wehub-resource-sync
236 lines
5.3 KiB
ArmAsm
236 lines
5.3 KiB
ArmAsm
//
|
|
// MNNConvRunForLineDepthWiseUint8.S
|
|
// MNN
|
|
//
|
|
// Created by MNN on 2018/10/25.
|
|
// Copyright © 2018, Alibaba Group Holding Limited
|
|
//
|
|
|
|
#ifdef __aarch64__
|
|
|
|
#include "MNNAsmGlobal.h"
|
|
|
|
.text
|
|
.align 5
|
|
/*
|
|
struct MNN::ConstConvolutionParameter
|
|
{
|
|
size_t kw;
|
|
size_t kh;
|
|
size_t weight_y_step;
|
|
size_t dilate_x_step;
|
|
size_t dilate_y_step;
|
|
size_t stride_x_step;
|
|
int32_t output_multiplier;
|
|
int32_t output_shift_before;
|
|
int32_t output_shift_after;
|
|
int32_t output_offset;
|
|
int32_t output_activation_min;
|
|
int32_t output_activation_max;
|
|
};
|
|
*/
|
|
|
|
asm_function MNNConvRunForLineDepthWiseUint8
|
|
//void MNNConvRunForLineDepthWiseUint8(uint8_t* dst, const int16_t* src, const int16_t* weight, size_t width,
|
|
// MNN::ConstConvolutionParameter* parameters,
|
|
// const int32_t* bias_data)
|
|
|
|
//Auto: x0:dst, x1:src, x2:weight, x3:width
|
|
//x4: parameters, x5: bias_data
|
|
|
|
//Load All parameters
|
|
//v23.4s: output_multiplier, v22.4s: output_shift_before, v21.4s: output_shift_after
|
|
//v20.4s: output_offset, v19.4h: output_activation_min, v20.4h: output_activation_max, v17.4s: bias
|
|
|
|
ld1 {v17.4s}, [x5]
|
|
add x6, x4, #48
|
|
ld1 {v0.s}[0], [x6], #4
|
|
dup v23.4s, v0.s[0]
|
|
ld1 {v0.s}[0], [x6], #4
|
|
dup v22.4s, v0.s[0]
|
|
ld1 {v0.s}[0], [x6], #4
|
|
dup v21.4s, v0.s[0]
|
|
ld1 {v0.s}[0], [x6], #4
|
|
dup v20.4s, v0.s[0]
|
|
ld1 {v0.s}[0], [x6], #4
|
|
dup v19.4s, v0.s[0]
|
|
ld1 {v0.s}[0], [x6], #4
|
|
sqxtn v19.4h, v19.4s
|
|
dup v18.4s, v0.s[0]
|
|
sqxtn v18.4h, v18.4s
|
|
dup v19.2d, v19.d[0]
|
|
dup v18.2d, v18.d[0]
|
|
|
|
|
|
//Now x5 is no used, set x5 as stride_x_step
|
|
ldr x5, [x4, #40]
|
|
//x6:fw, x7:fh
|
|
ldr x6, [x4, #0]
|
|
ldr x7, [x4, #8]
|
|
|
|
//x8:dilate_x_step, x9:dilate_y_step
|
|
ldr x8, [x4, #24]
|
|
ldr x9, [x4, #32]
|
|
|
|
//dilate_y_step->dilate_y_step-fw*dilate_x_step
|
|
mul x12, x6, x8
|
|
sub x9, x9, x12
|
|
|
|
L8:
|
|
cmp x3, #8
|
|
blt L1
|
|
|
|
mov x12, #8
|
|
mul x10, x5, x12
|
|
|
|
LoopL8:
|
|
mov x13, x1
|
|
mov x14, x2
|
|
mov v0.16b, v17.16b
|
|
mov v1.16b,v17.16b
|
|
mov v2.16b,v17.16b
|
|
mov v3.16b,v17.16b
|
|
mov v4.16b,v17.16b
|
|
mov v5.16b,v17.16b
|
|
mov v6.16b,v17.16b
|
|
mov v7.16b,v17.16b
|
|
|
|
mov x11, x7
|
|
LoopL8FY:
|
|
mov x12, x6
|
|
LoopL8FX:
|
|
ld1 {v16.4h}, [x2], #8
|
|
ld1 {v24.4h}, [x1], x5
|
|
smlal v0.4s, v16.4h, v24.4h
|
|
ld1 {v25.4h}, [x1], x5
|
|
smlal v1.4s, v16.4h, v25.4h
|
|
ld1 {v24.4h}, [x1], x5
|
|
smlal v2.4s, v16.4h, v24.4h
|
|
ld1 {v25.4h}, [x1], x5
|
|
smlal v3.4s, v16.4h, v25.4h
|
|
ld1 {v24.4h}, [x1], x5
|
|
smlal v4.4s, v16.4h, v24.4h
