MNN/source/backend/cpu/arm/arm32/MNNMatrixCopyUnit.S

71 lines
1.1 KiB
ArmAsm

//
// MNNMatrixCopyUnit.S
// MNN
//
// Created by MNN on 2020/01/21.
// Copyright © 2018, Alibaba Group Holding Limited
//
#ifdef __arm__
#ifndef __aarch64__
#include "MNNAsmGlobal.h"
.text
.align 5
asm_function MNNMatrixCopyUnit
//void MNNMatrixCopyUnit(float* C, const float* A, size_t cStride, size_t aStride, size_t height)
//Auto: r0: C, r1:A, r2:cStride
//r3:aStride, r4:height
push {r4, lr}
ldr r4, [sp, #8]
mov r12, #4 //sizeof(float)
mul r2, r12, r2
mul r3, r12, r3
sub r2, r2, #128
sub r3, r3, #128
subs r4, r4, #1
// Unit = 8 for armv7a
vldm r1!, {d16-d23}
beq LoopYEnd
LoopY:
vst1.64 {q8}, [r0]!
PLD [r1, #0xC0]
vldm r1!, {d24-d31}
vst1.64 {q9}, [r0]!
vst1.64 {q10}, [r0]!
vst1.64 {q11}, [r0]!
vst1.64 {q12}, [r0]!
add r1, r1, r3
vst1.64 {q13}, [r0]!
PLD [r1, #0xC0]
vldm r1!, {d16-d23}
vst1.64 {q14}, [r0]!
vst1.64 {q15}, [r0]!
add r0, r0, r2
subs r4, r4, #1
bne LoopY
LoopYEnd:
vst1.64 {q8, q9}, [r0]!
vld1.64 {q12, q13}, [r1]!
vst1.64 {q10, q11}, [r0]!
vld1.64 {q14, q15}, [r1]
vst1.64 {q12, q13}, [r0]!
vst1.64 {q14, q15}, [r0]
pop {r4, pc}
#endif
#endif