source: branches/reconfiguration/platforms/tsar_generic_iob/top.cpp @ 879

///////////////////////////////////////////////////////////////////////////////
// File: top.cpp  (for tsar_generic_iob platform)
// Author: Alain Greiner
// Copyright: UPMC/LIP6
// Date : august 2013
// This program is released under the GNU public license
//
// Modified by: Cesar Fuguet
///////////////////////////////////////////////////////////////////////////////
// This file define a generic TSAR architecture with an IO network emulating
// an external bus (i.e. Hypertransport) to access 7 external peripherals:
//
// - FBUF : Frame Buffer
// - MTTY : multi TTY (one channel)
// - MNIC : Network controller (up to 2 channels)
// - CDMA : Chained Buffer DMA controller (up to 4 channels)
// - BDEV : Dlock Device controler (one channel)
// - IOPI : HWI to SWI translator.
// - SIMH : Simulation Helper
//
// The internal physical address space is 40 bits, and the cluster index
// is defined by the 8 MSB bits, using a fixed format: X is encoded on 4 bits,
// Y is encodes on 4 bits, whatever the actual mesh size.
// => at most 16 * 16 clusters. Each cluster contains up to 4 processors.
//
// It contains 3 networks:
//
// 1) the "INT" network supports Read/Write transactions
//    between processors and L2 caches or peripherals.
//    (VCI ADDDRESS = 40 bits / VCI DATA width = 32 bits)
//    It supports also coherence transactions between L1 & L2 caches.
// 3) the "RAM" network emulates the 3D network between L2 caches
//    and L3 caches, and is implemented as a 2D mesh between the L2 caches,
//    the two IO bridges and the physical RAMs disributed in all clusters.
//    (VCI ADDRESS = 40 bits / VCI DATA = 64 bits)
// 4) the IOX network connects the two IO bridge components to the
//    7 external peripheral controllers.
//    (VCI ADDDRESS = 40 bits / VCI DATA width = 64 bits)
//
// The external peripherals HWI IRQs are translated to WTI IRQs by the
// external IOPIC component, that must be configured by the OS to route
// these WTI IRQS to one or several internal XICU components.
// - IOPIC HWI[1:0]     connected to IRQ_NIC_RX[1:0]
// - IOPIC HWI[3:2]     connected to IRQ_NIC_TX[1:0]
// - IOPIC HWI[7:4]     connected to IRQ_CMA_TX[3:0]]
// - IOPIC HWI[8]       connected to IRQ_BDEV
// - IOPIC HWI[31:16]   connected to IRQ_TTY_RX[15:0]
//
// Besides the external peripherals, each cluster contains one XICU component,
// and one multi channels DMA component.
// The XICU component is mainly used to handle WTI IRQs, as only 5 HWI IRQs
// are connected to XICU in each cluster:
// - IRQ_IN[0] : MMC
// - IRQ_IN[1] : DMA channel 0
// - IRQ_IN[2] : DMA channel 1
// - IRQ_IN[3] : DMA channel 2
// - IRQ_IN[4] : DMA channel 3
//
// All clusters are identical, but cluster(0, 0) and cluster(X_SIZE-1, Y_SIZE-1)
// contain an extra IO bridge component. These IOB0 & IOB1 components are
// connected to the three networks (INT, RAM, IOX).
//
// - It uses two dspin_local_crossbar per cluster to implement the
//   local interconnect correponding to the INT network.
// - It uses three dspin_local_crossbar per cluster to implement the
//   local interconnect correponding to the coherence INT network.
// - It uses two virtual_dspin_router per cluster to implement
//   the INT network (routing both the direct and coherence trafic).
// - It uses two dspin_router per cluster to implement the RAM network.
// - It uses the vci_cc_vcache_wrapper.
// - It uses the vci_mem_cache.
// - It contains one vci_xicu and one vci_multi_dma per cluster.
// - It contains one vci_simple ram per cluster to model the L3 cache.
//
// The TsarIobCluster component is defined in files
// tsar_iob_cluster.* (with * = cpp, h, sd)
//
// The main hardware parameters must be defined in the hard_config.h file :
// - X_SIZE           : number of clusters in a row
// - Y_SIZE           : number of clusters in a column
// - NB_PROCS_MAX     : number of processors per cluster (power of 2)
// - NB_TTY_CHANNELS  : number of TTY channels in I/O network (up to 16)
// - NB_NIC_CHANNELS  : number of NIC channels in I/O network (up to 2)
// - NB_CMA_CHANNELS  : number of CMA channels in I/O network (up to 4)
// - FBUF_X_SIZE      : width of frame buffer (pixels)
// - FBUF_Y_SIZE      : heigth of frame buffer (lines)
// - XCU_NB_INPUTS    : number of HWIs = number of WTIs = number of PTIs
//
// Some secondary hardware parameters must be defined in this top.cpp file:
// - XRAM_LATENCY     : external ram latency
// - MEMC_WAYS        : L2 cache number of ways
// - MEMC_SETS        : L2 cache number of sets
// - L1_IWAYS
// - L1_ISETS
// - L1_DWAYS
// - L1_DSETS
// - BDEV_IMAGE_NAME  : file pathname for block device
// - NIC_RX_NAME      : file pathname for NIC received packets
// - NIC_TX_NAME      : file pathname for NIC transmited packets
// - NIC_TIMEOUT      : max number of cycles before closing a container
//
// General policy for 40 bits physical address decoding:
// All physical segments base addresses are multiple of 1 Mbytes
// (=> the 24 LSB bits = 0, and the 16 MSB bits define the target)
// The (X_WIDTH + Y_WIDTH) MSB bits (left aligned) define
// the cluster index, and the LADR bits define the local index:
//      |X_ID|Y_ID|  LADR  |     OFFSET          |
//      |  4 |  4 |   8    |       24            |
//
// General policy for 14 bits SRCID decoding:
// Each component is identified by (x_id, y_id, l_id) tuple.
//      |X_ID|Y_ID| L_ID |
//      |  4 |  4 |  6   |
/////////////////////////////////////////////////////////////////////////

#include <systemc>
#include <sys/time.h>
#include <iostream>
#include <sstream>
#include <cstdlib>
#include <cstdarg>
#include <climits>
#include <stdint.h>

#include "gdbserver.h"
#include "mapping_table.h"

#include "tsar_iob_cluster.h"
#include "vci_chbuf_dma.h"
#include "vci_multi_tty.h"
#include "vci_multi_nic.h"
#include "vci_simple_rom.h"
#include "vci_block_device_tsar.h"
#include "vci_framebuffer.h"
#include "vci_iox_network.h"
#include "vci_iox_network.h"
#include "vci_iopic.h"
#include "vci_simhelper.h"

#include "alloc_elems.h"

///////////////////////////////////////////////////
//      OS
///////////////////////////////////////////////////
#define USE_ALMOS 0

#define almos_bootloader_pathname "bootloader.bin"
#define almos_kernel_pathname     "kernel-soclib.bin@0xbfc10000:D"
#define almos_archinfo_pathname   "arch-info.bin@0xBFC08000:D"

///////////////////////////////////////////////////
//               Parallelisation
///////////////////////////////////////////////////
#if USE_OPENMP
#include <omp.h>
#endif

///////////////////////////////////////////////////////////
//          DSPIN parameters
///////////////////////////////////////////////////////////

#define dspin_int_cmd_width   39
#define dspin_int_rsp_width   32

#define dspin_ram_cmd_width   64
#define dspin_ram_rsp_width   64

///////////////////////////////////////////////////////////
//         VCI fields width  for the 3 VCI networks
///////////////////////////////////////////////////////////

#define vci_cell_width_int    4
#define vci_cell_width_ext    8

#define vci_plen_width        8
#define vci_address_width     40
#define vci_rerror_width      1
#define vci_clen_width        1
#define vci_rflag_width       1
#define vci_srcid_width       14
#define vci_pktid_width       4
#define vci_trdid_width       4
#define vci_wrplen_width      1

////////////////////////////////////////////////////////////
//    Main Hardware Parameters values
//////////////////////i/////////////////////////////////////

#include "hard_config.h"

////////////////////////////////////////////////////////////
//    Secondary Hardware Parameters values
//////////////////////i/////////////////////////////////////

#define XRAM_LATENCY          0

#define MEMC_WAYS             16
#define MEMC_SETS             256

#define L1_IWAYS              4
#define L1_ISETS              64

#define L1_DWAYS              4
#define L1_DSETS              64

#define BDEV_IMAGE_NAME       "../../../giet_vm/hdd/virt_hdd.dmg"

#define NIC_RX_NAME           "/dev/null"
#define NIC_TX_NAME           "/dev/null"
#define NIC_TIMEOUT           10000

#define NORTH                 0
#define SOUTH                 1
#define EAST                  2
#define WEST                  3

#define cluster(x, y)   ((y) + ((x) << Y_WIDTH))

////////////////////////////////////////////////////////////
//    Software to be loaded in ROM & RAM
//////////////////////i/////////////////////////////////////

#define BOOT_SOFT_NAME        "../../softs/tsar_boot/preloader.elf"

////////////////////////////////////////////////////////////
//     DEBUG Parameters default values
//////////////////////i/////////////////////////////////////

#define MAX_FROZEN_CYCLES     200000

/////////////////////////////////////////////////////////
//    Physical segments definition
/////////////////////////////////////////////////////////

// All physical segments base addresses and sizes are defined
// in the hard_config.h file. For replicated segments, the
// base address is incremented by a cluster offset:
// offset  = cluster(x, y) << (address_width-X_WIDTH-Y_WIDTH);

////////////////////////////////////////////////////////////////////////
//          SRCID definition
////////////////////////////////////////////////////////////////////////
// All initiators are in the same indexing space (14 bits).
// The SRCID is structured in two fields:
// - The 8 MSB bits define the cluster index (left aligned)
// - The 6 LSB bits define the local index.
// Two different initiators cannot have the same SRCID, but a given
// initiator can have two alias SRCIDs:
// - Internal initiators (procs, mdma) are replicated in all clusters,
//   and each initiator has one single SRCID.
// - External initiators (bdev, cdma) are not replicated, but can be
//   accessed in 2 clusters : cluster_iob0 and cluster_iob1.
//   They have the same local index, but two different cluster indexes.
//
// As cluster_iob0 and cluster_iob1 contain both internal initiators
// and external initiators, they must have different local indexes.
// Consequence: For a local interconnect, the INI_ID port index
// is NOT equal to the SRCID local index, and the local interconnect
// must make a translation: SRCID => INI_ID
////////////////////////////////////////////////////////////////////////