|
|
ld1 {v25.4h}, [x1], x5
|
|
smlal v5.4s, v16.4h, v25.4h
|
|
ld1 {v24.4h}, [x1], x5
|
|
smlal v6.4s, v16.4h, v24.4h
|
|
ld1 {v25.4h}, [x1], x5
|
|
smlal v7.4s, v16.4h, v25.4h
|
|
sub x1, x1, x10
|
|
add x1, x1, x8
|
|
subs x12, x12, #1
|
|
bne LoopL8FX
|
|
subs x11, x11, #1
|
|
add x1, x1, x9
|
|
bne LoopL8FY
|
|
mov x1, x13
|
|
mov x2, x14
|
|
sqshl v0.4s, v0.4s, v22.4s
|
|
sqshl v1.4s, v1.4s, v22.4s
|
|
sqshl v2.4s, v2.4s, v22.4s
|
|
sqshl v3.4s, v3.4s, v22.4s
|
|
sqshl v4.4s, v4.4s, v22.4s
|
|
sqshl v5.4s, v5.4s, v22.4s
|
|
sqshl v6.4s, v6.4s, v22.4s
|
|
sqshl v7.4s, v7.4s, v22.4s
|
|
|
|
sqrdmulh v0.4s, v23.4s, v0.4s
|
|
sqrdmulh v1.4s, v23.4s, v1.4s
|
|
sqrdmulh v2.4s, v23.4s, v2.4s
|
|
sqrdmulh v3.4s, v23.4s, v3.4s
|
|
sqrdmulh v4.4s, v23.4s, v4.4s
|
|
sqrdmulh v5.4s, v23.4s, v5.4s
|
|
sqrdmulh v6.4s, v23.4s, v6.4s
|
|
sqrdmulh v7.4s, v23.4s, v7.4s
|
|
|
|
sqrshl v0.4s, v0.4s, v21.4s
|
|
sqrshl v1.4s, v1.4s, v21.4s
|
|
sqrshl v2.4s, v2.4s, v21.4s
|
|
sqrshl v3.4s, v3.4s, v21.4s
|
|
sqrshl v4.4s, v4.4s, v21.4s
|
|
sqrshl v5.4s, v5.4s, v21.4s
|
|
sqrshl v6.4s, v6.4s, v21.4s
|
|
sqrshl v7.4s, v7.4s, v21.4s
|
|
|
|
add v0.4s, v0.4s, v20.4s
|
|
add v1.4s, v1.4s, v20.4s
|
|
add v2.4s, v2.4s, v20.4s
|
|
add v3.4s, v3.4s, v20.4s
|
|
add v4.4s, v4.4s, v20.4s
|
|
add v5.4s, v5.4s, v20.4s
|
|
add v6.4s, v6.4s, v20.4s
|
|
add v7.4s, v7.4s, v20.4s
|
|
|
|
add x1, x1, x10
|
|
sqxtn v0.4h, v0.4s
|
|
sqxtn2 v0.8h, v1.4s
|
|
smax v0.8h, v0.8h, v19.8h
|
|
sqxtn v1.4h, v2.4s
|
|
sqxtn2 v1.8h, v3.4s
|
|
smin v0.8h, v0.8h, v18.8h
|
|
sqxtn v2.4h, v4.4s
|
|
sqxtn2 v2.8h, v5.4s
|
|
sqxtn v3.4h, v6.4s
|
|
sqxtn2 v3.8h, v7.4s
|
|
|
|
smax v1.8h, v1.8h, v19.8h
|
|
smax v2.8h, v2.8h, v19.8h
|
|
smax v3.8h, v3.8h, v19.8h
|
|
|
|
smin v1.8h, v1.8h, v18.8h
|
|
smin v2.8h, v2.8h, v18.8h
|
|
smin v3.8h, v3.8h, v18.8h
|
|
|
|
uqxtn v0.8b, v0.8h
|
|
uqxtn2 v0.16b, v1.8h
|
|
uqxtn v1.8b, v2.8h
|
|
uqxtn2 v1.16b, v3.8h
|
|
|
|
st1 {v0.4s, v1.4s}, [x0], #32
|
|
sub x3, x3, #8
|
|
cmp x3, #8
|
|
bge LoopL8
|
|
|
|
|
|
L1:
|
|
cmp x3, #0
|
|
beq End
|
|
|
|
LoopL1:
|
|
mov x13, x1
|
|
mov x14, x2
|
|
mov v0.16b,v17.16b
|
|
mov x11, x7
|
|
LoopL1FY:
|
|
mov x12, x6
|
|
LoopL1FX:
|
|
ld1 {v1.4h}, [x1], x8
|
|
ld1 {v2.4h}, [x2], #8
|
|
smlal v0.4s, v1.4h, v2.4h
|
|
subs x12, x12, #1
|
|
bne LoopL1FX
|
|
subs x11, x11, #1
|
|
add x1, x1, x9
|
|
bne LoopL1FY
|
|
mov x1, x13
|
|
mov x2, x14
|
|
sqshl v0.4s, v0.4s, v22.4s
|
|
sqrdmulh v0.4s, v23.4s, v0.4s
|
|
sqrshl v0.4s, v0.4s, v21.4s
|
|
add v0.4s, v0.4s, v20.4s
|
|
|
|
add x1, x1, x5
|
|
sqxtn v0.4h, v0.4s
|
|
smax v0.4h, v0.4h, v19.4h
|
|
smin v0.4h, v0.4h, v18.4h
|
|
uqxtn v0.8b, v0.8h
|
|
st1 {v0.s}[0], [x0], #4
|
|
subs x3, x3, #1
|
|
bne LoopL1
|
|
|
|
End:
|
|
|
|
ret
|
|
|
|
#endif
|