some cleanup,

added example_fast_pushpop
back out cache op better in burst send
burst send now sends pkts not requiring any netcp offload

diff --git a/ti/runtime/netapi/test/net_test.c b/ti/runtime/netapi/test/net_test.c
index 3e31adcb7ede9cdaea66a4352562ab086204133e..c01a8174ea20ef56ffe687d4f8cdaa521f97d3e3 100644 (file)
--- a/ti/runtime/netapi/test/net_test.c
+++ b/ti/runtime/netapi/test/net_test.c
@@ -83,7 +83,7 @@ void dump_header(unsigned long *p, int n, int a, int r)
 
 
 //************for multi pkt burst  xfer test in loopback mode
-#define TX_BURST 800 
+#define TX_BURST 700 
 int pktloopback=TUNE_NETAPI_NWAL_ENABLE_PASS_LOOPBACK;
 nwalTxPSCmdInfo_t   flowPSCmdInfo;
 
@@ -138,7 +138,9 @@ static uint8_t testPkt[] = {
 /* The pseudo header checksum of the packet except for the 16 bit length */
 #define TEST_PKT_PSEUDO_HDR_CHKSUM_SANS_LEN  0x0FFC
 
-void example_fast_poll( PKTIO_HANDLE_T * p_pktio);
+void example_fast_poll( PKTIO_HANDLE_T * p_pktio, int max_pkts);
+void example_fast_pushpop(Pktlib_HeapHandle p_heap, int ntrials);
+
 
 #if 1
 //#include "arpa/inet.h"
@@ -1310,6 +1312,9 @@ unsigned int nwal_sum_vv6=0;
 
 unsigned int nwal_sum_flow_vv1=0;
 unsigned int nwal_sum_flow_vv2=0;
+unsigned int cache_op_b1;
+unsigned int cache_op_b2;
+int n_c_ops;
 static int first =0;
 Cppi_HostDesc*      pPktDesc;
 
@@ -1346,6 +1351,8 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
 
    /* set up meta data */
     meta.sa_handle=nwal_HANDLE_INVALID;
+//#define BENCH_UDP_SEND
+#ifdef BEND_UDP_SEND
     meta_tx.txFlag1 = (NWAL_TX_FLAG1_DO_IPV4_CHKSUM | NWAL_TX_FLAG1_DO_UDP_CHKSUM| NWAL_TX_FLAG1_META_DATA_VALID);
     meta_tx.startOffset = 0;
     //GONE in V2 meta_tx.pktLen = len;
@@ -1359,7 +1366,11 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
     pIpHdr = pData + meta_tx.ipOffBytes;
     meta_tx.pseudoHdrChecksum =
         test_utilGetIpv4PsudoChkSum(pIpHdr,(TEST_PAYLOAD_LEN+TEST_PKT_UDP_HDR_LEN));
-
+#else
+    meta_tx.txFlag1 = NWAL_TX_FLAG1_META_DATA_VALID;
+    meta_tx.startOffset = 0;
+    meta_tx.ploadLen = TEST_PAYLOAD_LEN;
+#endif
    /* post it to netcp tx channel*/
    meta.u.tx_meta=&meta_tx;
 #ifdef DEBUG_DESC
@@ -1382,13 +1393,21 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
        nwal_sum_flow_vv2 += (nwal_flow_vv2-nwal_flow_vv1); 
    }
        
+   cache_op_b1= Osal_cache_op_measure(&n_c_ops);
    vv2= netapi_timing_stop();
+#ifdef BEND_UDP_SEND
    nwal_mCmdSetL4CkSumPort(  tip,
                              &flowPSCmdInfo,
                              TEST_PKT_UDP_OFFSET_BYTES,
                              (TEST_PKT_UDP_HDR_LEN + TEST_PAYLOAD_LEN),
                              meta_tx.pseudoHdrChecksum,
                              meta_tx.enetPort);
+#else
+   nwal_mCmdSetPort  (tip,
+                      &flowPSCmdInfo,  //could be NULL
+                      0);  //port 0 -> switch decides
+
+#endif
 
    pPktDesc = Pktlib_getDescFromPacket(tip);
    /* Send the packet out to transmit Q*/
@@ -1396,7 +1415,7 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
                         pPktDesc, 
                         NWAL_DESC_SIZE);
    vv3= netapi_timing_stop();
-   
+   cache_op_b2= Osal_cache_op_measure(&n_c_ops);
 
    sum_vv1 += (vv2-vv1);
    if(!house_pkts_gened)
@@ -1405,7 +1424,7 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
        sum_vv1 = sum_vv1 - nwal_sum_flow_vv2;
    }
 
