source: trunk/platforms/tsar_generic_leti/top.cpp@ 943

Last change on this file since 943 was 939, checked in by alain, 12 years ago

Introduce the distributed kernel heap vsegs.

File size: 54.2 KB
RevLine 
[621]1/////////////////////////////////////////////////////////////////////////
[664]2// File: top.cpp (for tsar_generic_leti)
[792]3// Author: Alain Greiner
[621]4// Copyright: UPMC/LIP6
[937]5// Date : february 2013 / updated january 2015
[621]6// This program is released under the GNU public license
7/////////////////////////////////////////////////////////////////////////
[681]8// This file define a generic TSAR architecture, fully compatible
9// with the VLSI Hardware prototype developped by CEA-LETI and LIP6
10// in the framework of the SHARP project.
[792]11//
[621]12// The processor is a MIPS32 processor wrapped in a GDB server
13// (this is defined in the tsar_xbar_cluster).
[792]14//
[937]15// The main hardware parameters are the mesh size (X_SIZE & Y_SIZE),
16// and the number of processors per cluster (NB_PROCS_MAX).
17// The NB_PROCS_MAX parameter cannot be larger than 4.
[621]18//
[937]19// All external peripherals are located in cluster[X_SIZE-1][Y_SIZE-1],
20// and are connected to an IO bus (implemented as a vci_local_crossbar):
21// - one disk controller
22// - one multi-channel ethernet controller
23// - one multi-channel chained buffer dma controller
24// - one multi-channel tty controller
25// - one frame buffer controller
26// - one iopic controller
27// This IO bus is directly connected to the north ports of the CMD/RSP
28// routers in cluster[X_SIZE-1][y_SIZE-2] through VCI/DSPIN wrappers.
29// All other clusters in the upper row are empty: no processors,
30// no ram, no routers.
31// The X_SIZE parameter must be larger than 0, but no larger than 16.
32// The Y_SIZE parameter must be larger than 1, but no larger than 16.
33//
34// We don't use an external ROM, as the boot code is (pre)loaded
35// in RAM in cluster[0][0] at address 0x0.
36//
37// An optional RAMDISK of 32 Mbytes can be used in RAM of cluster[0][0].
38//
[621]39// The physical address space is 40 bits.
40// The 8 address MSB bits define the cluster index.
41//
[937]42// Besides the processors, each cluster contains:
43// - 5 L1/L2 DSPIN routers implementing 5 separated NOCs
[628]44// - 1 vci_mem_cache
45// - 1 vci_xicu
[937]46// - 1 vci_simple_ram (to emulate the L3 cache).
[621]47//
[681]48// Each processor receives 4 consecutive IRQ lines from the local XICU.
[664]49//
[628]50// In all clusters, the MEMC IRQ line (signaling a late write error)
51// is connected to XICU HWI[8]
[621]52//
[792]53// This IOBUS is connected to the north port of the DIR_CMD
[681]54// and DIR_RSP routers, in cluster(X_SIZE-1, Y_SIZE-1).
[628]55// For all external peripherals, the hardware interrupts (HWI) are
[664]56// translated to write interrupts (WTI) by the iopic component:
57// - IOPIC HWI[1:0] connected to IRQ_NIC_RX[1:0]
[792]58// - IOPIC HWI[3:2] connected to IRQ_NIC_TX[1:0]
[664]59// - IOPIC HWI[7:4] connected to IRQ_CMA_TX[3:0]]
60// - IOPIC HWI[8] connected to IRQ_BDEV
61// - IOPIC HWI[15:9] unused (grounded)
62// - IOPIC HWI[23:16] connected to IRQ_TTY_RX[7:0]]
63// - IOPIC HWI[31:24] connected to IRQ_TTY_TX[7:0]]
[937]64//
65// The cluster internal architecture is defined in file tsar_leti_cluster,
66// that must be considered as an extension of this top.cpp file.
