ZYNQ学习之路——DMA PS(PL330)基础

本文转载自：亦梦云烟微信公众号

1. 什么是DMA
DMA是直接内存访问(Direct Memory Access)，DMA引擎可以将数据从一个地方传输到另一个地方，在传输过程中不经过CPU的控制。最简单的DMA用法是将数据从内存的一个区域搬运到另一个区域。DMA也可以将外设的数据(如ADC)搬运到内存中，或者将内存数据搬运到外设中(如DAC)。

Zynq-7000系列器件PS端的DMA控制器采用ARM的IP核DMA-330（PL-330）实现。

开发环境

Windows 10 64位

Vivado 2018.2

XC7Z010-1-CLG400

1.1 结构特点
DMA控制器具有以下的特点：

1. 8个独立的通道，4个可用于PL—PS间数据管理，每个通道有1024Byte的MFIFO；
2. 使用CPU_2x 时钟搬运数据，CPU_2x = （CPU frq/6）*2；
3. 执行自定义内存区域内的DMA指令运行DMA；
4. AHB控制寄存器支持安全和非安全模式；
5. 每个通道内置4字Cache；
6. 可以访问SoC的以下映射物理地址：

DDR、OCM、PL、Linear QSPI Read、SMC和M_AXI_GP设备，访问设备的互联结构如图1所示。

图1 DMA PS结构示意图

1.2 Zynq 访问互联结构图
从图1可以看出DMA控制器可以访问连接到Central Interconnect上的所有设备，并提供了四个通道的外设管理接口可用于控制PL的数据搬运。

Zynq系列器件中DMA控制器采用ARM PL-330 IP和r1p1版，结构框图如图2所示。

图2 ZYNQ DMA控制器结构示意图

如图2所示，DMA控制器由指令加速引擎，AXI Master数据接口，AXI APB寄存器访问接口以及可以连接到PL的外设请求接口，数据缓冲FIFO和控制及状态产生单元组成。

从图2可以看到，DMA PL330的设计思想是：DMA控制器通过DMA指令执行引擎执行自己的指令，并将执行状态通过APB总线和中断等形式反馈给CPU，达到数据搬运不占用CPU的目的。

DMA控制器共有八个通道，其中四个通道负责互联到Central Interconnectcun存储单元上的数据搬运；四个数据通道为外设请求接口，可用于PL AXI互联接口的数据访问管理。

每个DMA通道都执行自己的指令，拥有自己的独立线程，通道间互不影响。指令执行引擎有自己独立的Cache线。

2. 实例测试
首先构建AXI DMA例程使用的硬件环境，如图3所示，ZYNQ通过GP0端口读取Block RAM数据。

图3 Block RAM硬件结构

2.1 测试硬件完整性
首先使用SDK测试硬件的完成整性，编写如下代码测试BRAM读写情况。

#include <stdio.h>
#include "platform.h"
#include "xil_printf.h"
#include "xtime_l.h"
#include "xparameters.h"
 
void TC_BRAM();
#define RAM_W	XPAR_AXI_BRAM_CTRL_0_S_AXI_BASEADDR
#define RAM_R	XPAR_AXI_BRAM_CTRL_1_S_AXI_BASEADDR
 
int main()
{
    init_platform();
 
    TC_BRAM();
 
    cleanup_platform();
    return 0;
}
 
void TC_BRAM()
{
	printf("test for block RAM\n");
	XTime tb, te;
	double dt = 0.0;
	XTime_SetTime(0);
 
 
	for(int i=0; i<4*1024; i++)
	{
		*(int *)(RAM_W+4*i) = i;
	}
	XTime_GetTime(&tb);
	for(int i=0; i<4*1024; i++)
	{
		if(*(int *)(RAM_R+4*i) != i)
		{
			printf("Test Failed\n");
			break;
		}
	}
 
	XTime_GetTime(&te);
	printf("Test pass\n");
	dt = (te-tb)*1000000/COUNTS_PER_SECOND;
	printf("%fus\n",dt);
	printf("test for block RAM end!\n");
}