#define PROC_LOCAL_SRCID             0x0    // from 0 to 7
#define MDMA_LOCAL_SRCID             0x8
#define IOBX_LOCAL_SRCID             0x9
#define MEMC_LOCAL_SRCID             0xA
#define CDMA_LOCAL_SRCID             0xB
#define BDEV_LOCAL_SRCID             0xC
#define IOPI_LOCAL_SRCID             0xD

///////////////////////////////////////////////////////////////////////
//     TGT_ID and INI_ID port indexing for INT local interconnect
///////////////////////////////////////////////////////////////////////

#define INT_MEMC_TGT_ID              0
#define INT_XICU_TGT_ID              1
#define INT_MDMA_TGT_ID              2
#define INT_BROM_TGT_ID              3
#define INT_IOBX_TGT_ID              4

#define INT_PROC_INI_ID              0   // from 0 to (NB_PROCS_MAX-1)
#define INT_MDMA_INI_ID              (NB_PROCS_MAX)
#define INT_IOBX_INI_ID              (NB_PROCS_MAX+1)

///////////////////////////////////////////////////////////////////////
//     TGT_ID and INI_ID port indexing for RAM local interconnect
///////////////////////////////////////////////////////////////////////

#define RAM_XRAM_TGT_ID              0

#define RAM_MEMC_INI_ID              0
#define RAM_IOBX_INI_ID              1

///////////////////////////////////////////////////////////////////////
//     TGT_ID and INI_ID port indexing for I0X local interconnect
///////////////////////////////////////////////////////////////////////

#define IOX_FBUF_TGT_ID              0
#define IOX_BDEV_TGT_ID              1
#define IOX_MNIC_TGT_ID              2
#define IOX_CDMA_TGT_ID              3
#define IOX_MTTY_TGT_ID              4
#define IOX_IOPI_TGT_ID              5
#define IOX_SIMH_TGT_ID              6
#define IOX_IOB0_TGT_ID              7
#define IOX_IOB1_TGT_ID              8

#define IOX_BDEV_INI_ID              0
#define IOX_CDMA_INI_ID              1
#define IOX_IOPI_INI_ID              2
#define IOX_IOB0_INI_ID              3
#define IOX_IOB1_INI_ID              4

////////////////////////////////////////////////////////////////////////
int _main(int argc, char *argv[])
////////////////////////////////////////////////////////////////////////
{
   using namespace sc_core;
   using namespace soclib::caba;
   using namespace soclib::common;


   char     soft_name[256]   = BOOT_SOFT_NAME;    // pathname: binary code
   size_t   ncycles          = UINT_MAX;          // simulated cycles
   char     disk_name[256]   = BDEV_IMAGE_NAME;   // pathname: disk image
   char     nic_rx_name[256] = NIC_RX_NAME;       // pathname: rx packets file
   char     nic_tx_name[256] = NIC_TX_NAME;       // pathname: tx packets file
   ssize_t  threads_nr       = 1;                 // simulator's threads number
   size_t   faulty_router_id = 0xFFFFFFFF;        // faulty router coordinates
   size_t   faulty_mask      = 0x1F;              // interface mask for the faulty router
   bool     debug_ok         = false;             // trace activated
   size_t   debug_period     = 1;                 // trace period
   size_t   debug_memc_id    = 0xFFFFFFFF;        // index of traced memc
   size_t   debug_proc_id    = 0xFFFFFFFF;        // index of traced proc
   size_t   debug_xram_id    = 0xFFFFFFFF;        // index of traced xram
   bool     debug_iob        = false;             // trace iob0 & iob1 when true
   uint32_t debug_from       = 0;                 // trace start cycle
   uint32_t frozen_cycles    = MAX_FROZEN_CYCLES; // monitoring frozen processor

   assert( (X_WIDTH == 4) and (Y_WIDTH == 4) and
   "ERROR: we must have X_WIDTH == Y_WIDTH == 4");

   ////////////// command line arguments //////////////////////
   if (argc > 1)
   {
      for (int n = 1; n < argc; n = n + 2)
      {
         if ((strcmp(argv[n], "-NCYCLES") == 0) && (n+1<argc))
         {
            ncycles = strtol(argv[n+1], NULL, 0);
         }
         else if ((strcmp(argv[n], "-SOFT") == 0) && (n+1<argc) )
         {
            strcpy(soft_name, argv[n+1]);
         }
         else if ((strcmp(argv[n], "-DEBUG") == 0) && (n+1<argc) )
         {
            debug_ok = true;
            debug_from = strtol(argv[n+1], NULL, 0);
         }
         else if ((strcmp(argv[n], "-DISK") == 0) && (n+1<argc) )
         {
            strcpy(disk_name, argv[n+1]);
         }
         else if ((strcmp(argv[n], "-MEMCID") == 0) && (n+1<argc) )
         {
            debug_memc_id = strtol(argv[n+1], NULL, 0);
            size_t x = debug_memc_id >> Y_WIDTH;
            size_t y = debug_memc_id & ((1 << Y_WIDTH) - 1);
            if( (x>=X_SIZE) || (y>=Y_SIZE) )
            {
                std::cout << "MEMCID parameter does'nt fit X_SIZE/Y_SIZE" << std::endl;
                exit(0);
            }
         }
         else if ((strcmp(argv[n], "-XRAMID") == 0) && (n+1<argc) )
         {
            debug_xram_id = strtol(argv[n+1], NULL, 0);
            size_t x = debug_xram_id >> Y_WIDTH;
            size_t y = debug_xram_id & ((1 << Y_WIDTH) - 1);
            if( (x>=X_SIZE) || (y>=Y_SIZE) )
            {
                std::cout << "XRAMID parameter does'nt fit X_SIZE/Y_SIZE" << std::endl;
                exit(0);
            }
         }
         else if ((strcmp(argv[n], "-IOB") == 0) && (n+1<argc) )
         {
            debug_iob = strtol(argv[n+1], NULL, 0);
         }
         else if ((strcmp(argv[n], "-PROCID") == 0) && (n+1<argc) )
         {
            debug_proc_id     = strtol(argv[n+1], NULL, 0);
            size_t cluster_xy = debug_proc_id >> P_WIDTH;
            size_t x          = cluster_xy >> Y_WIDTH;
            size_t y          = cluster_xy & ((1 << Y_WIDTH) - 1);
            if( (x>=X_SIZE) || (y>=Y_SIZE) )
            {
                std::cout << "PROCID parameter does'nt fit X_SIZE/Y_SIZE" << std::endl;
                exit(0);
            }
         }
         else if ((strcmp(argv[n], "-THREADS") == 0) && ((n+1) < argc))
         {
            threads_nr = strtol(argv[n+1], NULL, 0);
            threads_nr = (threads_nr < 1) ? 1 : threads_nr;
         }
         else if ((strcmp(argv[n], "-FROZEN") == 0) && (n+1 < argc))
         {
            frozen_cycles = strtol(argv[n+1], NULL, 0);
         }
         else if ((strcmp(argv[n], "-PERIOD") == 0) && (n+1 < argc))
         {
            debug_period = strtol(argv[n+1], NULL, 0);
         }
         else if ((strcmp(argv[n], "-FAULTY_ROUTER") == 0) && (n+1 < argc) )
         {
            faulty_router_id = strtol(argv[n+1], NULL, 0);
            size_t x = faulty_router_id >> Y_WIDTH;
            size_t y = faulty_router_id & ((1 << Y_WIDTH) - 1);
            if( (x>=X_SIZE) || (y>=Y_SIZE) )
            {
                std::cout << "FAULTY_ROUTER parameter doesn't fit X_SIZE/Y_SIZE" << std::endl;
                exit(0);
            }
         }
         else if ((strcmp(argv[n], "-FAULTY_MASK") == 0) && (n+1 < argc) )
         {
            faulty_mask = strtol(argv[n+1], NULL, 0);
            if( faulty_mask > 0x1F )
            {
                std::cout << "FAULTY_MASK parameter max value is 0x1F" << std::endl;
                exit(0);
            }
         }
         else
         {
            std::cout << "   Arguments are (key, value) couples." << std::endl;
            std::cout << "   The order is not important." << std::endl;
            std::cout << "   Accepted arguments are :" << std::endl << std::endl;
            std::cout << "     -SOFT pathname_for_embedded_soft" << std::endl;
            std::cout << "     -DISK pathname_for_disk_image" << std::endl;
            std::cout << "     -NCYCLES number_of_simulated_cycles" << std::endl;
            std::cout << "     -DEBUG debug_start_cycle" << std::endl;
            std::cout << "     -THREADS simulator's threads number" << std::endl;
            std::cout << "     -FROZEN max_number_of_lines" << std::endl;
            std::cout << "     -PERIOD number_of_cycles between trace" << std::endl;
            std::cout << "     -MEMCID index_memc_to_be_traced" << std::endl;
            std::cout << "     -XRAMID index_xram_to_be_traced" << std::endl;
            std::cout << "     -PROCID index_proc_to_be_traced" << std::endl;
            std::cout << "     -IOB    non_zero_value" << std::endl;
            exit(0);
         }
      }
   }

   // Activate Distributed Boot (set by environment variable)
   // When this is activated, every processor boots with its instruction and data
   // physical address extension register initialized to its cluster index
   // (X_LOCAL, Y_LOCAL). To support this feature, a distributed ROM is
   // implemented in each cluster.

   const bool distributed_boot = (getenv("DISTRIBUTED_BOOT") != NULL);