[628]67////////////////////////////////////////////////////////////////////////////
68// The following parameters must be defined in the hard_config.h file :
[621]69// - X_WIDTH : number of bits for x coordinate (must be 4)
70// - Y_WIDTH : number of bits for y coordinate (must be 4)
[681]71// - X_SIZE : number of clusters in a row (1,2,4,8,16)
72// - Y_SIZE : number of clusters in a column (1,2,4,8)
[628]73// - NB_PROCS_MAX : number of processors per cluster (1, 2 or 4)
[664]74// - NB_CMA_CHANNELS : number of CMA channels in I/0 cluster (4 max)
75// - NB_TTY_CHANNELS : number of TTY channels in I/O cluster (8 max)
[628]76// - NB_NIC_CHANNELS : number of NIC channels in I/O cluster (2 max)
[937]77// - FBUF_X_SIZE : number of pixels per line for frame buffer
78// - FBUF_Y_SIZE : number of lines for frame buffer
[792]79//
[621]80// Some other hardware parameters are not used when compiling the OS,
[628]81// and are only defined in this top.cpp file:
[792]82// - XRAM_LATENCY : external ram latency
[628]83// - L1_IWAYS : L1 cache instruction number of ways
84// - L1_ISETS : L1 cache instruction number of sets
85// - L1_DWAYS : L1 cache data number of ways
86// - L1_DSETS : L1 cache data number of sets
[937]87// - BDEV_IMAGE_NAME : pathname for block device disk image
[621]88/////////////////////////////////////////////////////////////////////////
89// General policy for 40 bits physical address decoding:
90// All physical segments base addresses are multiple of 1 Mbytes
[792]91// (=> the 24 LSB bits = 0, and the 16 MSB bits define the target)
[621]92// The (X_WIDTH + Y_WIDTH) MSB bits (left aligned) define
93// the cluster index, and the LADR bits define the local index:
94// |X_ID|Y_ID| LADR | OFFSET |
95// | 4 | 4 | 8 | 24 |
96/////////////////////////////////////////////////////////////////////////
97// General policy for 14 bits SRCID decoding:
98// Each component is identified by (x_id, y_id, l_id) tuple.
99// |X_ID|Y_ID| L_ID |
100// | 4 | 4 | 6 |
101/////////////////////////////////////////////////////////////////////////
102
103#include <systemc>
104#include <sys/time.h>
105#include <iostream>
106#include <sstream>
107#include <cstdlib>
108#include <cstdarg>
109#include <stdint.h>
110
111#include "gdbserver.h"
112#include "mapping_table.h"
113#include "tsar_leti_cluster.h"
[628]114#include "vci_local_crossbar.h"
115#include "vci_dspin_initiator_wrapper.h"
116#include "vci_dspin_target_wrapper.h"
117#include "vci_multi_tty.h"
118#include "vci_multi_nic.h"
119#include "vci_chbuf_dma.h"
120#include "vci_block_device_tsar.h"
121#include "vci_framebuffer.h"
122#include "vci_iopic.h"
[621]123#include "alloc_elems.h"
124
[792]125#include "hard_config.h"
[621]126
127///////////////////////////////////////////////////
128// Parallelisation
129///////////////////////////////////////////////////
[708]130#define USE_OPENMP _OPENMP
[621]131
132#if USE_OPENMP
133#include <omp.h>
134#endif
135
136///////////////////////////////////////////////////
137// cluster index (from x,y coordinates)