-   sum_vv3 += (vv3-vv2);
+   sum_vv3 += (vv3-vv2)-(cache_op_b2-cache_op_b1); //sub out cache op cost
 
    // printf("pktio send. full=%d metadata=%d pktio_send=%d\n", vv3-vv1,  vv2-vv1,  vv3-vv2);
    
@@ -1415,18 +1434,18 @@ if ((house_pkts_gened>0) && (! (house_pkts_gened%1000)) )
  }
 {
    unsigned int ccycles;
-   int n_c_ops;
    ccycles =Osal_cache_op_measure(&n_c_ops);
    if (sum_vv1) 
    {
-      printf("BURST NWAL Fast send %d pkts.  metadata=%d Cmd Label Creation Cost=%d  nwal Fast Send Cost= %d n_c_ops=%d cache_op_time=%d (pp-> %d)\n", 
+      printf("BURST NWAL Fast send %d pkts.  metadata=%d Cmd Label Creation Cost=%d  nwal Fast Send Cost (less cacheop)= %d n_c_ops=%d cache_op_time=%d (pp-> %d)\n", 
               stats.itx, sum_vv1/stats.itx,  nwal_sum_flow_vv2, sum_vv3/stats.itx, 
               n_c_ops, ccycles, n_c_ops? (ccycles/(n_c_ops/2)) : 0);
-
+#if 0
       printf("NWAL Profile Cycles: Prof1= %d,Prof2=%d,Prof3=%d,Prof4=%d,Prof5=%d ,Prof6=%d \n",
               nwal_sum_vv1/stats.itx,nwal_sum_vv2/stats.itx,nwal_sum_vv3/stats.itx,
               nwal_sum_vv4/stats.itx,nwal_sum_vv5/stats.itx,nwal_sum_vv6/stats.itx);
-    
+  
+#endif  
 
       if(stats.itx2)
       {
@@ -1541,6 +1560,7 @@ printf("net_test> %d bytes left in our CMA area\n", netapi_getBufMemRemainder())
 our_sched =netapi_schedOpen(netapi_handle,&our_sched_cfg, &err);
 if (!our_sched) {printf("sched create failed\n"); exit(1);}
 
+#if 0
 /********************************************
 * Basic pkt loopback test
 *********************************************/
@@ -1551,6 +1571,7 @@ our_pktio_bench(1000);
 our_pktio_bench(1000);
 our_pktio_bench(1000);
 /*********************************************/
+#endif
 
 
 
@@ -1753,12 +1774,13 @@ if (err!=0) {printf("stats req failed\n");}
 /*********************************************/
 
 
+#if 0
 /********************************************
 * Basic pkt loopback test
 *********************************************/
 printf("...runnining pure push/pop benchmark\n");
 our_pktio_bench(100);
-
+#endif
 
 /**************unused stuff******************/
 /* create TRIE */
@@ -1791,7 +1813,8 @@ our_stats_cb(netapi_handle, NULL);
 
 #define DO_FAST_POLL
 #ifdef DO_FAST_POLL
-example_fast_poll(netcp_rx_chan);
+example_fast_pushpop(OurHeap, 500);
+example_fast_poll(netcp_rx_chan,100000);
 #endif
 