   // checking hardware parameters
   assert( (X_SIZE <= (1 << X_WIDTH)) and
           "The X_SIZE parameter cannot be larger than 16" );

   assert( (Y_SIZE <= (1 << Y_WIDTH)) and
           "The Y_SIZE parameter cannot be larger than 16" );

   assert( (NB_PROCS_MAX <= (1 << P_WIDTH)) and
           "NB_PROCS_MAX parameter cannot be larger than 2^P_WIDTH" );

   assert( (NB_DMA_CHANNELS <= 4) and
           "The NB_DMA_CHANNELS parameter cannot be larger than 4" );

   assert( (NB_TTY_CHANNELS >= 1) and (NB_TTY_CHANNELS <= 16) and
           "The NB_TTY_CHANNELS parameter cannot be larger than 16" );

   assert( (NB_NIC_CHANNELS == 2) and
           "The NB_NIC_CHANNELS parameter must be 2" );

   std::cout << std::endl << std::dec
             << " - X_SIZE          = " << X_SIZE << std::endl
             << " - Y_SIZE          = " << Y_SIZE << std::endl
             << " - NB_PROCS_MAX    = " << NB_PROCS_MAX <<  std::endl
             << " - NB_TTY_CHANNELS = " << NB_TTY_CHANNELS <<  std::endl
             << " - NB_DMA_CHANNELS = " << NB_DMA_CHANNELS <<  std::endl
             << " - NB_NIC_CHANNELS = " << NB_NIC_CHANNELS <<  std::endl
             << " - MEMC_WAYS       = " << MEMC_WAYS << std::endl
             << " - MEMC_SETS       = " << MEMC_SETS << std::endl
             << " - RAM_LATENCY     = " << XRAM_LATENCY << std::endl
             << " - MAX_FROZEN      = " << frozen_cycles << std::endl
             << " - DIST_BOOT       = " << distributed_boot << std::endl
             << " - DEBUG_PROCID    = " << debug_proc_id << std::endl
             << " - DEBUG_MEMCID    = " << debug_memc_id << std::endl
             << " - DEBUG_XRAMID    = " << debug_xram_id << std::endl;

   std::cout << std::endl;

#if USE_OPENMP
   omp_set_dynamic(false);
   omp_set_num_threads(threads_nr);
   std::cerr << "Built with openmp version " << _OPENMP << std::endl;
#endif

   // Define VciParams objects
   typedef soclib::caba::VciParams<vci_cell_width_int,
                                   vci_plen_width,
                                   vci_address_width,
                                   vci_rerror_width,
                                   vci_clen_width,
                                   vci_rflag_width,
                                   vci_srcid_width,
                                   vci_pktid_width,
                                   vci_trdid_width,
                                   vci_wrplen_width> vci_param_int;

   typedef soclib::caba::VciParams<vci_cell_width_ext,
                                   vci_plen_width,
                                   vci_address_width,
                                   vci_rerror_width,
                                   vci_clen_width,
                                   vci_rflag_width,
                                   vci_srcid_width,
                                   vci_pktid_width,
                                   vci_trdid_width,
                                   vci_wrplen_width> vci_param_ext;

   const size_t cluster_iob0 = cluster(0, 0);               // cluster containing IOB0
   const size_t cluster_iob1 = cluster(X_SIZE-1, Y_SIZE-1); // cluster containing IOB1

   /////////////////////////////////////////////////////////////////////
   // INT network mapping table
   // - two levels address decoding for commands
   // - two levels srcid decoding for responses
   // - NB_PROCS_MAX + 2 (MDMA, IOBX) local initiators per cluster
   // - 4 local targets (MEMC, XICU, MDMA, IOBX) per cluster
   /////////////////////////////////////////////////////////////////////
   MappingTable maptab_int( vci_address_width,
                            IntTab(X_WIDTH + Y_WIDTH, 16 - X_WIDTH - Y_WIDTH),
                            IntTab(X_WIDTH + Y_WIDTH, vci_srcid_width - X_WIDTH - Y_WIDTH),
                            0x00FF000000);

   for (size_t x = 0; x < X_SIZE; x++)
   {
      for (size_t y = 0; y < Y_SIZE; y++)
      {
         uint64_t offset = ((uint64_t)cluster(x, y))
                              << (vci_address_width-X_WIDTH-Y_WIDTH);
         bool config    = true;
         bool cacheable = true;

         // the four following segments are defined in all clusters

         std::ostringstream    smemc_conf;
         smemc_conf << "int_seg_memc_conf_" << x << "_" << y;
         maptab_int.add(Segment(smemc_conf.str(), SEG_MMC_BASE+offset, SEG_MMC_SIZE,
                     IntTab(cluster(x, y), INT_MEMC_TGT_ID), not cacheable, config ));

         std::ostringstream    smemc_xram;
         smemc_xram << "int_seg_memc_xram_" << x << "_" << y;
         maptab_int.add(Segment(smemc_xram.str(), SEG_RAM_BASE+offset, SEG_RAM_SIZE,
                     IntTab(cluster(x, y), INT_MEMC_TGT_ID), cacheable));

         std::ostringstream    sxicu;
         sxicu << "int_seg_xicu_" << x << "_" << y;
         maptab_int.add(Segment(sxicu.str(), SEG_XCU_BASE+offset, SEG_XCU_SIZE,
                     IntTab(cluster(x, y), INT_XICU_TGT_ID), not cacheable));

         std::ostringstream    smdma;
         smdma << "int_seg_mdma_" << x << "_" << y;
         maptab_int.add(Segment(smdma.str(), SEG_DMA_BASE+offset, SEG_DMA_SIZE,
                     IntTab(cluster(x, y), INT_MDMA_TGT_ID), not cacheable));

         std::ostringstream    sbrom;
         sbrom << "int_seg_brom_" << x << "_" << y;
         maptab_int.add(Segment(sbrom.str(), SEG_ROM_BASE+offset, SEG_ROM_SIZE,
                     IntTab(cluster(x, y), INT_BROM_TGT_ID), cacheable));

         // the following segments are only defined in cluster_iob0 or in cluster_iob1

         if ( (cluster(x, y) == cluster_iob0) or (cluster(x, y) == cluster_iob1) )
         {
            std::ostringstream    siobx;
            siobx << "int_seg_iobx_" << x << "_" << y;
            maptab_int.add(Segment(siobx.str(), SEG_IOB_BASE+offset, SEG_IOB_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable, config ));

            std::ostringstream    stty;
            stty << "int_seg_mtty_" << x << "_" << y;
            maptab_int.add(Segment(stty.str(), SEG_TTY_BASE+offset, SEG_TTY_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    sfbf;
            sfbf << "int_seg_fbuf_" << x << "_" << y;
            maptab_int.add(Segment(sfbf.str(), SEG_FBF_BASE+offset, SEG_FBF_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    sbdv;
            sbdv << "int_seg_bdev_" << x << "_" << y;
            maptab_int.add(Segment(sbdv.str(), SEG_IOC_BASE+offset, SEG_IOC_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    snic;
            snic << "int_seg_mnic_" << x << "_" << y;
            maptab_int.add(Segment(snic.str(), SEG_NIC_BASE+offset, SEG_NIC_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    sdma;
            sdma << "int_seg_cdma_" << x << "_" << y;
            maptab_int.add(Segment(sdma.str(), SEG_CMA_BASE+offset, SEG_CMA_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    spic;
            spic << "int_seg_iopi_" << x << "_" << y;
            maptab_int.add(Segment(spic.str(), SEG_PIC_BASE+offset, SEG_PIC_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));

            std::ostringstream    ssim;
            ssim << "int_seg_simh_" << x << "_" << y;
            maptab_int.add(Segment(ssim.str(), SEG_SIM_BASE+offset, SEG_SIM_SIZE,
                        IntTab(cluster(x, y), INT_IOBX_TGT_ID), not cacheable));
         }

         // This define the mapping between the SRCIDs
         // and the port index on the local interconnect.

         maptab_int.srcid_map( IntTab( cluster(x, y), MDMA_LOCAL_SRCID ),
                               IntTab( cluster(x, y), INT_MDMA_INI_ID ) );

         maptab_int.srcid_map( IntTab( cluster(x, y), IOBX_LOCAL_SRCID ),
                               IntTab( cluster(x, y), INT_IOBX_INI_ID ) );

         maptab_int.srcid_map( IntTab( cluster(x, y), IOPI_LOCAL_SRCID ),
                               IntTab( cluster(x, y), INT_IOBX_INI_ID ) );

         for ( size_t p = 0 ; p < NB_PROCS_MAX ; p++ )
         maptab_int.srcid_map( IntTab( cluster(x, y), PROC_LOCAL_SRCID+p ),
                               IntTab( cluster(x, y), INT_PROC_INI_ID+p ) );
      }
   }
   std::cout << "INT network " << maptab_int << std::endl;

    /////////////////////////////////////////////////////////////////////////
    // RAM network mapping table
    // - two levels address decoding for commands
    // - two levels srcid decoding for responses
    // - 2 local initiators (MEMC, IOBX) per cluster
    //   (IOBX component only in cluster_iob0 and cluster_iob1)
    // - 1 local target (XRAM) per cluster
    ////////////////////////////////////////////////////////////////////////
    MappingTable maptab_ram( vci_address_width,
                             IntTab(X_WIDTH+Y_WIDTH, 0),
                             IntTab(X_WIDTH+Y_WIDTH, vci_srcid_width - X_WIDTH - Y_WIDTH),
                             0x00FF000000);

    for (size_t x = 0; x < X_SIZE; x++)
    {
        for (size_t y = 0; y < Y_SIZE ; y++)
        {
            uint64_t offset = ((uint64_t)cluster(x, y))
                                << (vci_address_width-X_WIDTH-Y_WIDTH);

            std::ostringstream sxram;
            sxram << "ext_seg_xram_" << x << "_" << y;
            maptab_ram.add(Segment(sxram.str(), SEG_RAM_BASE+offset,
                           SEG_RAM_SIZE, IntTab(cluster(x, y), RAM_XRAM_TGT_ID), false));
        }
    }

