DSP C6678 EDMA 多核并行处理耗时问题

Chi Wang1

各位专家，大家好！

关于6678中EDMA模块8核并行传输。我采用了如下配置：核0~1使用EDMA CC0，核2~5使用EDMA CC1，核6~7使用EDMA CC2。核0使用Que0；核1使用Que1；核2使用Que0；核3使用Que1；核4使用Que2；核5使用Que3；核6使用Que0；核7使用Que1。

每个核使用与核号相同的通道号，完成标志TCC，PaRAM。每个通道采用shadow_region方式，以核号的形式进行区域划分。

现在的情况是8个核并行传输，都可以完成EDMA传输。但是达不到并行模式，耗时没有得到优化。根据6678 datasheet。EDMA模块共有10个TC模块，所有理论上可以达到8个核并行传输8个通道的数据。但是单核传输大块数据，与8核均分分块传输数据，耗时相差不大（传输64M的数据时，单核耗时为200ms；8核耗时为169ms）。没有达到并行传输的目的。虽然数据传输完成并且没有错误。

请问要达到8核完全并行运行EDMA，使8核耗时达到单核的1/8时，需要更改设置哪些参数？非常感谢！请各位专家指点。希望能够得到各位专家的宝贵建议。

附上主要函数：

#include <ti/csl/csl_chip.h>
#include <ti/csl/csl_edma3.h>
#include <ti/csl/csl_edma3Aux.h>
#include <ti/csl/csl_cacheAux.h>

CSL_Edma3ChannelObj chObj;
CSL_Edma3Obj edmaObj;
CSL_Status status;
CSL_Edma3Context context;
extern unsigned int coreNum;

/***************************************************************
* void DMA_Init_region(int ChannelNum)
* use region shadow
**************************************************************/
void DMA_Init_region()
{
int instNum;
if(coreNum < 2)
{
instNum = 0;
}
else if(coreNum < 6)
{
instNum = 1;
}
else
{
instNum = 2;
}
CSL_Edma3ChannelAttr chAttr;
CSL_Edma3Handle hModule;
CSL_Edma3ChannelHandle hChannel;
CSL_Edma3CmdDrae regionAccess;
CSL_Edma3CmdIntr regionIntr;

/* Module initialization */
CSL_edma3Init(&context);

/* Module level open */
hModule = CSL_edma3Open(&edmaObj,instNum, NULL, &status);

regionAccess.region = coreNum;
regionAccess.drae = 0x1<<coreNum;
regionAccess.draeh = 0x0;
CSL_edma3HwControl(hModule,CSL_EDMA3_CMD_DMAREGION_ENABLE, &regionAccess);

regionIntr.region = coreNum;
regionIntr.intr = 0x1<<coreNum;
regionIntr.intrh = 0x0;
CSL_edma3HwControl(hModule,CSL_EDMA3_CMD_INTR_ENABLE, &regionIntr);

CSL_edma3MapEventQueueToTC(hModule, 0, 0);
if(coreNum < 2)
{
CSL_edma3MapEventQueueToTC(hModule, coreNum, coreNum);
}
else if(coreNum < 6)
{
CSL_edma3MapEventQueueToTC(hModule, coreNum-2, coreNum-2);
}
else
{
CSL_edma3MapEventQueueToTC(hModule, coreNum-6, coreNum-6);
}

chAttr.regionNum = coreNum; //coreNum; //CSL_EDMA3_REGION_GLOBAL CSL_EDMA3_REGION_0
chAttr.chaNum = coreNum;
hChannel = CSL_edma3ChannelOpen(&chObj, instNum, &chAttr, &status);
/* Map the DMA Channel to the appropriate PARAM Block. We start with PING
* which is located at PARAM Block cNum. */
CSL_edma3HwChannelSetupParam(hChannel, coreNum); // ParamNum
/* Enable channel */
CSL_edma3HwChannelControl(hChannel,CSL_EDMA3_CMD_CHANNEL_ENABLE, NULL);

}

/***************************************************************
* void DMA_transport_ab_region(int ChannelNum,Uint32 srcBuff , Uint32 dstBuff, int acnt, int bcnt, int srcBidx, int dstBidx)
* use region shadow
**************************************************************/
void DMA_transport_ab_region(Uint32 srcBuff , Uint32 dstBuff, int acnt, int bcnt, int srcBidx, int dstBidx)
{
int instNum;
if(coreNum < 2)
{
instNum = 0;
}
else if(coreNum < 6)
{
instNum = 1;
}
else
{
instNum = 2;
}
CSL_Edma3ChannelHandle hChannel;
CSL_Edma3ParamSetup myParamSetup;
CSL_Edma3ParamHandle hParamPing;
CSL_Edma3ChannelAttr chAttr;

chAttr.regionNum = coreNum; //coreNum; //CSL_EDMA3_REGION_GLOBAL CSL_EDMA3_REGION_0
chAttr.chaNum = coreNum;
hChannel = CSL_edma3ChannelOpen(&chObj, instNum, &chAttr, &status);

hParamPing = CSL_edma3GetParamHandle(hChannel, coreNum, &status); // panum

myParamSetup.option = CSL_EDMA3_OPT_MAKE(CSL_EDMA3_ITCCH_DIS, \
CSL_EDMA3_TCCH_DIS, \
CSL_EDMA3_ITCINT_DIS, \
CSL_EDMA3_TCINT_EN, \
coreNum, CSL_EDMA3_TCC_NORMAL,\
CSL_EDMA3_FIFOWIDTH_NONE, \
CSL_EDMA3_STATIC_DIS, \
CSL_EDMA3_SYNC_AB, \
CSL_EDMA3_ADDRMODE_INCR, \
CSL_EDMA3_ADDRMODE_INCR );