138///////////////////////////////////////////////////
139
[692]140#define cluster(x,y) ((y) + ((x) << Y_WIDTH))
[621]141
142///////////////////////////////////////////////////////////
[792]143// DSPIN parameters
[621]144///////////////////////////////////////////////////////////
145
146#define dspin_cmd_width 39
147#define dspin_rsp_width 32
148
149///////////////////////////////////////////////////////////
[792]150// VCI parameters
[621]151///////////////////////////////////////////////////////////
152
153#define vci_cell_width_int 4
154#define vci_cell_width_ext 8
155#define vci_address_width 40
156#define vci_plen_width 8
157#define vci_rerror_width 1
158#define vci_clen_width 1
159#define vci_rflag_width 1
160#define vci_srcid_width 14
161#define vci_pktid_width 4
162#define vci_trdid_width 4
163#define vci_wrplen_width 1
164
165
[937]166///////////////////////////////////////////////////////////////////////////////////////
[792]167// Secondary Hardware Parameters
[937]168///////////////////////////////////////////////////////////////////////////////////////
[621]169
[937]170#define XMAX X_SIZE // actual number of columns in 2D mesh
171#define YMAX (Y_SIZE - 1) // actual number of rows in 2D mesh
[664]172
[621]173#define XRAM_LATENCY 0
174
175#define MEMC_WAYS 16
176#define MEMC_SETS 256
177
178#define L1_IWAYS 4
179#define L1_ISETS 64
180
181#define L1_DWAYS 4
182#define L1_DSETS 64
183
[937]184#define BDEV_IMAGE_NAME "../../../giet_vm/hdd/virt_hdd.dmg"
[621]185
[937]186#define ROM_SOFT_NAME "../../softs/tsar_boot/preloader.elf"
187
[621]188#define NORTH 0
189#define SOUTH 1
190#define EAST 2
191#define WEST 3
192
[664]193///////////////////////////////////////////////////////////////////////////////////////
[792]194// DEBUG Parameters default values
[664]195///////////////////////////////////////////////////////////////////////////////////////
[621]196
[681]197#define MAX_FROZEN_CYCLES 500000
[621]198
[664]199///////////////////////////////////////////////////////////////////////////////////////
[792]200// LOCAL TGTID & SRCID definition
[621]201// For all components: global TGTID = global SRCID = cluster_index
[664]202///////////////////////////////////////////////////////////////////////////////////////
[621]203
[664]204#define MEMC_TGTID 0
205#define XICU_TGTID 1
206#define MTTY_TGTID 2
207#define BDEV_TGTID 3
208#define FBUF_TGTID 4
209#define MNIC_TGTID 5
210#define CDMA_TGTID 6
211#define IOPI_TGTID 7
[621]212
[664]213#define BDEV_SRCID NB_PROCS_MAX
214#define CDMA_SRCID NB_PROCS_MAX + 1
215#define IOPI_SRCID NB_PROCS_MAX + 2
[628]216
[621]217bool stop_called = false;
218
219/////////////////////////////////
220int _main(int argc, char *argv[])
221{
222 using namespace sc_core;
223 using namespace soclib::caba;
224 using namespace soclib::common;
225
[937]226 uint32_t ncycles = 0xFFFFFFFF; // max simulated cycles
227 size_t threads = 1; // simulator's threads number
228 bool trace_ok = false; // trace activated
229 uint32_t trace_from = 0; // trace start cycle