在串口终端中如果没有输出"Test Failed"则说明硬件设计无误。

2.2 测试内存读取速度
在使用DMA之前，首先在不使用DMA的情况下测试内存读取的速度。本例程首先写入0~4095，然后全部读取出来。

指针循环访问：

void TC_PointerSpeed()
{
	XTime tb, te;
	double dt = 0.0;
	int a[4*1024];
	XTime_SetTime(0);
    for(int i=0; i<4*1024; i++)
    {
        *(int *)(RAM_W+4*i) = i;
    }
    XTime_GetTime(&tb);
    for(int i=0; i<4*1024; i++)
    {
        a[i] = *(int *)(RAM_R+4*i);
    }
    XTime_GetTime(&te);
    dt = (te-tb)*1000000/COUNTS_PER_SECOND;
    printf("%fus\n",dt);
}

memcpy:

void TC_MemcpySpeed()
{
	XTime tb, te;
	double dt = 0.0;
	int a[4*1024];
	XTime_SetTime(0);
 
    for(int i=0; i<4*1024; i++)
    {
        *(int *)(RAM_W+4*i) = i;
    }
    XTime_GetTime(&tb);
    memcpy(a, (void*)RAM_R, 4*1024*4);
    XTime_GetTime(&te);
    dt = (te-tb)*1000000/COUNTS_PER_SECOND;
    printf("%fus\n",dt);
}

速度如下表所示。

可以看出使用CPU进行的内存复制效率非常低。

3. DMAPS应用
3.1 编程模型

本文不考虑外设请求接口，DMA控制器编程分为以下几个部分：

1. DMA控制器初始化；
2. 组织DMA引擎执行代码；
3. 启动或停止DMA传输；
4. 异常处理。

官方例程在Vivado安装路径下:

Vivado2018.2\SDK\2018.2\data\embeddedsw\XilinxProcessorIPLib\drivers\dmaps_v2_3\examples

#include
#include "platform.h"
#include "xil_printf.h"
#include "sleep.h"
#include "xparameters.h"
#include "xil_types.h"
#include "xil_assert.h"
#include "xil_io.h"
#include "xil_exception.h"
#include "xil_cache.h"
#include "xil_printf.h"
#include "xscugic.h"
#include "xdmaps.h"

/************************** Constant Definitions *****************************/
/*
* The following constants map to the XPAR parameters created in the
* xparameters.h file. They are defined here such that a user can easily
* change all the needed parameters in one place.
*/
#define DMA_DEVICE_ID XPAR_XDMAPS_1_DEVICE_ID
#define INTC_DEVICE_ID XPAR_SCUGIC_SINGLE_DEVICE_ID

#define DMA_DONE_INTR_0 XPAR_XDMAPS_0_DONE_INTR_0
#define DMA_DONE_INTR_1 XPAR_XDMAPS_0_DONE_INTR_1
#define DMA_DONE_INTR_2 XPAR_XDMAPS_0_DONE_INTR_2
#define DMA_DONE_INTR_3 XPAR_XDMAPS_0_DONE_INTR_3
#define DMA_DONE_INTR_4 XPAR_XDMAPS_0_DONE_INTR_4
#define DMA_DONE_INTR_5 XPAR_XDMAPS_0_DONE_INTR_5
#define DMA_DONE_INTR_6 XPAR_XDMAPS_0_DONE_INTR_6
#define DMA_DONE_INTR_7 XPAR_XDMAPS_0_DONE_INTR_7
#define DMA_FAULT_INTR XPAR_XDMAPS_0_FAULT_INTR

#define TEST_ROUNDS 1 /* Number of loops that the Dma transfers run.*/
#define DMA_LENGTH 1024 /* Length of the Dma Transfers */
#define TIMEOUT_LIMIT 0x2000 /* Loop count for timeout */

/************************** Function Prototypes ******************************/

int XDmaPs_Example_W_Intr(XScuGic *GicPtr, u16 DeviceId);
int SetupInterruptSystem(XScuGic *GicPtr, XDmaPs *DmaPtr);
void DmaDoneHandler(unsigned int Channel, XDmaPs_Cmd *DmaCmd,
void *CallbackRef);

