EVM 6472 large amount of data transfer

Mary,

One option would be to use MessageQ to pass a pointer to the data you want to process between cores.  The data would have to be in shared memory like DDR2.  In order to add the pointer to the MessageQ message you need to define your own message structure.  Following is a quick example of how to do this.

Define new message structure:

typedef struct TstMsg
{
  MessageQ_MsgHeader header;      /* 32 bytes */
  UInt32 *dataPtr;
} TstMsg;

Allocate, populate, and send message to another core

/* send core */
  {
    TstMsg *sendMsg;

    sendMsg = (TstMsg *)MessageQ_alloc(heapId, sizeof(TstMsg));

    sendMsg->dataPtr = &array[0];

    MessageQ_put(remoteCoreId, (MessageQ_Msg)sendMsg);
  }

  /* receive core */
  {
    TstMsg *rxMsg;
    UInt32 *rxData;

    MessageQ_get(coreRcvMsgQ, (MessageQ_Msg*)&rxMsg, MessageQ_FOREVER);

    rxData = rxMsg->dataPtr;
   
    MessageQ_free((MessageQ_Msg) rxMsg);

    /* process data using rxData pointer */
  }

This example is extremely simple so you'll need to handle things like cache writeback (send core) and invalidate (receive core) to make sure your data is coherent across cores.  You'll also need to add some sort of arbitration to make sure the sender and receiver aren't modifying/viewing the data at the same time.

Hope this helps.

Justin

Thanks,Justin. I just want to know how to transfer my  own message.  This example tells me .  But when the message is very large , it will waste much time for putting and getting.  So I define array as a global variable and placed at DDRII. Core0 and core1 can deal with array directly instead of tranfering .

Moreover , can you tell me whether  array running in core0  and array running  in core1 has the same value?  In other words, if I define a global value named array which is placed in DDRII.   Does the value of array change simultaneously  in core1 when core0 changes the value of array? From the watch window of CCS , I watched array in core0 and array in core1 has the same address.

Thanks for  your suggestion,Todd. I'll take it.

Justin,

Thanks!  Your explanation is so detailed that I know how to transfer my own message and how to keep the cache cohenrence with Cache_inv and Cache_wb(suggested by Todd and Karl). I have verified your method by modifying the programm message_single.c (<ipc_install>/ipc_1_21_xx_xx /packages/ti/sdo/ipc/examples/multicore/c6472).

Justin,

In the example meesge_single.c , core1 must wait and get a message until core 0 put a message . Core 0  can't  put another message until array is free.To improve the efficiency of core0 and core1, I want to use three buffers array, array1 and array2 to store datas.  Core0 store data in array, array1 and array2 by  turns with link. When core1 process data in array, core0 can store  data in array0 and array1 by turns.  When core1 process data in array1, core0 can store data in array and array0. In short,I want to process data in array, array1,array2 by ping-pong operation.  Can you give me some suggestion?

Mary,

Your idea is good and will allow you to get better throughput.  I advise to use Todd's suggestion of messages as tokens.  Whoever has the token can manipulate the data.  You'd have three token messages, one for each ping-pong buffer.  Whichever core has the ping-pong buffer token message can read/write the data.

Justin