combining OpenMP and pre-optimized multicore codecs / algorithms (such as H.264)

Raja-

> We don’t support OpenMP Runtime 1.0 and expect you to follow the
> steps outlined in the wiki to build your own OpenMP Runtime 2.0.

Thanks for your reply to me (and Paula).   So let me confirm what we should be doing:

1) Build OpenMP runtime 2.0 library.

2) Make source code changes to reserve cores 0 and 1 (force starting core to be 2).

3) Test, combined with TI H.264 codec.

Does this sound correct ?

-Jeff

Raja-

I ported OpenMP 2.x for KeyStone 1 following instructions in the link  http://processors.wiki.ti.com/index.php/Porting_OpenMP_2.x_to_KeyStone_1, which involved rebuilding the C6678 PDK and openmp_dsp_2_01_16_02. 

Do I have to make any more changes other than the below cfg change to guarantee OpenMP uses only cores from 2 - 8?

var OpenMP = xdc.useModule('ti.runtime.ompbios.OpenMP');
OpenMP.masterCoreIdx = 2;
OpenMP.numCores = 6;

Thanks
Anish

Raja-

Can you let me and Anish know on this ?  Thanks.

-Jeff

Jeff,

The current release allows you to set OpenMP.masterCoreIdx to be zero without any issue, and the index error occurs  when OpenMP.masterCoreIdx is **non** zero (from Anish's post). The index error has been fixed and we will push to git.ti.com/.../ti-openmp-dsp-runtime shortly (in a day or so). If you need immediate access, here is the changes:

In OpenMP.xs

--- a/src/ti/runtime/ompbios/OpenMP.xs
+++ b/src/ti/runtime/ompbios/OpenMP.xs
@@ -133,17 +133,20 @@ function configureIpc(masterCoreIdx, numCores)
     // Configure MultiProc with the base index and number of cores
     var MultiProc   = xdc.useModule('ti.sdo.utils.MultiProc');
     MultiProc.baseIdOfCluster = masterCoreIdx;
-    // Total number of processors
-    MultiProc.numProcessors = numCores;
+
+    // Total number of processors available in the system
+    MultiProc.numProcessors = dspNames.length;
     // Number of processors in cluster set by setConfig
+    var endIndex = MultiProc.baseIdOfCluster + numCores;
     MultiProc.setConfig(null,
-                 dspNames.slice(MultiProc.baseIdOfCluster, numCores));
+                 dspNames.slice(MultiProc.baseIdOfCluster, endIndex));
+

     // Avoid wasting shared memory for MessageQ transports or notify.
     // Ipc is used exclusively to set up a HeapMemMP in the specified
     // Shared Region to support a shared heap for malloc's
     var Ipc         = xdc.useModule('ti.sdo.ipc.Ipc');
-    for (var i = 0; i < MultiProc.numProcessors; i++) {
+    for (var i = 0; i < MultiProc.numProcsInCluster; i++) {
         Ipc.setEntryMeta({
                 remoteProcId: i,
                 setupMessageQ: false,


In the config file omp_config.cfg:
     SharedRegion.setEntryMeta( sharedRegionId,
                                {   base: ddr3.base,
                                    len:  sharedHeapSize,
-                                   ownerProcId: 0,
+                                   ownerProcId: OpenMP.masterCoreIdx,
                                    cacheEnable: true,
                                    createHeap: true,
                                    
Regards,
Garrett

Thanks Garrett. I will try OpenMP with the new code changes.

-Anish

Garrett,

We got OpenMP hello and basic convolution programs running with below settings

OpenMP.masterCoreIdx = 0;
OpenMP.numCores = 8;

Next we will try with 

OpenMP.masterCoreIdx = 2;
OpenMP.numCores = 6;

-Anish

Hello,

When I tried to use OpenMP with master core = 2 and OpenMP.numCores = 4, core 0 is getting stuck at module initialization step in SYS/BIOS startup seqeunce. I have tried adding Startup firstFxns and core 0 gets to that function with no problem.I have already made OpenMP.xs, omp_config.cfg changes suggested by Garrett and rebuilt the library.

I am using OpenMP in RTSC mode and loading the same code to cores 0, 1, 2, 3, 4 and 5 using TI dsp_utils program.  __StartCores function in omplib.c is used to kickstart other cores by core 0. I am able to run OpenMP without any issues with  master core = 0.

I was reading about module initialization at http://rtsc.eclipse.org/docs-tip/Using_xdc.runtime_Startup and my program is also not reaching main.

I am loading same .out file to all cores. Do I need a different .out file just for core 0 which do not have OpenMP code? 

Do OpenMP module has initialization functions? What else can be done to debug this problem?