    // This define the mapping between the initiators SRCID
    // and the port index on the RAM local interconnect.
    // External initiator have two alias SRCID (iob0 / iob1)

    maptab_ram.srcid_map( IntTab( cluster_iob0, CDMA_LOCAL_SRCID ),
                          IntTab( cluster_iob0, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob1, CDMA_LOCAL_SRCID ),
                          IntTab( cluster_iob1, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob0, BDEV_LOCAL_SRCID ),
                          IntTab( cluster_iob0, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob1, BDEV_LOCAL_SRCID ),
                          IntTab( cluster_iob1, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob0, IOPI_LOCAL_SRCID ),
                          IntTab( cluster_iob0, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob1, IOPI_LOCAL_SRCID ),
                          IntTab( cluster_iob1, RAM_IOBX_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob0, MEMC_LOCAL_SRCID ),
                          IntTab( cluster_iob0, RAM_MEMC_INI_ID ) );

    maptab_ram.srcid_map( IntTab( cluster_iob1, MEMC_LOCAL_SRCID ),
                          IntTab( cluster_iob1, RAM_MEMC_INI_ID ) );

    std::cout << "RAM network " << maptab_ram << std::endl;

    ///////////////////////////////////////////////////////////////////////
    // IOX network mapping table
    // - two levels address decoding for commands (9, 7) bits
    // - two levels srcid decoding for responses
    // - 5 initiators (IOB0, IOB1, BDEV, CDMA, IOPI)
    // - 9 targets (IOB0, IOB1, BDEV, CDMA, MTTY, FBUF, BROM, MNIC, IOPI)
    //
    // Address bit 32 is used to determine if a command must be routed to
    // IOB0 or IOB1.
    ///////////////////////////////////////////////////////////////////////
    MappingTable maptab_iox(
          vci_address_width,
          IntTab(X_WIDTH + Y_WIDTH - 1, 16 - X_WIDTH - Y_WIDTH + 1),
          IntTab(X_WIDTH + Y_WIDTH    , vci_param_ext::S - X_WIDTH - Y_WIDTH),
          0x00FF000000);

    // External peripherals segments
    // When there is more than one cluster, external peripherals can be accessed
    // through two segments, depending on the used IOB (IOB0 or IOB1).

    const uint64_t iob0_base = ((uint64_t)cluster_iob0)
       << (vci_address_width - X_WIDTH - Y_WIDTH);

    maptab_iox.add(Segment("iox_seg_mtty_0", SEG_TTY_BASE + iob0_base, SEG_TTY_SIZE,
                   IntTab(0, IOX_MTTY_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_fbuf_0", SEG_FBF_BASE + iob0_base, SEG_FBF_SIZE,
                   IntTab(0, IOX_FBUF_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_bdev_0", SEG_IOC_BASE + iob0_base, SEG_IOC_SIZE,
                   IntTab(0, IOX_BDEV_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_mnic_0", SEG_NIC_BASE + iob0_base, SEG_NIC_SIZE,
                   IntTab(0, IOX_MNIC_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_cdma_0", SEG_CMA_BASE + iob0_base, SEG_CMA_SIZE,
                   IntTab(0, IOX_CDMA_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_iopi_0", SEG_PIC_BASE + iob0_base, SEG_PIC_SIZE,
                   IntTab(0, IOX_IOPI_TGT_ID), false));
    maptab_iox.add(Segment("iox_seg_simh_0", SEG_SIM_BASE + iob0_base, SEG_SIM_SIZE,
                   IntTab(0, IOX_SIMH_TGT_ID), false));

    if ( cluster_iob0 != cluster_iob1 )
    {
       const uint64_t iob1_base = ((uint64_t)cluster_iob1)
          << (vci_address_width - X_WIDTH - Y_WIDTH);

        maptab_iox.add(Segment("iox_seg_mtty_1", SEG_TTY_BASE + iob1_base, SEG_TTY_SIZE,
                   IntTab(0, IOX_MTTY_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_fbuf_1", SEG_FBF_BASE + iob1_base, SEG_FBF_SIZE,
                   IntTab(0, IOX_FBUF_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_bdev_1", SEG_IOC_BASE + iob1_base, SEG_IOC_SIZE,
                   IntTab(0, IOX_BDEV_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_mnic_1", SEG_NIC_BASE + iob1_base, SEG_NIC_SIZE,
                   IntTab(0, IOX_MNIC_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_cdma_1", SEG_CMA_BASE + iob1_base, SEG_CMA_SIZE,
                   IntTab(0, IOX_CDMA_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_iopi_1", SEG_PIC_BASE + iob1_base, SEG_PIC_SIZE,
                   IntTab(0, IOX_IOPI_TGT_ID), false));
        maptab_iox.add(Segment("iox_seg_simh_1", SEG_SIM_BASE + iob1_base, SEG_SIM_SIZE,
                   IntTab(0, IOX_SIMH_TGT_ID), false));
    }

    // If there is more than one cluster, external peripherals
    // can access RAM through two segments (IOB0 / IOB1).
    // As IOMMU is not activated, addresses are 40 bits (physical addresses),
    // and the choice depends on address bit A[32].
    for (size_t x = 0; x < X_SIZE; x++)
    {
        for (size_t y = 0; y < Y_SIZE ; y++)
        {
            const bool wti       = true;
            const bool cacheable = true;

            const uint64_t offset = ((uint64_t)cluster(x, y))
                << (vci_address_width-X_WIDTH-Y_WIDTH);

            const uint64_t xicu_base = SEG_XCU_BASE + offset;

            if ( (y & 0x1) == 0 ) // use IOB0
            {
                std::ostringstream sxcu0;
                sxcu0 << "iox_seg_xcu0_" << x << "_" << y;
                maptab_iox.add(Segment(sxcu0.str(), xicu_base, SEG_XCU_SIZE,
                            IntTab(0, IOX_IOB0_TGT_ID), not cacheable, wti));

                std::ostringstream siob0;
                siob0 << "iox_seg_ram0_" << x << "_" << y;
                maptab_iox.add(Segment(siob0.str(), offset, SEG_XCU_BASE,
                            IntTab(0, IOX_IOB0_TGT_ID), not cacheable, not wti));
            }
            else                  // USE IOB1
            {
                std::ostringstream sxcu1;
                sxcu1 << "iox_seg_xcu1_" << x << "_" << y;
                maptab_iox.add(Segment(sxcu1.str(), xicu_base, SEG_XCU_SIZE,
                            IntTab(0, IOX_IOB1_TGT_ID), not cacheable, wti));

                std::ostringstream siob1;
                siob1 << "iox_seg_ram1_" << x << "_" << y;
                maptab_iox.add(Segment(siob1.str(), offset, SEG_XCU_BASE,
                            IntTab(0, IOX_IOB1_TGT_ID), not cacheable, not wti));
            }
        }
    }

    // This define the mapping between the external initiators (SRCID)
    // and the port index on the IOX local interconnect.

    maptab_iox.srcid_map( IntTab( 0, CDMA_LOCAL_SRCID ) ,
                          IntTab( 0, IOX_CDMA_INI_ID  ) );
    maptab_iox.srcid_map( IntTab( 0, BDEV_LOCAL_SRCID ) ,
                          IntTab( 0, IOX_BDEV_INI_ID  ) );
    maptab_iox.srcid_map( IntTab( 0, IOPI_LOCAL_SRCID ) ,
                          IntTab( 0, IOX_IOPI_INI_ID  ) );
    maptab_iox.srcid_map( IntTab( 0, IOX_IOB0_INI_ID  ) ,
                          IntTab( 0, IOX_IOB0_INI_ID  ) );

    if ( cluster_iob0 != cluster_iob1 )
    {
        maptab_iox.srcid_map( IntTab( 0, IOX_IOB1_INI_ID ) ,
                              IntTab( 0, IOX_IOB1_INI_ID ) );
    }

    std::cout << "IOX network " << maptab_iox << std::endl;

    ////////////////////
    // Signals
    ///////////////////

    sc_clock                          signal_clk("clk");
    sc_signal<bool>                   signal_resetn("resetn");

    sc_signal<bool>                   signal_irq_false;
    sc_signal<bool>                   signal_irq_bdev;
    sc_signal<bool>                   signal_irq_mtty_rx[NB_TTY_CHANNELS];
    sc_signal<bool>                   signal_irq_mnic_rx[NB_NIC_CHANNELS];
    sc_signal<bool>                   signal_irq_mnic_tx[NB_NIC_CHANNELS];
    sc_signal<bool>                   signal_irq_cdma[NB_CMA_CHANNELS];

    // VCI signals for IOX network
    VciSignals<vci_param_ext>         signal_vci_ini_iob0("signal_vci_ini_iob0");
    VciSignals<vci_param_ext>         signal_vci_ini_iob1("signal_vci_ini_iob1");
    VciSignals<vci_param_ext>         signal_vci_ini_bdev("signal_vci_ini_bdev");
    VciSignals<vci_param_ext>         signal_vci_ini_cdma("signal_vci_ini_cdma");
    VciSignals<vci_param_ext>         signal_vci_ini_iopi("signal_vci_ini_iopi");

    VciSignals<vci_param_ext>         signal_vci_tgt_iob0("signal_vci_tgt_iob0");
    VciSignals<vci_param_ext>         signal_vci_tgt_iob1("signal_vci_tgt_iob1");
    VciSignals<vci_param_ext>         signal_vci_tgt_mtty("signal_vci_tgt_mtty");
    VciSignals<vci_param_ext>         signal_vci_tgt_fbuf("signal_vci_tgt_fbuf");
    VciSignals<vci_param_ext>         signal_vci_tgt_mnic("signal_vci_tgt_mnic");
    VciSignals<vci_param_ext>         signal_vci_tgt_bdev("signal_vci_tgt_bdev");
    VciSignals<vci_param_ext>         signal_vci_tgt_cdma("signal_vci_tgt_cdma");
    VciSignals<vci_param_ext>         signal_vci_tgt_iopi("signal_vci_ini_iopi");
    VciSignals<vci_param_ext>         signal_vci_tgt_simh("signal_vci_ini_simh");