 /*************************************************
@@ -1854,13 +1877,73 @@ static int l3_off[M], l4_off[M], L3_chk_ok[M], L4_chk_ok[M], len[M] ;
 static nwal_AppId appid[M];
 static unsigned char * buf[M];
 #define N2POLL 8 //batch size
-void example_fast_poll( PKTIO_HANDLE_T * p_pktio)
+#define NTOPOP 190
+Ti_Pkt * pHd[NTOPOP];
+void example_fast_pushpop(Pktlib_HeapHandle h , int n_trials)
+{
+int i,j;
+unsigned long t1;
+unsigned long t2;
+unsigned long sumf=0;
+unsigned long sumt=0;
+unsigned long sump=0;
+#define NTOPOP 190
+Ti_Pkt * tempVA;
+Qmss_QueueHnd freeQ=Pktlib_getInternalHeapQueue(h);
+
+//n_trials of NTOPOP actions
+for(i=0;i<n_trials;i++)
+{
+    //raw pop
+    t1= netapi_timing_start();
+    pHd[0] = (Ti_Pkt *)QMSS_DESC_PTR(PKTIO_QMSS_QUEUE_POP_RAW (freeQ));
+    for(j=1;j<NTOPOP;j++)
+    {
+       tempVA  =  _Osal_qmssPhyToVirt(pHd[j-1]);
+       netapi_pld(tempVA);
+        //__builtin_prefetch(tempVA);
+       pHd[j] = (Ti_Pkt *)QMSS_DESC_PTR(PKTIO_QMSS_QUEUE_POP_RAW (freeQ));
+       sumf+= (int) (   ((Cppi_HostDesc *) tempVA)->origBuffPtr -
+                  ((Cppi_HostDesc *) tempVA)->buffPtr);
+    }
+    tempVA  =  _Osal_qmssPhyToVirt(pHd[j-1]);
+    sumf+= (int) (   ((Cppi_HostDesc *) tempVA)->origBuffPtr -
+                  ((Cppi_HostDesc *) tempVA)->buffPtr);
+    t2=netapi_timing_stop();
+    sumt+= (t2-t1);
+
+    //invalidate/wb 
+    for(j=0;j<NTOPOP;j++)
+    {
+        tempVA  =  _Osal_qmssPhyToVirt(pHd[j]);
+        netapi_utilCacheWbInv(tempVA,128);
+    }
+
+    //raw push 
+    t1=netapi_timing_start();
+    for(j=0;j<NTOPOP;j++)
+    {
+        PKTIO_QMSS_QUEUE_PUSH_DESC_SIZE_RAW (freeQ,
+                                             (void *) pHd[j],
+                                             128);
+    }
+    t2=netapi_timing_stop();
+    sump += (t2-t1);
+}
+printf("\nfast poppush: np=%d  pop=%d  push=%d sumf=%d\n", (n_trials*NTOPOP),
+        sumt/(n_trials*NTOPOP),  sump/(n_trials*NTOPOP), sumf);
+sleep(1);
+return;
+}
+
+
+
+void example_fast_poll( PKTIO_HANDLE_T * p_pktio, int max_pkts)
 {
 int j=0;
 int jj=0;
 int i,k,l=0,ltot=0;
 int n= N2POLL; //max # of pkts to poll
-Ti_Pkt * pHd[N2POLL];
 Ti_Pkt * tempVA;
 pasahoLongInfo_t* pinfo;
 unsigned long t1;
@@ -1878,7 +1961,7 @@ int max_batch=0;
 Qmss_QueueHnd rxQ= PKTIO_GET_DEFAULT_NETCP_Q(p_pktio);
 Qmss_QueueHnd freeQ;
 //loop forever
-for(;;)
+for(;max_pkts>0;)
 {
     t1= netapi_timing_stop();
     pHd[0] = (Ti_Pkt *)QMSS_DESC_PTR(PKTIO_QMSS_QUEUE_POP_RAW (rxQ));
@@ -1975,6 +2058,8 @@ for(;;)
     }
     sumt += (t2-t1);
     sump +=(pHd[i-1]? i: (i-1));
+    max_pkts -= (pHd[i-1]? i: (i-1));
+
     //printf("end cleanup %d %d %d\n",sumt,sumf,sump );
     if (sump > 10000) {
        printf("pkts rx %d batches=%d  appid=%x l3_off=%d l4_off=%d len=%d buf=0x%x rxcycle= %d pkts/batchx1000=%d maxbatch=%d cycles per rawpush = %d\n",