230 bool trace_proc_ok = false; // detailed proc trace activated
231 size_t trace_memc_ok = false; // detailed memc trace activated
232 size_t trace_memc_id = 0; // index of memc to be traced
233 size_t trace_proc_id = 0; // index of proc to be traced
234 char soft_name[256] = ROM_SOFT_NAME; // pathname for ROM binary code
235 char disk_name[256] = BDEV_IMAGE_NAME; // pathname for DISK image
236 uint32_t frozen_cycles = MAX_FROZEN_CYCLES; // for debug
[621]237 struct timeval t1,t2;
238 uint64_t ms1,ms2;
239
240 ////////////// command line arguments //////////////////////
241 if (argc > 1)
242 {
243 for (int n = 1; n < argc; n = n + 2)
244 {
245 if ((strcmp(argv[n], "-NCYCLES") == 0) && (n + 1 < argc))
246 {
247 ncycles = (uint64_t) strtol(argv[n + 1], NULL, 0);
248 }
249 else if ((strcmp(argv[n],"-DEBUG") == 0) && (n + 1 < argc))
250 {
251 trace_ok = true;
252 trace_from = (uint32_t) strtol(argv[n + 1], NULL, 0);
253 }
254 else if ((strcmp(argv[n], "-MEMCID") == 0) && (n + 1 < argc))
255 {
256 trace_memc_ok = true;
257 trace_memc_id = (size_t) strtol(argv[n + 1], NULL, 0);
258 size_t x = trace_memc_id >> Y_WIDTH;
259 size_t y = trace_memc_id & ((1<<Y_WIDTH)-1);
260
[937]261 assert( (x < XMAX) and (y < (YMAX)) and
[621]262 "MEMCID parameter refers a not valid memory cache");
263 }
264 else if ((strcmp(argv[n], "-PROCID") == 0) && (n + 1 < argc))
265 {
266 trace_proc_ok = true;
267 trace_proc_id = (size_t) strtol(argv[n + 1], NULL, 0);
[803]268 size_t cluster_xy = trace_proc_id >> P_WIDTH ;
[621]269 size_t x = cluster_xy >> Y_WIDTH;
270 size_t y = cluster_xy & ((1<<Y_WIDTH)-1);
[803]271 size_t l = trace_proc_id & ((1<<P_WIDTH)-1) ;
[621]272
[937]273 assert( (x < XMAX) and (y < YMAX) and (l < NB_PROCS_MAX) and
[621]274 "PROCID parameter refers a not valid processor");
275 }
[937]276 else if ((strcmp(argv[n], "-ROM") == 0) && ((n + 1) < argc))
[692]277 {
[792]278 strcpy(soft_name, argv[n + 1]);
[692]279 }
280 else if ((strcmp(argv[n], "-DISK") == 0) && ((n + 1) < argc))
281 {
282 strcpy(disk_name, argv[n + 1]);
283 }
[621]284 else if ((strcmp(argv[n], "-THREADS") == 0) && ((n + 1) < argc))
285 {
[628]286 threads = (size_t) strtol(argv[n + 1], NULL, 0);
287 threads = (threads < 1) ? 1 : threads;
[621]288 }
289 else if ((strcmp(argv[n], "-FROZEN") == 0) && (n + 1 < argc))
290 {
291 frozen_cycles = (uint32_t) strtol(argv[n + 1], NULL, 0);
292 }
293 else
294 {
295 std::cout << " Arguments are (key,value) couples." << std::endl;
296 std::cout << " The order is not important." << std::endl;
297 std::cout << " Accepted arguments are :" << std::endl << std::endl;
[937]298 std::cout << " - NCYCLES number_of_simulated_cycles" << std::endl;
299 std::cout << " - DEBUG debug_start_cycle" << std::endl;
300 std::cout << " - ROM path to ROM image" << std::endl;
301 std::cout << " - DISK path to disk image" << std::endl;
302 std::cout << " - THREADS simulator's threads number" << std::endl;
303 std::cout << " - FROZEN max_number_of_lines" << std::endl;