   // Horizontal inter-clusters INT network DSPIN
   DspinSignals<dspin_int_cmd_width>*** signal_dspin_int_cmd_h_inc =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_int_cmd_h_inc", X_SIZE-1, Y_SIZE, 3);
   DspinSignals<dspin_int_cmd_width>*** signal_dspin_int_cmd_h_dec =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_int_cmd_h_dec", X_SIZE-1, Y_SIZE, 3);
   DspinSignals<dspin_int_rsp_width>*** signal_dspin_int_rsp_h_inc =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_int_rsp_h_inc", X_SIZE-1, Y_SIZE, 2);
   DspinSignals<dspin_int_rsp_width>*** signal_dspin_int_rsp_h_dec =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_int_rsp_h_dec", X_SIZE-1, Y_SIZE, 2);

   // Vertical inter-clusters INT network DSPIN
   DspinSignals<dspin_int_cmd_width>*** signal_dspin_int_cmd_v_inc =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_int_cmd_v_inc", X_SIZE, Y_SIZE-1, 3);
   DspinSignals<dspin_int_cmd_width>*** signal_dspin_int_cmd_v_dec =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_int_cmd_v_dec", X_SIZE, Y_SIZE-1, 3);
   DspinSignals<dspin_int_rsp_width>*** signal_dspin_int_rsp_v_inc =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_int_rsp_v_inc", X_SIZE, Y_SIZE-1, 2);
   DspinSignals<dspin_int_rsp_width>*** signal_dspin_int_rsp_v_dec =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_int_rsp_v_dec", X_SIZE, Y_SIZE-1, 2);

   // Mesh boundaries INT network DSPIN
   DspinSignals<dspin_int_cmd_width>**** signal_dspin_false_int_cmd_in =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_false_int_cmd_in", X_SIZE, Y_SIZE, 4, 3);
   DspinSignals<dspin_int_cmd_width>**** signal_dspin_false_int_cmd_out =
      alloc_elems<DspinSignals<dspin_int_cmd_width> >("signal_dspin_false_int_cmd_out", X_SIZE, Y_SIZE, 4, 3);
   DspinSignals<dspin_int_rsp_width>**** signal_dspin_false_int_rsp_in =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_false_int_rsp_in", X_SIZE, Y_SIZE, 4, 2);
   DspinSignals<dspin_int_rsp_width>**** signal_dspin_false_int_rsp_out =
      alloc_elems<DspinSignals<dspin_int_rsp_width> >("signal_dspin_false_int_rsp_out", X_SIZE, Y_SIZE, 4, 2);


   // Horizontal inter-clusters RAM network DSPIN
   DspinSignals<dspin_ram_cmd_width>** signal_dspin_ram_cmd_h_inc =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_ram_cmd_h_inc", X_SIZE-1, Y_SIZE);
   DspinSignals<dspin_ram_cmd_width>** signal_dspin_ram_cmd_h_dec =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_ram_cmd_h_dec", X_SIZE-1, Y_SIZE);
   DspinSignals<dspin_ram_rsp_width>** signal_dspin_ram_rsp_h_inc =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_ram_rsp_h_inc", X_SIZE-1, Y_SIZE);
   DspinSignals<dspin_ram_rsp_width>** signal_dspin_ram_rsp_h_dec =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_ram_rsp_h_dec", X_SIZE-1, Y_SIZE);

   // Vertical inter-clusters RAM network DSPIN
   DspinSignals<dspin_ram_cmd_width>** signal_dspin_ram_cmd_v_inc =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_ram_cmd_v_inc", X_SIZE, Y_SIZE-1);
   DspinSignals<dspin_ram_cmd_width>** signal_dspin_ram_cmd_v_dec =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_ram_cmd_v_dec", X_SIZE, Y_SIZE-1);
   DspinSignals<dspin_ram_rsp_width>** signal_dspin_ram_rsp_v_inc =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_ram_rsp_v_inc", X_SIZE, Y_SIZE-1);
   DspinSignals<dspin_ram_rsp_width>** signal_dspin_ram_rsp_v_dec =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_ram_rsp_v_dec", X_SIZE, Y_SIZE-1);

   // Mesh boundaries RAM network DSPIN
   DspinSignals<dspin_ram_cmd_width>*** signal_dspin_false_ram_cmd_in =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_false_ram_cmd_in", X_SIZE, Y_SIZE, 4);
   DspinSignals<dspin_ram_cmd_width>*** signal_dspin_false_ram_cmd_out =
      alloc_elems<DspinSignals<dspin_ram_cmd_width> >("signal_dspin_false_ram_cmd_out", X_SIZE, Y_SIZE, 4);
   DspinSignals<dspin_ram_rsp_width>*** signal_dspin_false_ram_rsp_in =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_false_ram_rsp_in", X_SIZE, Y_SIZE, 4);
   DspinSignals<dspin_ram_rsp_width>*** signal_dspin_false_ram_rsp_out =
      alloc_elems<DspinSignals<dspin_ram_rsp_width> >("signal_dspin_false_ram_rsp_out", X_SIZE, Y_SIZE, 4);

   ////////////////////////////
   //      Loader
   ////////////////////////////

#if USE_ALMOS
   soclib::common::Loader loader(almos_bootloader_pathname,
                                 almos_archinfo_pathname,
                                 almos_kernel_pathname);
#else
   soclib::common::Loader loader(soft_name);
#endif

   // initialize memory with a value different than 0 (expose software errors
   // dues to uninitialized data)
   loader.memory_default(0xA0);

   typedef soclib::common::GdbServer<soclib::common::Mips32ElIss> proc_iss;
   proc_iss::set_loader(loader);

   ////////////////////////////////////////
   //  Instanciated Hardware Components
   ////////////////////////////////////////

   std::cout << std::endl << "External Bus and Peripherals" << std::endl << std::endl;

   const size_t nb_iox_initiators = (cluster_iob0 != cluster_iob1) ? 5 : 4;
   const size_t nb_iox_targets = (cluster_iob0 != cluster_iob1) ? 9 : 8;

   // IOX network
   VciIoxNetwork<vci_param_ext>* iox_network;
   iox_network = new VciIoxNetwork<vci_param_ext>( "iox_network",
                                                   maptab_iox,
                                                   nb_iox_targets,
                                                   nb_iox_initiators );

   // Network Controller
   VciMultiNic<vci_param_ext>*  mnic;
   mnic = new VciMultiNic<vci_param_ext>( "mnic",
                                          IntTab(0, IOX_MNIC_TGT_ID),
                                          maptab_iox,
                                          NB_NIC_CHANNELS,
                                          0,           // mac_4 address
                                          0,           // mac_2 address
                                          nic_rx_name,
                                          nic_tx_name);

   // Frame Buffer
   VciFrameBuffer<vci_param_ext>*  fbuf;
   fbuf = new VciFrameBuffer<vci_param_ext>( "fbuf",
                                             IntTab(0, IOX_FBUF_TGT_ID),
                                             maptab_iox,
                                             FBUF_X_SIZE, FBUF_Y_SIZE );

   // Block Device
   // for AHCI
   // std::vector<std::string> filenames;
   // filenames.push_back(disk_name);            // one single disk
   VciBlockDeviceTsar<vci_param_ext>*  bdev;
   bdev = new VciBlockDeviceTsar<vci_param_ext>( "bdev",
                                                  maptab_iox,
                                                  IntTab(0, BDEV_LOCAL_SRCID),
                                                  IntTab(0, IOX_BDEV_TGT_ID),
                                                  disk_name,
                                                  512,        // block size
                                                  64,         // burst size (bytes)
                                                  0 );        // disk latency

   // Chained Buffer DMA controller
   VciChbufDma<vci_param_ext>*  cdma;
   cdma = new VciChbufDma<vci_param_ext>( "cdma",
                                          maptab_iox,
                                          IntTab(0, CDMA_LOCAL_SRCID),
                                          IntTab(0, IOX_CDMA_TGT_ID),
                                          64,          // burst size (bytes)
                                          2*NB_NIC_CHANNELS );
   // Multi-TTY controller
   std::vector<std::string> vect_names;
   for( size_t tid = 0 ; tid < NB_TTY_CHANNELS ; tid++ )
   {
      std::ostringstream term_name;
         term_name <<  "term" << tid;
         vect_names.push_back(term_name.str().c_str());
      }
      VciMultiTty<vci_param_ext>*  mtty;
      mtty = new VciMultiTty<vci_param_ext>( "mtty",
                                             IntTab(0, IOX_MTTY_TGT_ID),
                                             maptab_iox,
                                             vect_names);

   // IOPIC
   VciIopic<vci_param_ext>* iopi;
   iopi = new VciIopic<vci_param_ext>( "iopi",
                                       maptab_iox,
                                       IntTab(0, IOPI_LOCAL_SRCID),
                                       IntTab(0, IOX_IOPI_TGT_ID),
                                       32 );        // number of input HWI

   // Simhelper
   VciSimhelper<vci_param_ext>* simh;
   simh = new VciSimhelper<vci_param_ext>("simh",
                                          IntTab(0, IOX_SIMH_TGT_ID),
                                          maptab_iox );

   // Clusters
   TsarIobCluster<vci_param_int,
                  vci_param_ext,
                  dspin_int_cmd_width,
                  dspin_int_rsp_width,
                  dspin_ram_cmd_width,
                  dspin_ram_rsp_width>* clusters[X_SIZE][Y_SIZE];

#if USE_OPENMP
#pragma omp parallel
    {
#pragma omp for
#endif
        for(size_t i = 0; i  < (X_SIZE * Y_SIZE); i++)
        {
            size_t x = i / Y_SIZE;
            size_t y = i % Y_SIZE;

#if USE_OPENMP
#pragma omp critical
            {
#endif
            std::cout << std::endl;
            std::cout << "Cluster_" << std::dec << x << "_" << y << std::endl;
            std::cout << std::endl;

            const bool is_iob0 = (cluster(x, y) == cluster_iob0);
            const bool is_iob1 = (cluster(x, y) == cluster_iob1);
            const bool is_io_cluster = is_iob0 || is_iob1;

            const int iox_iob_ini_id = is_iob0 ?
                IOX_IOB0_INI_ID :
                IOX_IOB1_INI_ID ;
            const int iox_iob_tgt_id = is_iob0 ?
                IOX_IOB0_TGT_ID :
                IOX_IOB1_TGT_ID ;

            std::ostringstream sc;
            sc << "cluster_" << x << "_" << y;
            clusters[x][y] = new TsarIobCluster<vci_param_int,
                                                vci_param_ext,
                                                dspin_int_cmd_width,
                                                dspin_int_rsp_width,
                                                dspin_ram_cmd_width,
                                                dspin_ram_rsp_width>
            (
                sc.str().c_str(),
                NB_PROCS_MAX,
                NB_DMA_CHANNELS,
                x,
                y,
                X_SIZE,
                Y_SIZE,