Thanks
Anish

1184.omp_config.cfg

#include <stdio.h>
#include <string.h>
#include "/opt/ti/ti-cgt-c6000_8.0.1/include/c6x.h" 
#include <ti/sysbios/family/c66/Cache.h>
#include <ti/csl/csl_bootcfgAux.h>
#include <ti/csl/csl_ipcAux.h>

#pragma DATA_SECTION (mainProg,"L2SRAM");
volatile int mainProg = 0;
#pragma DATA_SECTION (ompProg,"L2SRAM");
volatile int ompProg = 0;
#define LOG_LEN 0x100000
#pragma DATA_SECTION(log_buffer, "DDR3")
#pragma DATA_ALIGN(log_buffer,8)
volatile unsigned char log_buffer[8][LOG_LEN];
volatile unsigned int log_idx = 0;
va_list va;

void core_print(uint16_t loglevel, char *fmt, ...)
{
   char outputString[1024];
   va_start(va, fmt);
   vsnprintf(outputString, 1024, fmt, va);

   if (log_idx < LOG_LEN)  {
      memcpy((void *)&log_buffer[DNUM][log_idx], outputString, strlen(outputString));
      log_idx += strlen(outputString);
    }
}

void __StartCores(void){
    
   int core = 0;
   volatile unsigned int numCores = 0;
   //volatile int nCoreList = 0x1f;
   volatile int nCoreList = 0x03f;
   unsigned int tmp_coreList;
    
   mainProg |= 0x1;   
   tmp_coreList = nCoreList;
   mainProg |= 0x2;
   
   TSCL = 0; /* start TSC register counting */
   
/* calculate number of cores */

	do {
      numCores++;
      tmp_coreList >>= 1;
    } while (tmp_coreList > 0);
   
   if (DNUM == 0) {  /* core zero only */
      CSL_BootCfgUnlockKicker();
      tmp_coreList = nCoreList;
      do  {
         if (tmp_coreList & 1) {
            *(unsigned int *)(0x0C3FFF00 + core*4) = 0;
            if (core != 0) CSL_IPC_genGEMInterrupt( core, 0 );
            while(TSCL < core * 1000000);
         }
         core++;
         tmp_coreList >>= 1;
        } while(tmp_coreList > 0);
    }

   mainProg |= 0x4;

   while(TSCL < 10000000);

   mainProg |= 0x8;

/* The core has reached __StartCores */

   *(unsigned int *)(0x0C3FFF00 + DNUM*4) = 1;

   mainProg |= 0x10;
  
/* All cores wait here to sync up */

   while(1) {
        
      int i = 0;
      core = 0;
      tmp_coreList = nCoreList;
      mainProg |= 0x20;
      Cache_wbInvAll();
      mainProg |= 0x40;

      do {
         if (*(unsigned int *)(0x0C3FFF00 + core*4) == 1) i++;
            core++;
            tmp_coreList >>= 1;
       } while (tmp_coreList > 0);

       if (i == numCores) break;
    }
	
    mainProg |= 0x20;       
}

Anish,

Which version of the OpenMP runtime are you using? v2.01.17.02 fixes a defect related to using a non zero master core index in RTSC mode.

Also, if you have any cores that are not configured to be part of the OpenMP runtime, you'll have to use your own binaries for those cores.

Ajay

Ajay,

The version I am running is openmp_dsp_2_01_16_03 which got installed when I ran mcsdk-hpc_03_00_01_12_setuplinux.bin(MCSDK HPC Package for Linux)

I am not able to find where to download openmp_dsp_2_01_17_02 from. Can you please provide the download url for openmp_dsp_2_01_17_02?

Thanks

Ajay-

We understand the need for separate .out (binaries) and that's our long-term objective:  one .out with multimedia codecs and another with our c66x OpenCV port, accelerated with TI OpenMP

However, for the immediate test purpose, isn't there a way to use the same openmp-enabled .out for all cores?  For example, if master core index = 2, then cores 0 and 1 should skip any OpenMP related initialization and not hang during SYSBIOS module init.  Can you guys add that ?

Thanks.

-Jeff
Signalogic

Jeff,

The intent is that any core not part of the OpenMP runtime will have it's own custom binary. 

As a quick test, you can override the default definition of __TI_omp_initialize_rtsc_mode provided by the OpenMP runtime and supply your own version. In this version, you can disable OpenMP initialization for cores not participating in the OpenMP runtime.

Ajay

Ajay,

As far as I can tell, OpenMP runtime version 2.01.17.02 is not publicly available. I am not able to find that version in TI website.
Where can I download OpenMP runtime version 2.01.17.02 from?


Thanks
Anish