;
85     }
86     if (n & (1 * sizeof(float))) {
87       *y = wasm_f32x4_extract_lane(vy, 0);
88     }
89   }
90 }
91