                P_WIDTH,

                maptab_int,
                maptab_ram,
                maptab_iox,

                X_WIDTH,
                Y_WIDTH,
                vci_srcid_width - X_WIDTH - Y_WIDTH,            // l_id width,

                INT_MEMC_TGT_ID,
                INT_XICU_TGT_ID,
                INT_MDMA_TGT_ID,
                INT_BROM_TGT_ID,
                INT_IOBX_TGT_ID,

                INT_PROC_INI_ID,
                INT_MDMA_INI_ID,
                INT_IOBX_INI_ID,

                RAM_XRAM_TGT_ID,

                RAM_MEMC_INI_ID,
                RAM_IOBX_INI_ID,

                is_io_cluster,
                iox_iob_tgt_id,
                iox_iob_ini_id,

                MEMC_WAYS,
                MEMC_SETS,
                L1_IWAYS,
                L1_ISETS,
                L1_DWAYS,
                L1_DSETS,
                XRAM_LATENCY,
                XCU_NB_INPUTS,

                distributed_boot,

                loader,

                frozen_cycles,
                debug_from,
                debug_ok and (cluster(x, y) == debug_memc_id),
                debug_ok and (cluster(x, y) == (debug_proc_id >> P_WIDTH)),
                debug_ok and debug_iob
            );

#if USE_OPENMP
            } // end critical
#endif
        } // end for
#if USE_OPENMP
    }
#endif

    // disable all interfaces of the faulty router
    if (faulty_router_id != 0xFFFFFFFF)
    {
       int faulty_x = faulty_router_id >> Y_WIDTH;
       int faulty_y = faulty_router_id & ((1 << Y_WIDTH) - 1);
       clusters[faulty_x][faulty_y]->int_router_cmd[0]->set_disable_mask(faulty_mask);
    }

    std::cout << std::endl;

    ///////////////////////////////////////////////////////////////////////////////
    //     Net-list
    ///////////////////////////////////////////////////////////////////////////////

    // IOX network connexion
    iox_network->p_clk                                   (signal_clk);
    iox_network->p_resetn                                (signal_resetn);
    iox_network->p_to_ini[IOX_IOB0_INI_ID]               (signal_vci_ini_iob0);
    iox_network->p_to_ini[IOX_BDEV_INI_ID]               (signal_vci_ini_bdev);
    iox_network->p_to_ini[IOX_CDMA_INI_ID]               (signal_vci_ini_cdma);
    iox_network->p_to_ini[IOX_IOPI_INI_ID]               (signal_vci_ini_iopi);

    iox_network->p_to_tgt[IOX_IOB0_TGT_ID]               (signal_vci_tgt_iob0);
    iox_network->p_to_tgt[IOX_MTTY_TGT_ID]               (signal_vci_tgt_mtty);
    iox_network->p_to_tgt[IOX_FBUF_TGT_ID]               (signal_vci_tgt_fbuf);
    iox_network->p_to_tgt[IOX_MNIC_TGT_ID]               (signal_vci_tgt_mnic);
    iox_network->p_to_tgt[IOX_BDEV_TGT_ID]               (signal_vci_tgt_bdev);
    iox_network->p_to_tgt[IOX_CDMA_TGT_ID]               (signal_vci_tgt_cdma);
    iox_network->p_to_tgt[IOX_IOPI_TGT_ID]               (signal_vci_tgt_iopi);
    iox_network->p_to_tgt[IOX_SIMH_TGT_ID]               (signal_vci_tgt_simh);

    if (cluster_iob0 != cluster_iob1)
    {
        iox_network->p_to_ini[IOX_IOB1_INI_ID]           (signal_vci_ini_iob1);
        iox_network->p_to_tgt[IOX_IOB1_TGT_ID]           (signal_vci_tgt_iob1);
    }

    // BDEV connexion
    bdev->p_clk                                          (signal_clk);
    bdev->p_resetn                                       (signal_resetn);
    bdev->p_irq                                          (signal_irq_bdev);
    bdev->p_vci_target                                   (signal_vci_tgt_bdev);
    bdev->p_vci_initiator                                (signal_vci_ini_bdev);

    std::cout << "  - BDEV connected" << std::endl;

    // FBUF connexion
    fbuf->p_clk                                          (signal_clk);
    fbuf->p_resetn                                       (signal_resetn);
    fbuf->p_vci                                          (signal_vci_tgt_fbuf);

    std::cout << "  - FBUF connected" << std::endl;

    // MNIC connexion
    mnic->p_clk                                          (signal_clk);
    mnic->p_resetn                                       (signal_resetn);
    mnic->p_vci                                          (signal_vci_tgt_mnic);
    for ( size_t i=0 ; i<NB_NIC_CHANNELS ; i++ )
    {
         mnic->p_rx_irq[i]                               (signal_irq_mnic_rx[i]);
         mnic->p_tx_irq[i]                               (signal_irq_mnic_tx[i]);
    }

    std::cout << "  - MNIC connected" << std::endl;

    // MTTY connexion
    mtty->p_clk                                          (signal_clk);
    mtty->p_resetn                                       (signal_resetn);
    mtty->p_vci                                          (signal_vci_tgt_mtty);
    for ( size_t i=0 ; i<NB_TTY_CHANNELS ; i++ )
    {
        mtty->p_irq[i]                                   (signal_irq_mtty_rx[i]);
    }
    std::cout << "  - MTTY connected" << std::endl;

    // CDMA connexion
    cdma->p_clk                                          (signal_clk);
    cdma->p_resetn                                       (signal_resetn);
    cdma->p_vci_target                                   (signal_vci_tgt_cdma);
    cdma->p_vci_initiator                                (signal_vci_ini_cdma);
    for ( size_t i=0 ; i<(NB_NIC_CHANNELS*2) ; i++)
    {
        cdma->p_irq[i]                                   (signal_irq_cdma[i]);
    }

    std::cout << "  - CDMA connected" << std::endl;

    // IOPI connexion
    iopi->p_clk                                          (signal_clk);
    iopi->p_resetn                                       (signal_resetn);
    iopi->p_vci_target                                   (signal_vci_tgt_iopi);
    iopi->p_vci_initiator                                (signal_vci_ini_iopi);
    for ( size_t i=0 ; i<32 ; i++)
    {
       if     (i < NB_NIC_CHANNELS)    iopi->p_hwi[i] (signal_irq_mnic_rx[i]);
       else if(i < 2 )                 iopi->p_hwi[i] (signal_irq_false);
       else if(i < 2+NB_NIC_CHANNELS)  iopi->p_hwi[i] (signal_irq_mnic_tx[i-2]);
       else if(i < 4 )                 iopi->p_hwi[i] (signal_irq_false);
       else if(i < 4+NB_CMA_CHANNELS)  iopi->p_hwi[i] (signal_irq_cdma[i-4]);
       else if(i < 8)                  iopi->p_hwi[i] (signal_irq_false);
       else if(i < 9)                  iopi->p_hwi[i] (signal_irq_bdev);
       else if(i < 16)                 iopi->p_hwi[i] (signal_irq_false);
       else if(i < 16+NB_TTY_CHANNELS) iopi->p_hwi[i] (signal_irq_mtty_rx[i-16]);
       else                            iopi->p_hwi[i] (signal_irq_false);
    }

    std::cout << "  - IOPIC connected" << std::endl;

    // Simhelper connexion
    simh->p_clk(signal_clk);
    simh->p_resetn(signal_resetn);
    simh->p_vci(signal_vci_tgt_simh);

    // IOB0 cluster connexion to IOX network
    (*clusters[0][0]->p_vci_iob_iox_ini) (signal_vci_ini_iob0);
    (*clusters[0][0]->p_vci_iob_iox_tgt) (signal_vci_tgt_iob0);

    // IOB1 cluster connexion to IOX network
    // (only when there is more than 1 cluster)
    if ( cluster_iob0 != cluster_iob1 )
    {
        (*clusters[X_SIZE-1][Y_SIZE-1]->p_vci_iob_iox_ini) (signal_vci_ini_iob1);
        (*clusters[X_SIZE-1][Y_SIZE-1]->p_vci_iob_iox_tgt) (signal_vci_tgt_iob1);
    }

    // All clusters Clock & RESET connexions
    for ( size_t x = 0; x < (X_SIZE); x++ )
    {
        for (size_t y = 0; y < Y_SIZE; y++)
        {
            clusters[x][y]->p_clk     (signal_clk);
            clusters[x][y]->p_resetn  (signal_resetn);
        }
    }