/************************** Variable Definitions *****************************/
#ifdef __ICCARM__
#pragma data_alignment=32
static int Src[DMA_LENGTH];
static int Dst[DMA_LENGTH];
#pragma data_alignment=4
#else
static int Src[DMA_LENGTH] __attribute__ ((aligned (32)));
static int Dst[DMA_LENGTH] __attribute__ ((aligned (32)));
#endif

XDmaPs DmaInstance;
#ifndef TESTAPP_GEN
XScuGic GicInstance;
#endif

#ifndef TESTAPP_GEN
int main(void)
{
int Status;

Status = XDmaPs_Example_W_Intr(&GicInstance,DMA_DEVICE_ID);
if (Status != XST_SUCCESS) {
xil_printf("Error: XDMaPs_Example_W_Intr failed\r\n");
return XST_FAILURE;
}

xil_printf("Successfully ran XDMaPs_Example_W_Intr\r\n");
return XST_SUCCESS;

}
#endif

/*****************************************************************************/
/**
*
* Interrupt Example to test the DMA.
*
* @param DeviceId is the Device ID of the DMA controller.
*
* @return XST_SUCCESS to indicate success, otherwise XST_FAILURE.
*
* @note None.
*
****************************************************************************/
int XDmaPs_Example_W_Intr(XScuGic *GicPtr, u16 DeviceId)
{
int Index;
unsigned int Channel = 0;
int Status;
int TestStatus;
int TestRound;
int TimeOutCnt;
volatile int Checked[XDMAPS_CHANNELS_PER_DEV];
XDmaPs_Config *DmaCfg;
XDmaPs *DmaInst = &DmaInstance;
XDmaPs_Cmd DmaCmd;

memset(&DmaCmd, 0, sizeof(XDmaPs_Cmd));

DmaCmd.ChanCtrl.SrcBurstSize = 4;
DmaCmd.ChanCtrl.SrcBurstLen = 4;
DmaCmd.ChanCtrl.SrcInc = 1;
DmaCmd.ChanCtrl.DstBurstSize = 4;
DmaCmd.ChanCtrl.DstBurstLen = 4;
DmaCmd.ChanCtrl.DstInc = 1;
DmaCmd.BD.SrcAddr = (u32) Src;
DmaCmd.BD.DstAddr = (u32) Dst;
DmaCmd.BD.Length = DMA_LENGTH * sizeof(int);

/*
* Initialize the DMA Driver
*/
DmaCfg = XDmaPs_LookupConfig(DeviceId);
if (DmaCfg == NULL) {
return XST_FAILURE;
}

Status = XDmaPs_CfgInitialize(DmaInst,
DmaCfg,
DmaCfg->BaseAddress);
if (Status != XST_SUCCESS) {
return XST_FAILURE;
}

/*
* Setup the interrupt system.
*/
Status = SetupInterruptSystem(GicPtr, DmaInst);
if (Status != XST_SUCCESS) {
return XST_FAILURE;
}

TestStatus = XST_SUCCESS;

for (TestRound = 0; TestRound < TEST_ROUNDS; TestRound++) {
xil_printf("Test round %d\r\n", TestRound);
for (Channel = 0;
Channel < XDMAPS_CHANNELS_PER_DEV;
Channel++) {

/* Initialize source */
for (Index = 0; Index < DMA_LENGTH; Index++)
Src[Index] = DMA_LENGTH - Index;

/* Clear destination */
for (Index = 0; Index < DMA_LENGTH; Index++)
Dst[Index] = 0;

Checked[Channel] = 0;

/* Set the Done interrupt handler */
XDmaPs_SetDoneHandler(DmaInst,
Channel,
DmaDoneHandler,
(void *)Checked);

Status = XDmaPs_Start(DmaInst, Channel, &DmaCmd, 0);
if (Status != XST_SUCCESS) {
return XST_FAILURE;
}

TimeOutCnt = 0;

/* Now the DMA is done */
while (!Checked[Channel]
&& TimeOutCnt < TIMEOUT_LIMIT) {
TimeOutCnt++;
}

if (TimeOutCnt >= TIMEOUT_LIMIT) {
TestStatus = XST_FAILURE;
}

if (Checked[Channel] < 0) {
/* DMA controller failed */
TestStatus = XST_FAILURE;
}
}
}

return TestStatus;

