1//
2//  NEON_MNNConvRunForUnitDepthWise_BF16.S
3//  MNN
4//
5//  Created by MNN on 2021/03/09.
6//  Copyright © 2018-2021 Alibaba Group Holding Limited
7//
8
9#ifdef __aarch64__
10
11#include "MNNAsmGlobal.h"
12
13.text
14.align 5
15
16asm_function NEON_MNNConvRunForUnitDepthWise_BF16
17//void NEON_MNNConvRunForUnitDepthWise_BF16(float* dst, const float* src, const float* weight, size_t fw, size_t fh, size_t weight_y_step, size_t dilate_x_step, size_t dilate_y_step)
18
19//Auto: x0:dst, x1:src, x2:weight, x3:fw
20//x4:fh, x5:weight_y_step, x6:dilate_x_step, x7:dilate_y_step
21
22cmp x3, #0
23movi v0.4s, #0
24beq UnitEnd
25cmp x4, #0
26beq UnitEnd
27
28mov x9, #2
29mul x5, x9, x5 // x5(weight_y_step in byte) = sizeof(int16_t) * weight_y_step
30mul x6, x9, x6 // x6(dilate_x_step in byte) = sizeof(int16_t) * dilate_x_step
31mul x7, x9, x7 // x7(dilate_y_step in byte) = sizeof(int16_t) * dilate_y_step
32
33//dilate_y_step -> dilate_y_step - dilate_x_step*fw
34mul x9, x3, x6
35sub x7, x7, x9 // because x1 has already been auto-increased at 'ld1 {v1.4h}, [x1], x6', here we should rewind by x6*fw
36
37//weight_y_step -> weight_y_step - 4*sizeof(int16_t)*fw
38mov x9, #8
39mul x9, x3, x9
40sub x5, x5, x9
41
42
43UnitLoopH:
44mov x9, x3
45UnitLoopW:
46ld1 {v1.4h}, [x1], x6
47ld1 {v2.4h}, [x2], #8 // 4 * sizeof(int16_t)
48shll v1.4s, v1.4h, #16
49shll v2.4s, v2.4h, #16
50
51fmla v0.4s, v1.4s, v2.4s
52subs x9, x9, #1
53bne UnitLoopW
54subs x4, x4, #1
55add x1, x1, x7
56add x2, x2, x5
57bne UnitLoopH
58
59
60UnitEnd:
61shrn v0.4h, v0.4s, #16
62st1 {v0.4h}, [x0]
63
64ret
65
66#endif
67