myParamSetup.srcAddr = srcBuff;
myParamSetup.aCntbCnt = CSL_EDMA3_CNT_MAKE(acnt,bcnt);
myParamSetup.dstAddr = dstBuff;
myParamSetup.srcDstBidx = CSL_EDMA3_BIDX_MAKE(srcBidx, dstBidx);
myParamSetup.srcDstCidx = CSL_EDMA3_CIDX_MAKE(0,0);
myParamSetup.cCnt = 1;
myParamSetup.linkBcntrld= CSL_EDMA3_LINKBCNTRLD_MAKE(CSL_EDMA3_LINK_NULL,0);
CSL_edma3ParamSetup(hParamPing,&myParamSetup);

switch(coreNum)
{
case 0:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_0);
break;
case 1:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_1);
break;
case 2:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_0);
break;
case 3:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_1);
break;
case 4:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_2);
break;
case 5:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_3);
break;
case 6:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_0);
break;
case 7:
CSL_edma3HwChannelSetupQue(hChannel,CSL_EDMA3_QUE_1);
break;
}

/* Manually trigger the channel 设置同步事件*/
CSL_edma3HwChannelControl(hChannel,CSL_EDMA3_CMD_CHANNEL_SET,NULL);
}

/***************************************************************
* void waitDMAover_region(int ChannelNum)
* use region shadow
**************************************************************/
void waitDMAover_region()
{
CSL_Edma3Handle hModule;
CSL_Edma3CmdIntr regionIntr;
int instNum;
if(coreNum < 2)
{
instNum = 0;
}
else if(coreNum < 6)
{
instNum = 1;
}
else
{
instNum = 2;
}
int query=0x1<<coreNum;
hModule = CSL_edma3Open(&edmaObj,instNum,NULL,&status);
regionIntr.region = coreNum;//coreNum
regionIntr.intr = 0;
regionIntr.intrh = 0;
/* Poll on IPR bit 0 */
do {
CSL_edma3GetHwStatus(hModule,CSL_EDMA3_QUERY_INTRPEND,&regionIntr);
} while (!(regionIntr.intr & query));

/* Clear the pending bit */
CSL_edma3HwControl(hModule,CSL_EDMA3_CMD_INTRPEND_CLEAR, &regionIntr);
}

/***************************************************************
* void DMA_Close_region(int ChannelNum)
* use region shadow
**************************************************************/
void DMA_Close_region()
{
int instNum;
if(coreNum < 2)
{
instNum = 0;
}
else if(coreNum < 6)
{
instNum = 1;
}
else
{
instNum = 2;
}
CSL_Edma3Handle hModule;
CSL_Edma3ChannelHandle hChannel;
CSL_Edma3ChannelAttr chAttr;

chAttr.regionNum = coreNum;
chAttr.chaNum = coreNum;
hChannel = CSL_edma3ChannelOpen(&chObj, instNum, &chAttr, &status);
hModule = CSL_edma3Open(&edmaObj,instNum,NULL,&status);

/* Close channel */
CSL_edma3ChannelClose(hChannel);
/* Close EDMA module */
CSL_edma3Close(hModule);
}

unsigned int convert_coreLocalToGlobalAddr( unsigned int addr)
{
unsigned int coreNum0;
coreNum0 = CSL_chipReadReg(CSL_CHIP_DNUM);
return ((1 << 28) | (coreNum0 << 24) | (addr & 0x00ffffff));
}

* main.c

#include <stdint.h>

#include <math.h>

#include <stdio.h>

#include <string.h>

#include <stdlib.h>

#include <ti/csl/csl_ChipAux.h>

#include <ti/csl/csl_semAux.h>

#include "RSP_EDMA.h"

#pragma DATA_ALIGN(x_sp, 8);

float x_sp[2*NRN];

unsigned int coreNum; //当前程序运行的核号

void main ()

{

coreNum=CSL_chipReadReg(CSL_CHIP_DNUM);//获得各核编号

coreNum = coreNum % 8;

Uint32 srcAddr = 0x84000000 + coreNum*NAN*2*size_float/8;

DMA_Init_region();

int i = 0;

int loopNum = NAN/8;

for(i=0; i<loopNum; i++)

{

//传入

DMA_transport_ab_region(srcAddr + 2*size_float*i, convert_coreLocalToGlobalAddr((Uint32)&x_sp),

size_float*2, NRN, 2*size_float*NAN, 2*size_float);

waitDMAover_region();

}

CSL_semAcquireDirect(coreNum+1); //每个核获得对应核号加1编号的信号量

DMA_Close_region();

}

7 年多前

0 Ryan BL 7 年多前

Genius 3929 points

流量优化是个细致核麻烦以及有技术含量的活儿;没仔细看你的代码;
建议你搜搜TI专家们发布过的memory performance文档，上面有EDMA使用的技巧，以进行针对性的优化;

0 Ryan BL 7 年多前

Genius 3929 points

e2echina.ti.com/.../download
帮你找到了,或许能解决你的问题.

0 Chi Wang1 7 年多前回复 Ryan BL

Prodigy 50 points

非常感谢。

0 Ryan BL 7 年多前回复 Chi Wang1

Genius 3929 points

不客气.

处理器

处理器论坛

DSP C6678 EDMA 多核并行处理耗时问题