}

/******************************************************************************/
/**
*
* This function connects the interrupt handler of the interrupt controller to
* the processor. This function is seperate to allow it to be customized for
* each application. Each processor or RTOS may require unique processing to
* connect the interrupt handler.
*
* @param GicPtr is the GIC instance pointer.
* @param DmaPtr is the DMA instance pointer.
*
* @return None.
*
* @note None.
*
****************************************************************************/
int SetupInterruptSystem(XScuGic *GicPtr, XDmaPs *DmaPtr)
{
int Status;
#ifndef TESTAPP_GEN
XScuGic_Config *GicConfig;

Xil_ExceptionInit();

/*
* Initialize the interrupt controller driver so that it is ready to
* use.
*/
GicConfig = XScuGic_LookupConfig(INTC_DEVICE_ID);
if (NULL == GicConfig) {
return XST_FAILURE;
}

Status = XScuGic_CfgInitialize(GicPtr, GicConfig,
GicConfig->CpuBaseAddress);
if (Status != XST_SUCCESS) {
return XST_FAILURE;
}

/*
* Connect the interrupt controller interrupt handler to the hardware
* interrupt handling logic in the processor.
*/
Xil_ExceptionRegisterHandler(XIL_EXCEPTION_ID_IRQ_INT,
(Xil_ExceptionHandler)XScuGic_InterruptHandler,
GicPtr);
#endif
/*
* Connect the device driver handlers that will be called when an interrupt
* for the device occurs, the device driver handler performs the specific
* interrupt processing for the device
*/

/*
* Connect the Fault ISR
*/
Status = XScuGic_Connect(GicPtr,
DMA_FAULT_INTR,
(Xil_InterruptHandler)XDmaPs_FaultISR,
(void *)DmaPtr);
if (Status != XST_SUCCESS) {
return XST_FAILURE;
}

/*
* Connect the Done ISR for all 8 channels of DMA 0
*/
Status = XScuGic_Connect(GicPtr,
DMA_DONE_INTR_0,
(Xil_InterruptHandler)XDmaPs_DoneISR_0,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_1,
(Xil_InterruptHandler)XDmaPs_DoneISR_1,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_2,
(Xil_InterruptHandler)XDmaPs_DoneISR_2,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_3,
(Xil_InterruptHandler)XDmaPs_DoneISR_3,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_4,
(Xil_InterruptHandler)XDmaPs_DoneISR_4,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_5,
(Xil_InterruptHandler)XDmaPs_DoneISR_5,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_6,
(Xil_InterruptHandler)XDmaPs_DoneISR_6,
(void *)DmaPtr);
Status |= XScuGic_Connect(GicPtr,
DMA_DONE_INTR_7,
(Xil_InterruptHandler)XDmaPs_DoneISR_7,
(void *)DmaPtr);

if (Status != XST_SUCCESS)
return XST_FAILURE;

/*
* Enable the interrupts for the device
*/
XScuGic_Enable(GicPtr, DMA_DONE_INTR_0);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_1);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_2);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_3);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_4);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_5);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_6);
XScuGic_Enable(GicPtr, DMA_DONE_INTR_7);
XScuGic_Enable(GicPtr, DMA_FAULT_INTR);

Xil_ExceptionEnable();
return XST_SUCCESS;
}

/*****************************************************************************/
/**
*
* DmaDoneHandler.
*
* @param Channel is the Channel number.
* @param DmaCmd is the Dma Command.
* @param CallbackRef is the callback reference data.
*
* @return None.
*
* @note None.
*
******************************************************************************/
void DmaDoneHandler(unsigned int Channel, XDmaPs_Cmd *DmaCmd, void *CallbackRef)
{

/* done handler */
volatile int *Checked = (volatile int *)CallbackRef;
int Index;
int Status = 1;
int *Src;
int *Dst;

Src = (int *)DmaCmd->BD.SrcAddr;
Dst = (int *)DmaCmd->BD.DstAddr;

