Index: /trunk/platforms/tsar_generic_leti/top.cpp =================================================================== --- /trunk/platforms/tsar_generic_leti/top.cpp (revision 791) +++ /trunk/platforms/tsar_generic_leti/top.cpp (revision 792) @@ -1,5 +1,5 @@ ///////////////////////////////////////////////////////////////////////// // File: top.cpp (for tsar_generic_leti) -// Author: Alain Greiner +// Author: Alain Greiner // Copyright: UPMC/LIP6 // Date : february 2014 @@ -9,8 +9,8 @@ // with the VLSI Hardware prototype developped by CEA-LETI and LIP6 // in the framework of the SHARP project. -// +// // The processor is a MIPS32 processor wrapped in a GDB server // (this is defined in the tsar_xbar_cluster). -// +// // It does not use an external ROM, as the boot code is (pre)loaded // in cluster (0,0) memory at address 0x0. @@ -23,5 +23,5 @@ // The number of clusters cannot be larger than 128. // The number of processors per cluster cannot be larger than 4. -// +// // Each cluster contains: // - 5 dspin_local_crossbar (local interconnect) @@ -43,19 +43,19 @@ // that must be considered as an extension of this top.cpp file. // -// Besides the hardware components in clusters, "external" peripherals +// Besides the hardware components in clusters, "external" peripherals // are connected to an external IO bus (implemented as a vci_local_crossbar): -// - one disk controller -// - one multi-channel ethernet controller +// - one disk controller +// - one multi-channel ethernet controller // - one multi-channel chained buffer dma controller -// - one multi-channel tty controller -// - one frame buffer controller +// - one multi-channel tty controller +// - one frame buffer controller // - one 32 channels iopic controller // -// This IOBUS is connected to the north port of the DIR_CMD +// This IOBUS is connected to the north port of the DIR_CMD // and DIR_RSP routers, in cluster(X_SIZE-1, Y_SIZE-1). // For all external peripherals, the hardware interrupts (HWI) are // translated to write interrupts (WTI) by the iopic component: // - IOPIC HWI[1:0] connected to IRQ_NIC_RX[1:0] -// - IOPIC HWI[3:2] connected to IRQ_NIC_TX[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 @@ -73,8 +73,8 @@ // - NB_TTY_CHANNELS : number of TTY channels in I/O cluster (8 max) // - NB_NIC_CHANNELS : number of NIC channels in I/O cluster (2 max) -// +// // Some other hardware parameters are not used when compiling the OS, // and are only defined in this top.cpp file: -// - XRAM_LATENCY : external ram latency +// - XRAM_LATENCY : external ram latency // - MEMC_WAYS : L2 cache number of ways // - MEMC_SETS : L2 cache number of sets @@ -85,5 +85,5 @@ // - FBUF_X_SIZE : width of frame buffer (pixels) // - FBUF_Y_SIZE : heigth of frame buffer (lines) -// - BDEV_IMAGE_NAME : file pathname for block device +// - 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 @@ -93,5 +93,5 @@ // 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 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: @@ -127,36 +127,5 @@ #include "alloc_elems.h" -////////////////////////////////////////////////////////////////////////////////////////// -// Parameters depending on the OS and software application -// - path to hard_config file -// - path to binary code for the RAM loader -// - path to disk image for RAMDISK loader -// - path to disk image for IOC device -////////////////////////////////////////////////////////////////////////////////////////// - -#define USE_GIET_VM 1 -#define USE_GIET_TSAR 0 - -#if ( USE_GIET_VM and USE_GIET_TSAR ) -#error "Can't use