   // Inter Clusters horizontal connections
   if (X_SIZE > 1)
   {
      for (size_t x = 0; x < (X_SIZE-1); x++)
      {
         for (size_t y = 0; y < Y_SIZE; y++)
         {
            for (size_t k = 0; k < 3; k++)
            {
               clusters[x][y]->p_dspin_int_cmd_out[EAST][k]      (signal_dspin_int_cmd_h_inc[x][y][k]);
               clusters[x+1][y]->p_dspin_int_cmd_in[WEST][k]     (signal_dspin_int_cmd_h_inc[x][y][k]);
               clusters[x][y]->p_dspin_int_cmd_in[EAST][k]       (signal_dspin_int_cmd_h_dec[x][y][k]);
               clusters[x+1][y]->p_dspin_int_cmd_out[WEST][k]    (signal_dspin_int_cmd_h_dec[x][y][k]);
            }

            for (size_t k = 0; k < 2; k++)
            {
               clusters[x][y]->p_dspin_int_rsp_out[EAST][k]      (signal_dspin_int_rsp_h_inc[x][y][k]);
               clusters[x+1][y]->p_dspin_int_rsp_in[WEST][k]     (signal_dspin_int_rsp_h_inc[x][y][k]);
               clusters[x][y]->p_dspin_int_rsp_in[EAST][k]       (signal_dspin_int_rsp_h_dec[x][y][k]);
               clusters[x+1][y]->p_dspin_int_rsp_out[WEST][k]    (signal_dspin_int_rsp_h_dec[x][y][k]);
            }

            clusters[x][y]->p_dspin_ram_cmd_out[EAST]      (signal_dspin_ram_cmd_h_inc[x][y]);
            clusters[x+1][y]->p_dspin_ram_cmd_in[WEST]     (signal_dspin_ram_cmd_h_inc[x][y]);
            clusters[x][y]->p_dspin_ram_cmd_in[EAST]       (signal_dspin_ram_cmd_h_dec[x][y]);
            clusters[x+1][y]->p_dspin_ram_cmd_out[WEST]    (signal_dspin_ram_cmd_h_dec[x][y]);
            clusters[x][y]->p_dspin_ram_rsp_out[EAST]      (signal_dspin_ram_rsp_h_inc[x][y]);
            clusters[x+1][y]->p_dspin_ram_rsp_in[WEST]     (signal_dspin_ram_rsp_h_inc[x][y]);
            clusters[x][y]->p_dspin_ram_rsp_in[EAST]       (signal_dspin_ram_rsp_h_dec[x][y]);
            clusters[x+1][y]->p_dspin_ram_rsp_out[WEST]    (signal_dspin_ram_rsp_h_dec[x][y]);
         }
      }
   }

   std::cout << std::endl << "Horizontal connections established" << std::endl;

   // Inter Clusters vertical connections
   if (Y_SIZE > 1)
   {
      for (size_t y = 0; y < (Y_SIZE-1); y++)
      {
         for (size_t x = 0; x < X_SIZE; x++)
         {
            for (size_t k = 0; k < 3; k++)
            {
               clusters[x][y]->p_dspin_int_cmd_out[NORTH][k]     (signal_dspin_int_cmd_v_inc[x][y][k]);
               clusters[x][y+1]->p_dspin_int_cmd_in[SOUTH][k]    (signal_dspin_int_cmd_v_inc[x][y][k]);
               clusters[x][y]->p_dspin_int_cmd_in[NORTH][k]      (signal_dspin_int_cmd_v_dec[x][y][k]);
               clusters[x][y+1]->p_dspin_int_cmd_out[SOUTH][k]   (signal_dspin_int_cmd_v_dec[x][y][k]);
            }

            for (size_t k = 0; k < 2; k++)
            {
               clusters[x][y]->p_dspin_int_rsp_out[NORTH][k]     (signal_dspin_int_rsp_v_inc[x][y][k]);
               clusters[x][y+1]->p_dspin_int_rsp_in[SOUTH][k]    (signal_dspin_int_rsp_v_inc[x][y][k]);
               clusters[x][y]->p_dspin_int_rsp_in[NORTH][k]      (signal_dspin_int_rsp_v_dec[x][y][k]);
               clusters[x][y+1]->p_dspin_int_rsp_out[SOUTH][k]   (signal_dspin_int_rsp_v_dec[x][y][k]);
            }

            clusters[x][y]->p_dspin_ram_cmd_out[NORTH]     (signal_dspin_ram_cmd_v_inc[x][y]);
            clusters[x][y+1]->p_dspin_ram_cmd_in[SOUTH]    (signal_dspin_ram_cmd_v_inc[x][y]);
            clusters[x][y]->p_dspin_ram_cmd_in[NORTH]      (signal_dspin_ram_cmd_v_dec[x][y]);
            clusters[x][y+1]->p_dspin_ram_cmd_out[SOUTH]   (signal_dspin_ram_cmd_v_dec[x][y]);
            clusters[x][y]->p_dspin_ram_rsp_out[NORTH]     (signal_dspin_ram_rsp_v_inc[x][y]);
            clusters[x][y+1]->p_dspin_ram_rsp_in[SOUTH]    (signal_dspin_ram_rsp_v_inc[x][y]);
            clusters[x][y]->p_dspin_ram_rsp_in[NORTH]      (signal_dspin_ram_rsp_v_dec[x][y]);
            clusters[x][y+1]->p_dspin_ram_rsp_out[SOUTH]   (signal_dspin_ram_rsp_v_dec[x][y]);
         }
      }
   }

   std::cout << "Vertical connections established" << std::endl;

   // East & West boundary cluster connections
   for (size_t y = 0; y < Y_SIZE; y++)
   {
      for (size_t k = 0; k < 3; k++)
      {
         clusters[0][y]->p_dspin_int_cmd_in[WEST][k]          (signal_dspin_false_int_cmd_in[0][y][WEST][k]);
         clusters[0][y]->p_dspin_int_cmd_out[WEST][k]         (signal_dspin_false_int_cmd_out[0][y][WEST][k]);
         clusters[X_SIZE-1][y]->p_dspin_int_cmd_in[EAST][k]   (signal_dspin_false_int_cmd_in[X_SIZE-1][y][EAST][k]);
         clusters[X_SIZE-1][y]->p_dspin_int_cmd_out[EAST][k]  (signal_dspin_false_int_cmd_out[X_SIZE-1][y][EAST][k]);
      }

      for (size_t k = 0; k < 2; k++)
      {
         clusters[0][y]->p_dspin_int_rsp_in[WEST][k]          (signal_dspin_false_int_rsp_in[0][y][WEST][k]);
         clusters[0][y]->p_dspin_int_rsp_out[WEST][k]         (signal_dspin_false_int_rsp_out[0][y][WEST][k]);
         clusters[X_SIZE-1][y]->p_dspin_int_rsp_in[EAST][k]   (signal_dspin_false_int_rsp_in[X_SIZE-1][y][EAST][k]);
         clusters[X_SIZE-1][y]->p_dspin_int_rsp_out[EAST][k]  (signal_dspin_false_int_rsp_out[X_SIZE-1][y][EAST][k]);
      }

     clusters[0][y]->p_dspin_ram_cmd_in[WEST]       (signal_dspin_false_ram_cmd_in[0][y][WEST]);
     clusters[0][y]->p_dspin_ram_cmd_out[WEST]      (signal_dspin_false_ram_cmd_out[0][y][WEST]);
     clusters[0][y]->p_dspin_ram_rsp_in[WEST]       (signal_dspin_false_ram_rsp_in[0][y][WEST]);
     clusters[0][y]->p_dspin_ram_rsp_out[WEST]      (signal_dspin_false_ram_rsp_out[0][y][WEST]);

     clusters[X_SIZE-1][y]->p_dspin_ram_cmd_in[EAST]  (signal_dspin_false_ram_cmd_in[X_SIZE-1][y][EAST]);
     clusters[X_SIZE-1][y]->p_dspin_ram_cmd_out[EAST] (signal_dspin_false_ram_cmd_out[X_SIZE-1][y][EAST]);
     clusters[X_SIZE-1][y]->p_dspin_ram_rsp_in[EAST]  (signal_dspin_false_ram_rsp_in[X_SIZE-1][y][EAST]);
     clusters[X_SIZE-1][y]->p_dspin_ram_rsp_out[EAST] (signal_dspin_false_ram_rsp_out[X_SIZE-1][y][EAST]);
   }

   std::cout << "East & West boundaries established" << std::endl;

   // North & South boundary clusters connections
   for (size_t x = 0; x < X_SIZE; x++)
   {
      for (size_t k = 0; k < 3; k++)
      {
         clusters[x][0]->p_dspin_int_cmd_in[SOUTH][k]         (signal_dspin_false_int_cmd_in[x][0][SOUTH][k]);
         clusters[x][0]->p_dspin_int_cmd_out[SOUTH][k]        (signal_dspin_false_int_cmd_out[x][0][SOUTH][k]);
         clusters[x][Y_SIZE-1]->p_dspin_int_cmd_in[NORTH][k]  (signal_dspin_false_int_cmd_in[x][Y_SIZE-1][NORTH][k]);
         clusters[x][Y_SIZE-1]->p_dspin_int_cmd_out[NORTH][k] (signal_dspin_false_int_cmd_out[x][Y_SIZE-1][NORTH][k]);
      }

      for (size_t k = 0; k < 2; k++)
      {
         clusters[x][0]->p_dspin_int_rsp_in[SOUTH][k]         (signal_dspin_false_int_rsp_in[x][0][SOUTH][k]);
         clusters[x][0]->p_dspin_int_rsp_out[SOUTH][k]        (signal_dspin_false_int_rsp_out[x][0][SOUTH][k]);
         clusters[x][Y_SIZE-1]->p_dspin_int_rsp_in[NORTH][k]  (signal_dspin_false_int_rsp_in[x][Y_SIZE-1][NORTH][k]);
         clusters[x][Y_SIZE-1]->p_dspin_int_rsp_out[NORTH][k] (signal_dspin_false_int_rsp_out[x][Y_SIZE-1][NORTH][k]);
      }

      clusters[x][0]->p_dspin_ram_cmd_in[SOUTH]       (signal_dspin_false_ram_cmd_in[x][0][SOUTH]);
      clusters[x][0]->p_dspin_ram_cmd_out[SOUTH]      (signal_dspin_false_ram_cmd_out[x][0][SOUTH]);
      clusters[x][0]->p_dspin_ram_rsp_in[SOUTH]       (signal_dspin_false_ram_rsp_in[x][0][SOUTH]);
      clusters[x][0]->p_dspin_ram_rsp_out[SOUTH]      (signal_dspin_false_ram_rsp_out[x][0][SOUTH]);

      clusters[x][Y_SIZE-1]->p_dspin_ram_cmd_in[NORTH]  (signal_dspin_false_ram_cmd_in[x][Y_SIZE-1][NORTH]);
      clusters[x][Y_SIZE-1]->p_dspin_ram_cmd_out[NORTH] (signal_dspin_false_ram_cmd_out[x][Y_SIZE-1][NORTH]);
      clusters[x][Y_SIZE-1]->p_dspin_ram_rsp_in[NORTH]  (signal_dspin_false_ram_rsp_in[x][Y_SIZE-1][NORTH]);
      clusters[x][Y_SIZE-1]->p_dspin_ram_rsp_out[NORTH] (signal_dspin_false_ram_rsp_out[x][Y_SIZE-1][NORTH]);
   }

   std::cout << "North & South boundaries established" << std::endl << std::endl;