/* DMA successful */
/* compare the src and dst buffer */
for (Index = 0; Index < DMA_LENGTH; Index++) {
if ((Src[Index] != Dst[Index]) ||
(Dst[Index] != DMA_LENGTH - Index)) {
Status = -XST_FAILURE;
}
}

Checked[Channel] = Status;
}

3.2 修改DMA PS
修改DMA配置，使其将PL中的数据传输到内存中。

修改DMA的源地址：
DmaCmd.BD.SrcAddr = (u32) RAM_R;

测量DMA传输16KB数据，时间约为180us，远远高于memcpy。

4. Linux DMA驱动
4.1 编程方法
配置DMA
void dma_init(u32 s, int size)
{
dma_cap_mask_t mask;
//alloc 512B src memory and dst memory
dma_src = s;
printk(KERN_INFO "dma_src = 0x%x\n",src);
//src = dma_alloc_coherent(NULL, MM_SIZE, &dma_src, GFP_KERNEL);
dst = dma_alloc_coherent(NULL, size, &dma_dst, GFP_KERNEL);
printk(KERN_INFO "dst = 0x%x, dma_dst = 0x%x\n",dst, dma_dst);

dma_cap_zero(mask);
dma_cap_set(DMA_MEMCPY, mask);//direction:memory to memory
chan = dma_request_channel(mask,NULL,NULL); //request a dma channel
printk(KERN_INFO "dma channel id = %d\n",chan->chan_id);

flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
dev = chan->device;
}

释放DMA
void dma_del(void)
{
//free memory and dma channel
dma_free_coherent(NULL, MM_SIZE, dst, &dma_dst);
dma_release_channel(chan);
}

向DMA引擎发起一个传输请求
void dma_read(u32 dst,u32 src,int size)
{
//alloc a desc,and set dst_addr,src_addr,data_size.
/*获取时间*/
do_gettimeofday(&tb);
tx = dev->device_prep_dma_memcpy(chan, dst, src, size, flags);
if (!tx){
printk(KERN_INFO "Failed to prepare DMA memcpy");
}

tx->callback = dma_callback_func;//set call back function
tx->callback_param = NULL;
cookie = tx->tx_submit(tx); //submit the desc
if (dma_submit_error(cookie)){
printk(KERN_INFO "Failed to do DMA tx_submit");
}
dma_async_issue_pending(chan);//begin dma transfer
}

4.2 实例代码
将Block RAM中的数据先使用ioremap映射的地址src，写入一些字符，然后使用DMA从Block RAM中传输16KB数据到分配的内存dst中。传输完成后调用dma_callback_func函数，在该函数中比较传输的数据和发送的数据是否相同，并测量DMA消耗的时间。
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include

#include
#include

#define DEVICE_NAME "dma_driver"
#define ImageReadAddress0 0x40000000
volatile unsigned int *CaptureReadAddr0;

struct timeval tb, te;

#define MM_SIZE (1440*10)

void dma_callback_func(void *dma_async_param);
void dma_read(u32 dma_dst,u32 dma_src,int size);
void dma_init(u32 s, int size);
void dma_del(void);

struct dma_chan *chan;
//bus address
dma_addr_t dma_src;
dma_addr_t dma_dst;
//virtual address
char *src = NULL;
char *dst = NULL ;
struct dma_device *dev;
struct dma_async_tx_descriptor *tx = NULL;
enum dma_ctrl_flags flags;
dma_cookie_t cookie;

//When dma transfer finished,this function will be called.
void dma_callback_func(void *dma_async_param)
{
int i=0;
do_gettimeofday(&te);
printk("DMA\n");
printk("T:%ld, %ld\n", tb.tv_sec, tb.tv_usec);
printk("T2:%ld, %ld\n", te.tv_sec, te.tv_usec);
printk(KERN_ALERT "time use:%ld, %ld\n",
(te.tv_sec-tb.tv_sec),
(te.tv_usec-tb.tv_usec));

printk("memcpy\n");
do_gettimeofday(&tb);
memcpy(dst ,src, MM_SIZE);
do_gettimeofday(&te);
printk("T:%ld, %ld\n", tb.tv_sec, tb.tv_usec);
printk("T2:%ld, %ld\n", te.tv_sec, te.tv_usec);
printk(KERN_ALERT "time use:%ld, %ld\n",
(te.tv_sec-tb.tv_sec),
(te.tv_usec-tb.tv_usec));