Two different OS" -#endif - -#if ( (not USE_GIET_VM) and (not USE_GIET_TSAR) ) -#error "You need to specify one OS" -#endif - -#if USE_GIET_TSAR -#include "hard_config.h" -#define BINARY_PATH_FOR_LOADER "../../softs/soft_transpose_giet/bin.soft" -#define DISK_IMAGE_PATH_FOR_IOC "../../softs/soft_transpose_giet/images.raw" -#define RAMDISK_PATH_FOR_LOADER DISK_IMAGE_PATH_FOR_IOC "@0x00800000:" -#endif - -#if USE_GIET_VM -#include "hard_config.h" -#define BINARY_PATH_FOR_LOADER "../../softs/tsar_boot/preloader.elf" -#define DISK_IMAGE_PATH_FOR_IOC "../../../giet_vm/hdd/virt_hdd.dmg" -#define RAMDISK_PATH_FOR_LOADER DISK_IMAGE_PATH_FOR_IOC "@0x02000000:" -#endif +#include "hard_config.h" /////////////////////////////////////////////////// @@ -176,5 +145,5 @@ /////////////////////////////////////////////////////////// -// DSPIN parameters +// DSPIN parameters /////////////////////////////////////////////////////////// @@ -183,5 +152,5 @@ /////////////////////////////////////////////////////////// -// VCI parameters +// VCI parameters /////////////////////////////////////////////////////////// @@ -200,8 +169,6 @@ ///////////////////////////////////////////////////////////////////////////////////////// -// Secondary Hardware Parameters +// Secondary Hardware Parameters ///////////////////////////////////////////////////////////////////////////////////////// - -#define RESET_ADDRESS 0x0 #define MAX_TTY_CHANNELS 8 @@ -210,5 +177,4 @@ #define XRAM_LATENCY 0 -#define XRAM_SIZE 0x04000000 // 64 Mbytes per cluster #define MEMC_WAYS 16 @@ -220,7 +186,4 @@ #define L1_DWAYS 4 #define L1_DSETS 64 - -#define FBUF_X_SIZE 128 -#define FBUF_Y_SIZE 128 #define NIC_MAC4 0XBABEF00D @@ -235,5 +198,5 @@ /////////////////////////////////////////////////////////////////////////////////////// -// DEBUG Parameters default values +// DEBUG Parameters default values /////////////////////////////////////////////////////////////////////////////////////// @@ -241,5 +204,5 @@ /////////////////////////////////////////////////////////////////////////////////////// -// LOCAL TGTID & SRCID definition +// LOCAL TGTID & SRCID definition // For all components: global TGTID = global SRCID = cluster_index /////////////////////////////////////////////////////////////////////////////////////// @@ -258,48 +221,4 @@ #define IOPI_SRCID NB_PROCS_MAX + 2 -////////////////////////////////////////////////////////////////////////////////////// -// Physical segments definition -////////////////////////////////////////////////////////////////////////////////////// -// - 3 segments are replicated in all clusters -// - 2 segments are only in cluster[0,0] -// - 4 segments are only in cluster [X_SIZE-1,Y_SIZE] -// The following values are for segments in cluster 0, -// and these 32 bits values must be concatenate with the cluster -// index (on 8 bits) to obtain the 40 bits address. -////////////////////////////////////////////////////////////////////////////////////// - -// in cluster [0,0] & [X_SIZE-1,Y_SIZE] - -#define MTTY_BASE 0xF4000000 -#define MTTY_SIZE 0x00001000 // 4 Kbytes - -#define BDEV_BASE 0xF2000000 -#define BDEV_SIZE 0x00001000 // 4 Kbytes - -// in cluster [X_SIZE-1,Y_SIZE] - -#define FBUF_BASE 0xF3000000 -#define FBUF_SIZE (FBUF_X_SIZE * FBUF_Y_SIZE * 2) - -#define MNIC_BASE 0xF7000000 -#define MNIC_SIZE 0x00800000 // 512 Kbytes (for 8 channels) - -#define CDMA_BASE 0xF8000000 -#define CDMA_SIZE 0x00004000 * NB_CMA_CHANNELS - -#define IOPI_BASE 0xF9000000 -#define IOPI_SIZE 0x00001000 // 4 Kbytes - -// replicated segments : address is extended to 40 