304 std::cout << " - PERIOD number_of_cycles between trace" << std::endl;
305 std::cout << " - MEMCID index_memc_to_be_traced" << std::endl;
306 std::cout << " - PROCID index_proc_to_be_traced" << std::endl;
[621]307 exit(0);
308 }
309 }
310 }
311
312 // checking hardware parameters
[937]313 assert( ((X_SIZE <= 16) and (X_SIZE > 0)) and
[681]314 "Illegal X_SIZE parameter" );
[621]315
[937]316 assert( ((Y_SIZE <= 16) and (Y_SIZE > 1)) and
[681]317 "Illegal Y_SIZE parameter" );
[621]318
[803]319 assert( (P_WIDTH <= 2) and
320 "P_WIDTH parameter cannot be larger than 2" );
321
[664]322 assert( (NB_PROCS_MAX <= 4) and
[681]323 "Illegal NB_PROCS_MAX parameter" );
[621]324
[937]325 assert( (NB_CMA_CHANNELS <= 4) and
[628]326 "The NB_CMA_CHANNELS parameter cannot be larger than 4" );
[621]327
[937]328 assert( (NB_TTY_CHANNELS <= 8) and
329 "The NB_TTY_CHANNELS parameter cannot be larger than 16" );
[621]330
[937]331 assert( (NB_NIC_CHANNELS <= 2) and
[628]332 "The NB_NIC_CHANNELS parameter cannot be larger than 2" );
[621]333
334 assert( (vci_address_width == 40) and
335 "VCI address width with the GIET must be 40 bits" );
336
337 assert( (X_WIDTH == 4) and (Y_WIDTH == 4) and
[937]338 "You must have X_WIDTH == Y_WIDTH == 4");
[792]339
[621]340 std::cout << std::endl;
341
[937]342 std::cout << " - XMAX = " << XMAX << std::endl;
343 std::cout << " - YMAX = " << YMAX << std::endl;
[621]344 std::cout << " - NB_PROCS_MAX = " << NB_PROCS_MAX << std::endl;
345 std::cout << " - NB_TTY_CHANNELS = " << NB_TTY_CHANNELS << std::endl;
346 std::cout << " - NB_NIC_CHANNELS = " << NB_NIC_CHANNELS << std::endl;
[937]347 std::cout << " - NB_CMA_CHANNELS = " << NB_CMA_CHANNELS << std::endl;
[621]348 std::cout << " - MEMC_WAYS = " << MEMC_WAYS << std::endl;
349 std::cout << " - MEMC_SETS = " << MEMC_SETS << std::endl;
350 std::cout << " - RAM_LATENCY = " << XRAM_LATENCY << std::endl;
351 std::cout << " - MAX_FROZEN = " << frozen_cycles << std::endl;
352 std::cout << " - MAX_CYCLES = " << ncycles << std::endl;
353 std::cout << " - RESET_ADDRESS = " << RESET_ADDRESS << std::endl;
[692]354 std::cout << " - SOFT_FILENAME = " << soft_name << std::endl;
355 std::cout << " - DISK_IMAGENAME = " << disk_name << std::endl;
356 std::cout << " - OPENMP THREADS = " << threads << std::endl;
[621]357
358 std::cout << std::endl;
359
360 // Internal and External VCI parameters definition
361 typedef soclib::caba::VciParams<vci_cell_width_int,
362 vci_plen_width,
363 vci_address_width,
364 vci_rerror_width,
365 vci_clen_width,
366 vci_rflag_width,
367 vci_srcid_width,
368 vci_pktid_width,
369 vci_trdid_width,
370 vci_wrplen_width> vci_param_int;
371
372 typedef soclib::caba::VciParams<vci_cell_width_ext,
373 vci_plen_width,
374 vci_address_width,
375 vci_rerror_width,
376 vci_clen_width,
377 vci_rflag_width,
378 vci_srcid_width,
379 vci_pktid_width,
380 vci_trdid_width,
381 vci_wrplen_width> vci_param_ext;
382
383#if USE_OPENMP
384 omp_set_dynamic(false);
[628]385 omp_set_num_threads(threads);
[621]386 std::cerr << "Built with openmp version " << _OPENMP << std::endl;