   ////////////////////////////////////////////////////////
   //   Simulation
   ///////////////////////////////////////////////////////

   sc_start(sc_core::sc_time(0, SC_NS));

   signal_resetn = false;
   signal_irq_false = false;

   // network boundaries signals
   for (size_t x = 0; x < X_SIZE ; x++)
   {
      for (size_t y = 0; y < Y_SIZE ; y++)
      {
         for (size_t a = 0; a < 4; a++)
         {
            for (size_t k = 0; k < 3; k++)
            {
               signal_dspin_false_int_cmd_in[x][y][a][k].write = false;
               signal_dspin_false_int_cmd_in[x][y][a][k].read = true;
               signal_dspin_false_int_cmd_out[x][y][a][k].write = false;
               signal_dspin_false_int_cmd_out[x][y][a][k].read = true;
            }

            for (size_t k = 0; k < 2; k++)
            {
               signal_dspin_false_int_rsp_in[x][y][a][k].write = false;
               signal_dspin_false_int_rsp_in[x][y][a][k].read = true;
               signal_dspin_false_int_rsp_out[x][y][a][k].write = false;
               signal_dspin_false_int_rsp_out[x][y][a][k].read = true;
            }

            signal_dspin_false_ram_cmd_in[x][y][a].write = false;
            signal_dspin_false_ram_cmd_in[x][y][a].read = true;
            signal_dspin_false_ram_cmd_out[x][y][a].write = false;
            signal_dspin_false_ram_cmd_out[x][y][a].read = true;

            signal_dspin_false_ram_rsp_in[x][y][a].write = false;
            signal_dspin_false_ram_rsp_in[x][y][a].read = true;
            signal_dspin_false_ram_rsp_out[x][y][a].write = false;
            signal_dspin_false_ram_rsp_out[x][y][a].read = true;
         }
      }
   }

   sc_start(sc_core::sc_time(1, SC_NS));
   signal_resetn = true;

   // simulation loop
   struct timeval t1, t2;

   // cycles between stats
   const size_t stats_period = 100000;
   const size_t simul_period = debug_ok ? debug_period : stats_period;

   for (size_t n = 0; n < ncycles; n += simul_period)
   {
      // stats display
      if((n % stats_period) == 0)
      {
         if (n > 0)
         {
            gettimeofday(&t2, NULL);

            uint64_t ms1 = (uint64_t) t1.tv_sec  * 1000ULL +
               (uint64_t) t1.tv_usec / 1000;
            uint64_t ms2 = (uint64_t) t2.tv_sec  * 1000ULL +
               (uint64_t) t2.tv_usec / 1000;
            std::cerr << "### cycle = " << n << " / frequency (Khz) = "
               << (double) stats_period / (double) (ms2 - ms1) << std::endl;
         }

         gettimeofday(&t1, NULL);
      }

      // Monitor a specific address for one L1 cache
      // clusters[1][1]->proc[0]->cache_monitor(0x50090ULL);

      // Monitor a specific address for one L2 cache
      // clusters[0][0]->memc->cache_monitor( 0x170000ULL);

      // Monitor a specific address for one XRAM
      // if (n == 3000000) clusters[0][0]->xram->start_monitor( 0x170000ULL , 64);

      if (debug_ok and (n > debug_from) and (n % debug_period == 0))
      {
         std::cout << "****************** cycle " << std::dec << n ;
         std::cout << " ************************************************" << std::endl;

         // trace proc[debug_proc_id]
         if ( debug_proc_id != 0xFFFFFFFF )
         {
            size_t l          = debug_proc_id & ((1 << P_WIDTH) - 1);
            size_t cluster_xy = debug_proc_id >> P_WIDTH ;
            size_t x          = cluster_xy >> Y_WIDTH;
            size_t y          = cluster_xy & ((1 << Y_WIDTH) - 1);

            clusters[x][y]->proc[l]->print_trace(1);
            std::ostringstream proc_signame;
            proc_signame << "[SIG]PROC_" << x << "_" << y << "_" << l ;
            clusters[x][y]->signal_int_vci_ini_proc[l].print_trace(proc_signame.str());

            clusters[x][y]->xicu->print_trace(l);
            std::ostringstream xicu_signame;
            xicu_signame << "[SIG]XICU_" << x << "_" << y;
            clusters[x][y]->signal_int_vci_tgt_xicu.print_trace(xicu_signame.str());

            //              clusters[x][y]->mdma->print_trace();
            //              std::ostringstream mdma_signame;
            //              mdma_signame << "[SIG]MDMA_" << x << "_" << y;
            //              clusters[x][y]->signal_int_vci_tgt_mdma.print_trace(mdma_signame.str());

            if( clusters[x][y]->signal_proc_it[l].read() )
               std::cout << "### IRQ_PROC_" << std::dec
                  << x << "_" << y << "_" << l << " ACTIVE" << std::endl;
         }

         // trace memc[debug_memc_id]
         if ( debug_memc_id != 0xFFFFFFFF )
         {
            size_t x = debug_memc_id >> Y_WIDTH;
            size_t y = debug_memc_id & ((1 << Y_WIDTH) - 1);

            clusters[x][y]->memc->print_trace(0);
            std::ostringstream smemc_tgt;
            smemc_tgt << "[SIG]MEMC_TGT_" << x << "_" << y;
            clusters[x][y]->signal_int_vci_tgt_memc.print_trace(smemc_tgt.str());
            std::ostringstream smemc_ini;
            smemc_ini << "[SIG]MEMC_INI_" << x << "_" << y;
            clusters[x][y]->signal_ram_vci_ini_memc.print_trace(smemc_ini.str());

            clusters[x][y]->xram->print_trace();
            std::ostringstream sxram_tgt;
            sxram_tgt << "[SIG]XRAM_TGT_" << x << "_" << y;
            clusters[x][y]->signal_ram_vci_tgt_xram.print_trace(sxram_tgt.str());
         }


         // trace XRAM and XRAM network routers in cluster[debug_xram_id]
         if ( debug_xram_id != 0xFFFFFFFF )
         {
            size_t x = debug_xram_id >> Y_WIDTH;
            size_t y = debug_xram_id & ((1 << Y_WIDTH) - 1);

            clusters[x][y]->xram->print_trace();
            std::ostringstream sxram_tgt;
            sxram_tgt << "[SIG]XRAM_TGT_" << x << "_" << y;
            clusters[x][y]->signal_ram_vci_tgt_xram.print_trace(sxram_tgt.str());

            clusters[x][y]->ram_router_cmd->print_trace();
            clusters[x][y]->ram_router_rsp->print_trace();
         }

         // trace iob, iox and external peripherals
         if ( debug_iob )
         {
            clusters[0][0]->iob->print_trace();
            clusters[X_SIZE-1][Y_SIZE-1]->iob->print_trace();
            //              clusters[0][0]->signal_int_vci_tgt_iobx.print_trace( "[SIG]IOB0_INT_TGT");
            //              clusters[0][0]->signal_int_vci_ini_iobx.print_trace( "[SIG]IOB0_INT_INI");
            //              clusters[0][0]->signal_ram_vci_ini_iobx.print_trace( "[SIG]IOB0_RAM_INI");

            signal_vci_ini_iob0.print_trace("[SIG]IOB0_IOX_INI");
            signal_vci_tgt_iob0.print_trace("[SIG]IOB0_IOX_TGT");

            //              cdma->print_trace();
            //              signal_vci_tgt_cdma.print_trace("[SIG]IOX_CDMA_TGT");
            //              signal_vci_ini_cdma.print_trace("[SIG]IOX_CDMA_INI");

            //              mtty->print_trace();
            //              signal_vci_tgt_mtty.print_trace("[SIG]IOX_MTTY_TGT");

            bdev->print_trace();
            signal_vci_tgt_bdev.print_trace("[SIG]BDEV_TGT");
            signal_vci_ini_bdev.print_trace("[SIG]BDEV_INI");

            mnic->print_trace();
            signal_vci_tgt_mnic.print_trace("[SIG]MNIC_TGT");

            //              fbuf->print_trace();
            //              signal_vci_tgt_fbuf.print_trace("[SIG]FBUF");

            iopi->print_trace();
            signal_vci_ini_iopi.print_trace("[SIG]IOPI_INI");
            signal_vci_tgt_iopi.print_trace("[SIG]IOPI_TGT");

            signal_vci_tgt_simh.print_trace("[SIG]SIMH_TGT");

            iox_network->print_trace();

            // interrupts
            if (signal_irq_bdev)       std::cout << "### IRQ_BDEV ACTIVE"       << std::endl;
            if (signal_irq_mtty_rx[0]) std::cout << "### IRQ_MTTY ACTIVE"       << std::endl;
            if (signal_irq_mnic_rx[0]) std::cout << "### IRQ_MNIC_RX[0] ACTIVE" << std::endl;
            if (signal_irq_mnic_rx[1]) std::cout << "### IRQ_MNIC_RX[1] ACTIVE" << std::endl;
            if (signal_irq_mnic_tx[0]) std::cout << "### IRQ_MNIC_TX[0] ACTIVE" << std::endl;
            if (signal_irq_mnic_tx[1]) std::cout << "### IRQ_MNIC_TX[1] ACTIVE" << std::endl;
         }
      }

      sc_start(sc_core::sc_time(simul_period, SC_NS));
   }
   return EXIT_SUCCESS;
}

int sc_main (int argc, char *argv[])
{
   try {
      return _main(argc, argv);
   } catch (soclib::exception::RunTimeError &e) {
      std::cout << "RunTimeError: " << e.what() << std::endl;
   } catch (std::exception &e) {
      std::cout << e.what() << std::endl;
   } catch (...) {
      std::cout << "Unknown exception occured" << std::endl;
      throw;
   }
   return 1;
}


// Local Variables:
// tab-width: 3
// c-basic-offset: 3
// c-file-offsets:((innamespace . 0)(inline-open . 0))
// indent-tabs-mode: nil
// End:

// vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3