bits by cluster_xy - -#define MEMC_BASE 0x00000000 -#define MEMC_SIZE XRAM_SIZE - -#define MCFG_BASE 0xE0000000 -#define MCFG_SIZE 0x00001000 // 4 Kbytes - -#define XICU_BASE 0xF0000000 -#define XICU_SIZE 0x00001000 // 4 Kbytes - bool stop_called = false; @@ -317,10 +236,10 @@ bool trace_proc_ok = false; // detailed proc trace activated size_t trace_memc_ok = false; // detailed memc trace activated - size_t trace_memc_id = 0; // index of memc to be traced + size_t trace_memc_id = 0; // index of memc to be traced size_t trace_proc_id = 0; // index of proc to be traced uint32_t frozen_cycles = MAX_FROZEN_CYCLES; - char soft_name[256] = BINARY_PATH_FOR_LOADER; - char disk_name[256] = DISK_IMAGE_PATH_FOR_IOC; - char ramdisk_name[256] = RAMDISK_PATH_FOR_LOADER; + char soft_name[256] = "soft.elf"; + char disk_name[256] = "disk.img"; + char ramdisk_name[256] = "disk.img@0x02000000:"; struct timeval t1,t2; uint64_t ms1,ms2; @@ -364,5 +283,5 @@ else if ((strcmp(argv[n], "-SOFT") == 0) && ((n + 1) < argc)) { - strcpy(soft_name, argv[n + 1]); + strcpy(soft_name, argv[n + 1]); } else if ((strcmp(argv[n], "-DISK") == 0) && ((n + 1) < argc)) @@ -372,5 +291,5 @@ else if ((strcmp(argv[n], "-RAMDISK") == 0) && ((n + 1) < argc)) { - strcpy(ramdisk_name, argv[n + 1]); + strcpy(ramdisk_name, argv[n + 1]); } else if ((strcmp(argv[n], "-THREADS") == 0) && ((n + 1) < argc)) @@ -404,8 +323,9 @@ // checking hardware parameters - assert( ((X_SIZE==1) or (X_SIZE==2) or (X_SIZE==4) or (X_SIZE==8) or (X_SIZE==16)) and + assert( ((X_SIZE==1) or (X_SIZE==2) or (X_SIZE==4) or (X_SIZE==8) or + (X_SIZE==16)) and "Illegal X_SIZE parameter" ); - assert( ((Y_SIZE==1) or (Y_SIZE==2) or (Y_SIZE==4) or (Y_SIZE==8)) and + assert( ((Y_SIZE==1) or (Y_SIZE==2) or (Y_SIZE==4) or (Y_SIZE==8)) and "Illegal Y_SIZE parameter" ); @@ -427,5 +347,5 @@ assert( (X_WIDTH == 4) and (Y_WIDTH == 4) and "ERROR: you must have X_WIDTH == Y_WIDTH == 4"); - + std::cout << std::endl; @@ -483,7 +403,7 @@ /////////////////////////////////////// - MappingTable maptabd(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), + MappingTable maptabd(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), 0x00FF000000ULL); @@ -498,15 +418,15 @@ std::ostringstream si; si << "seg_xicu_" << x << "_" << y; - maptabd.add(Segment(si.str(), XICU_BASE + offset, XICU_SIZE, + maptabd.add(Segment(si.str(), SEG_XCU_BASE + offset, SEG_XCU_SIZE, IntTab(cluster(x,y),XICU_TGTID), false)); std::ostringstream sd; sd << "seg_mcfg_" << x << "_" << y; - maptabd.add(Segment(sd.str(), MCFG_BASE + offset, MCFG_SIZE, + maptabd.add(Segment(sd.str(), SEG_MMC_BASE + offset, SEG_MMC_SIZE, IntTab(cluster(x,y),MEMC_TGTID), false)); std::ostringstream sh; sh << "seg_memc_" << x << "_" << y; - maptabd.add(Segment(sh.str(), MEMC_BASE + offset, MEMC_SIZE, + maptabd.add(Segment(sh.str(), SEG_RAM_BASE + offset, SEG_RAM_SIZE, IntTab(cluster(x,y),MEMC_TGTID), true)); } @@ -514,8 +434,8 @@ // segments for peripherals in cluster(0,0) - maptabd.add(Segment("seg_tty0", MTTY_BASE, MTTY_SIZE, + maptabd.add(Segment("seg_tty0", SEG_TTY_BASE, SEG_TTY_SIZE, IntTab(cluster(0,0),MTTY_TGTID), false)); - maptabd.add(Segment("seg_ioc0", BDEV_BASE, BDEV_SIZE, + maptabd.add(Segment("seg_ioc0", SEG_IOC_BASE, SEG_IOC_SIZE, IntTab(cluster(0,0),BDEV_TGTID), false)); @@ -524,20 +444,20 @@ offset = ((sc_uint)cluster(X_SIZE-1,Y_SIZE)) << 32; - maptabd.add(Segment("seg_mtty", MTTY_BASE + offset, MTTY_SIZE, + maptabd.add(Segment("seg_mtty", SEG_TTY_BASE + offset, SEG_TTY_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),MTTY_TGTID), false)); - maptabd.add(Segment("seg_fbuf", FBUF_BASE + offset, FBUF_SIZE, + maptabd.add(Segment("seg_fbuf", SEG_FBF_BASE + offset, SEG_FBF_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),FBUF_TGTID), false)); - maptabd.add(Segment("seg_bdev", BDEV_BASE + offset, BDEV_SIZE, + maptabd.add(Segment("seg_bdev", SEG_IOC_BASE + offset, SEG_IOC_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),BDEV_TGTID), false)); - maptabd.add(Segment("seg_mnic", MNIC_BASE + offset, MNIC_SIZE, + maptabd.add(Segment("seg_mnic", SEG_NIC_BASE + offset, SEG_NIC_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),MNIC_TGTID), false)); - maptabd.add(Segment("seg_cdma", CDMA_BASE + offset, CDMA_SIZE, + maptabd.add(Segment("seg_cdma", SEG_CMA_BASE + offset, SEG_CMA_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),CDMA_TGTID), false)); - maptabd.add(Segment("seg_iopi", IOPI_BASE + offset, IOPI_SIZE, + maptabd.add(Segment("seg_iopi", SEG_PIC_BASE + offset, SEG_PIC_SIZE, IntTab(cluster(X_SIZE-1, Y_SIZE),IOPI_TGTID), false)); @@ -548,7 +468,7 @@ ///////////////////////////////////////////////// - MappingTable maptabx(vci_address_width, - IntTab(X_WIDTH+Y_WIDTH), - IntTab(X_WIDTH+Y_WIDTH), + MappingTable maptabx(vci_address_width, + IntTab(X_WIDTH+Y_WIDTH), + IntTab(X_WIDTH+Y_WIDTH), 0x00FF000000ULL); @@ -556,7 +476,7 @@ { for (size_t y = 0; y < (Y_SIZE) ; y++) - { + { sc_uint offset; - offset = (sc_uint)cluster(x,y) + offset = (sc_uint)cluster(x,y) << (vci_address_width-X_WIDTH-Y_WIDTH); @@ -564,6 +484,6 @@ sh << "x_seg_memc_" << x << "_" << y; - maptabx.add(Segment(sh.str(), MEMC_BASE + offset, - MEMC_SIZE, IntTab(cluster(x,y)), false)); + maptabx.add(Segment(sh.str(), SEG_RAM_BASE + offset, + SEG_RAM_SIZE, IntTab(cluster(x,y)), false)); } } @@ -669,5 +589,5 @@ VciSignals signal_vci_ini_iopi("signal_vci_ini_iopi"); - VciSignals* signal_vci_ini_proc = + VciSignals* signal_vci_ini_proc = alloc_elems >("signal_vci_ini_proc", NB_PROCS_MAX ); @@ -683,10 +603,10 @@ VciSignals signal_vci_cmd_to_noc("signal_vci_cmd_to_noc"); VciSignals signal_vci_cmd_from_noc("signal_vci_cmd_from_noc"); - + //////////////////////////// - // Loader + // Loader //////////////////////////// -#if USE_IOC_RDK +#if USE_RAMDISK soclib::common::Loader loader( soft_name, ramdisk_name ); #else @@ -726,5 +646,5 @@ #endif std::cout << std::endl; - std::cout << "Cluster_" << std::dec << x << "_" << y + std::cout << "Cluster_" << std::dec << x << "_" << y << " with cluster_xy = " << std::hex << cluster(x,y) << std::endl; std::cout << std::endl; @@ -764,7 +684,7 @@ frozen_cycles, trace_from, - trace_proc_ok, + trace_proc_ok, trace_proc_id, - trace_memc_ok, + trace_memc_ok, trace_memc_id ); @@ -785,6 +705,6 @@ // but the crossbar has (NB_PROCS_MAX + 3) intiators and // 8 targets, in order to use the same SRCID and TGTID space - // (same mapping table for the internal components, - // and for the external peripherals) + // (same mapping table for the internal components, + // and for the external peripherals) ////////////////////////////////////////////////////////////////// @@ -792,8 +712,8 @@ std::cout << " Building IO cluster (external peripherals)" << std::endl; std::cout << std::endl; - + size_t cluster_io = cluster(X_SIZE-1, Y_SIZE); - //////////// vci_local_crossbar + //////////// vci_local_crossbar VciLocalCrossbar* iobus = new VciLocalCrossbar( @@ -805,5 +725,5 @@ BDEV_TGTID ); // default target index - //////////// vci_framebuffer + //////////// vci_framebuffer VciFrameBuffer* fbuf = new VciFrameBuffer( @@ -813,5 +733,5 @@ FBUF_X_SIZE, FBUF_Y_SIZE ); - //////////// vci_block_device + //////////// vci_block_device VciBlockDeviceTsar* bdev = new VciBlockDeviceTsar( @@ -824,5 +744,5 @@ 64 ); // burst size - //////////// vci_multi_nic + //////////// vci_multi_nic VciMultiNic* mnic = new VciMultiNic( @@ -836,5 +756,5 @@ NIC_TX_NAME ); - ///////////// vci_chbuf_dma + ///////////// vci_chbuf_dma VciChbufDma* cdma = new VciChbufDma( @@ -844,5 +764,5 @@ IntTab(cluster_io, CDMA_TGTID), 64, // burst size - NB_CMA_CHANNELS ); + NB_CMA_CHANNELS ); ////////////// vci_multi_tty @@ -855,5 +775,5 @@ } - VciMultiTty* + VciMultiTty* mtty = new VciMultiTty( "mtty", @@ -869,8 +789,7 @@ IntTab(cluster_io, IOPI_SRCID), IntTab(cluster_io, IOPI_TGTID), - 32, - 5000 ); - - ////////////// vci_dspin wrappers + 32 ); + + ////////////// vci_dspin wrappers VciDspinTargetWrapper* wt_iobus = new VciDspinTargetWrapper( @@ -884,9 +803,9 @@ /////////////////////////////////////////////////////////////// - // Net-list + // Net-list /////////////////////////////////////////////////////////////// - // iobus - iobus->p_clk (signal_clk); + // iobus + iobus->p_clk (signal_clk); iobus->p_resetn (signal_resetn); @@ -1049,8 +968,8 @@ } } - std::cout << std::endl << "Horizontal connections done" << std::endl; + std::cout << std::endl << "Horizontal connections done" << std::endl; // Inter Clusters vertical connections - if (Y_SIZE > 1) + if (Y_SIZE > 1) { for (size_t y = 0; y < (Y_SIZE-1); y++) @@ -1198,5 +1117,5 @@ } - // set default values for VCI signals connected to unused ports on iobus + // set default values for VCI signals connected to unused ports on iobus signal_vci_tgt_memc.rspval = false; signal_vci_tgt_xicu.rspval = false; @@ -1206,5 +1125,5 @@ signal_resetn = true; - if (gettimeofday(&t1, NULL) != 0) + if (gettimeofday(&t1, NULL) != 0) { perror("gettimeofday"); @@ -1233,5 +1152,5 @@ { - if (gettimeofday(&t2, NULL) != 0) + if (gettimeofday(&t2, NULL) != 0) { perror("gettimeofday"); @@ -1241,8 +1160,8 @@ ms1 = (uint64_t) t1.tv_sec * 1000ULL + (uint64_t) t1.tv_usec / 1000; ms2 = (uint64_t) t2.tv_sec * 1000ULL + (uint64_t) t2.tv_usec / 1000; - std::cerr << "platform clock frequency " + std::cerr << "platform clock frequency " << (double) 5000000 / (double) (ms2 - ms1) << "Khz" << std::endl; - if (gettimeofday(&t1, NULL) != 0) + if (gettimeofday(&t1, NULL) != 0) { perror("gettimeofday"); @@ -1291,5 +1210,5 @@ clusters[x][y]->signal_vci_tgt_memc.print_trace(smemc.str()); clusters[x][y]->signal_vci_xram.print_trace(sxram.str()); - } + } // trace coherence signals @@ -1309,5 +1228,5 @@ // clusters[0][1]->xbar_m2p->print_trace(); // clusters[1][1]->xbar_m2p->print_trace(); - + // trace router(s) m2p // clusters[0][0]->router_m2p->print_trace(); @@ -1315,5 +1234,5 @@ // clusters[0][1]->router_m2p->print_trace(); // clusters[1][1]->router_m2p->print_trace(); - + // trace external ioc bdev->print_trace(); @@ -1335,5 +1254,5 @@ { // trace BDV interrupts events - if ( signal_irq_bdev.read() != prev_irq_bdev ) + if ( signal_irq_bdev.read() != prev_irq_bdev ) { prev_irq_bdev = signal_irq_bdev.read(); @@ -1341,12 +1260,12 @@ << " at cycle " << n << std::endl; } - + // trace TTY interrupts events for ( size_t x = 0 ; x < 8 ; x++ ) { - if ( signal_irq_mtty_rx[x].read() != prev_irq_mtty_rx[x] ) + if ( signal_irq_mtty_rx[x].read() != prev_irq_mtty_rx[x] ) { prev_irq_mtty_rx[x] = signal_irq_mtty_rx[x].read(); - std::cout << std::dec << "@@@ IRQ_MTTY["<signal_proc_irq[i] != prev_irq_proc[x][y][i] ) - { + { prev_irq_proc[x][y][i] = clusters[x][y]->signal_proc_irq[i]; std::cout << std::dec << "@@@ IRQ_PROC["<