387#endif
388
389
[628]390 ///////////////////////////////////////
391 // Direct Network Mapping Table
392 ///////////////////////////////////////
[621]393
[792]394 MappingTable maptabd(vci_address_width,
395 IntTab(X_WIDTH + Y_WIDTH, 16 - X_WIDTH - Y_WIDTH),
396 IntTab(X_WIDTH + Y_WIDTH, vci_srcid_width - X_WIDTH - Y_WIDTH),
[628]397 0x00FF000000ULL);
[621]398
[628]399 // replicated segments
[937]400 for (size_t x = 0; x < XMAX; x++)
[621]401 {
[937]402 for (size_t y = 0; y < (YMAX) ; y++)
[621]403 {
404 sc_uint<vci_address_width> offset;
[628]405 offset = ((sc_uint<vci_address_width>)cluster(x,y)) << 32;
[621]406
407 std::ostringstream si;
408 si << "seg_xicu_" << x << "_" << y;
[792]409 maptabd.add(Segment(si.str(), SEG_XCU_BASE + offset, SEG_XCU_SIZE,
[621]410 IntTab(cluster(x,y),XICU_TGTID), false));
411
412 std::ostringstream sd;
[628]413 sd << "seg_mcfg_" << x << "_" << y;
[792]414 maptabd.add(Segment(sd.str(), SEG_MMC_BASE + offset, SEG_MMC_SIZE,
[628]415 IntTab(cluster(x,y),MEMC_TGTID), false));
[621]416
417 std::ostringstream sh;
418 sh << "seg_memc_" << x << "_" << y;
[792]419 maptabd.add(Segment(sh.str(), SEG_RAM_BASE + offset, SEG_RAM_SIZE,
[621]420 IntTab(cluster(x,y),MEMC_TGTID), true));
421 }
422 }
[628]423
[664]424 // segments for peripherals in cluster(0,0)
[792]425 maptabd.add(Segment("seg_tty0", SEG_TTY_BASE, SEG_TTY_SIZE,
[628]426 IntTab(cluster(0,0),MTTY_TGTID), false));
427
[792]428 maptabd.add(Segment("seg_ioc0", SEG_IOC_BASE, SEG_IOC_SIZE,
[628]429 IntTab(cluster(0,0),BDEV_TGTID), false));
430
[937]431 // segments for peripherals in cluster_io (XMAX-1,YMAX)
[628]432 sc_uint<vci_address_width> offset;
[937]433 offset = ((sc_uint<vci_address_width>)cluster(XMAX-1,YMAX)) << 32;
[628]434
[792]435 maptabd.add(Segment("seg_mtty", SEG_TTY_BASE + offset, SEG_TTY_SIZE,
[937]436 IntTab(cluster(XMAX-1, YMAX),MTTY_TGTID), false));
[628]437
[792]438 maptabd.add(Segment("seg_fbuf", SEG_FBF_BASE + offset, SEG_FBF_SIZE,
[937]439 IntTab(cluster(XMAX-1, YMAX),FBUF_TGTID), false));
[628]440
[792]441 maptabd.add(Segment("seg_bdev", SEG_IOC_BASE + offset, SEG_IOC_SIZE,
[937]442 IntTab(cluster(XMAX-1, YMAX),BDEV_TGTID), false));
[628]443
[792]444 maptabd.add(Segment("seg_mnic", SEG_NIC_BASE + offset, SEG_NIC_SIZE,
[937]445 IntTab(cluster(XMAX-1, YMAX),MNIC_TGTID), false));
[628]446
[792]447 maptabd.add(Segment("seg_cdma", SEG_CMA_BASE + offset, SEG_CMA_SIZE,
[937]448 IntTab(cluster(XMAX-1, YMAX),CDMA_TGTID), false));
[628]449
[792]450 maptabd.add(Segment("seg_iopi", SEG_PIC_BASE + offset, SEG_PIC_SIZE,
[937]451 IntTab(cluster(XMAX-1, YMAX),IOPI_TGTID), false));
[628]452
[621]453 std::cout << maptabd << std::endl;
454
[628]455 /////////////////////////////////////////////////
456 // Ram network mapping table
457 /////////////////////////////////////////////////
[621]458
[792]459 MappingTable maptabx(vci_address_width,
460 IntTab(X_WIDTH+Y_WIDTH),
461 IntTab(X_WIDTH+Y_WIDTH),
[628]462 0x00FF000000ULL);
[621]463
[937]464 for (size_t x = 0; x < XMAX; x++)