printk("DMA transfer finished!\n\r");
for(i=0; i {
if(*(dst + i) != (char)('a' + i%26))
{
printk("Failed\n");
return;
}
}
printk("PASS\n");
}

void dma_read(u32 dst,u32 src,int size)
{
//alloc a desc,and set dst_addr,src_addr,data_size.
/*获取时间*/
do_gettimeofday(&tb);
tx = dev->device_prep_dma_memcpy(chan, dst, src, size, flags);
if (!tx){
printk(KERN_INFO "Failed to prepare DMA memcpy");
}

tx->callback = dma_callback_func;//set call back function
tx->callback_param = NULL;
cookie = tx->tx_submit(tx); //submit the desc
if (dma_submit_error(cookie)){
printk(KERN_INFO "Failed to do DMA tx_submit");
}
dma_async_issue_pending(chan);//begin dma transfer
}

void dma_init(u32 s, int size)
{
dma_cap_mask_t mask;
//alloc 512B src memory and dst memory
dma_src = s;
printk(KERN_INFO "dma_src = 0x%x\n",src);
//src = dma_alloc_coherent(NULL, MM_SIZE, &dma_src, GFP_KERNEL);
dst = dma_alloc_coherent(NULL, size, &dma_dst, GFP_KERNEL);
printk(KERN_INFO "dst = 0x%x, dma_dst = 0x%x\n",dst, dma_dst);

dma_cap_zero(mask);
dma_cap_set(DMA_SLAVE, mask);//direction:memory to memory
chan = dma_request_channel(mask,NULL,NULL); //request a dma channel
printk(KERN_INFO "dma channel id = %d\n",chan->chan_id);

flags = DMA_CTRL_ACK | DMA_PREP_INTERRUPT;
dev = chan->device;
}

void dma_del(void)
{
//free memory and dma channel
dma_free_coherent(NULL, MM_SIZE, dst, &dma_dst);
dma_release_channel(chan);
}

static int device_open(struct inode *inode, struct file *file)
{
return 0;
}

static int device_close(struct inode *indoe, struct file *file)
{
printk("device close\n");
return 0;
}

static ssize_t device_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
{
int ret = 0;

dma_read(dma_dst, dma_src, MM_SIZE);

return ret;
}

static struct file_operations device_fops =
{
.owner = THIS_MODULE,
.open = device_open,
.release = device_close,
.read = device_read,
};

static struct miscdevice MMAP_misc =
{
.minor = MISC_DYNAMIC_MINOR,
.name = DEVICE_NAME,
.fops = &device_fops,
};

static int __init char_device_init( void )
{
int ret=0;
int i = 0;
printk("init module\n");
ret = misc_register(&MMAP_misc);
if(ret)
{
printk("Error:misc_register failed!\n");
return 0;
}

CaptureReadAddr0 = (volatile unsigned int*)ioremap(ImageReadAddress0, 1440*10);

printk("init module\n");
dma_init(ImageReadAddress0, MM_SIZE);

src = (char*)CaptureReadAddr0;
for (i = 0; i < MM_SIZE; i++){
*(src + i) = (char)('a' + i%26);
}

return 0;
}

static void __exit char_device_exit( void )
{
printk(KERN_ALERT"module exit\n");
misc_deregister(&MMAP_misc);

iounmap(CaptureReadAddr0);
dma_del();
}

MODULE_LICENSE("GPL");
MODULE_AUTHOR("DMA_test");

module_init(char_device_init);//模块加载
module_exit(char_device_exit);//模块退出

使用DMA搬运和memcpy搬运PL中的数据速度对比如下：
Z-turn# ./test
Test for dma
DMA
T:34, 358179
T2:34, 358290
time use:0, 111
memcpy
T:34, 364372
T2:34, 364796
time use:0, 424
DMA transfer finished!
PASS

DMA搬运消耗了111us，而memcpy需要使用424us，可见DMA速度远高于CPU对数据的搬运。