[628]465 {
[937]466 for (size_t y = 0; y < (YMAX) ; y++)
[792]467 {
[628]468 sc_uint<vci_address_width> offset;
[792]469 offset = (sc_uint<vci_address_width>)cluster(x,y)
[628]470 << (vci_address_width-X_WIDTH-Y_WIDTH);
[621]471
[628]472 std::ostringstream sh;
473 sh << "x_seg_memc_" << x << "_" << y;
[621]474
[792]475 maptabx.add(Segment(sh.str(), SEG_RAM_BASE + offset,
476 SEG_RAM_SIZE, IntTab(cluster(x,y)), false));
[628]477 }
478 }
479 std::cout << maptabx << std::endl;
[621]480
[628]481 ////////////////////
482 // Signals
483 ///////////////////
[621]484
[628]485 sc_clock signal_clk("clk");
486 sc_signal<bool> signal_resetn("resetn");
[621]487
[628]488 // IRQs from external peripherals
489 sc_signal<bool> signal_irq_bdev;
490 sc_signal<bool> signal_irq_mnic_rx[NB_NIC_CHANNELS];
491 sc_signal<bool> signal_irq_mnic_tx[NB_NIC_CHANNELS];
[664]492 sc_signal<bool> signal_irq_mtty_rx[NB_TTY_CHANNELS];
[628]493 sc_signal<bool> signal_irq_cdma[NB_CMA_CHANNELS];
494 sc_signal<bool> signal_irq_false;
495
[621]496 // Horizontal inter-clusters DSPIN signals
[628]497 DspinSignals<dspin_cmd_width>** signal_dspin_h_cmd_inc =
[937]498 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_cmd_inc", XMAX-1, YMAX);
[628]499 DspinSignals<dspin_cmd_width>** signal_dspin_h_cmd_dec =
[937]500 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_cmd_dec", XMAX-1, YMAX);
[621]501
[628]502 DspinSignals<dspin_rsp_width>** signal_dspin_h_rsp_inc =
[937]503 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_h_rsp_inc", XMAX-1, YMAX);
[628]504 DspinSignals<dspin_rsp_width>** signal_dspin_h_rsp_dec =
[937]505 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_h_rsp_dec", XMAX-1, YMAX);
[628]506
507 DspinSignals<dspin_cmd_width>** signal_dspin_h_m2p_inc =
[937]508 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_m2p_inc", XMAX-1, YMAX);
[628]509 DspinSignals<dspin_cmd_width>** signal_dspin_h_m2p_dec =
[937]510 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_m2p_dec", XMAX-1, YMAX);
[628]511
512 DspinSignals<dspin_rsp_width>** signal_dspin_h_p2m_inc =
[937]513 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_h_p2m_inc", XMAX-1, YMAX);
[628]514 DspinSignals<dspin_rsp_width>** signal_dspin_h_p2m_dec =
[937]515 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_h_p2m_dec", XMAX-1, YMAX);
[628]516
517 DspinSignals<dspin_cmd_width>** signal_dspin_h_cla_inc =
[937]518 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_cla_inc", XMAX-1, YMAX);
[628]519 DspinSignals<dspin_cmd_width>** signal_dspin_h_cla_dec =
[937]520 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_h_cla_dec", XMAX-1, YMAX);
[628]521
[621]522 // Vertical inter-clusters DSPIN signals
[628]523 DspinSignals<dspin_cmd_width>** signal_dspin_v_cmd_inc =
[937]524 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_cmd_inc", XMAX, YMAX-1);
[628]525 DspinSignals<dspin_cmd_width>** signal_dspin_v_cmd_dec =
[937]526 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_cmd_dec", XMAX, YMAX-1);
[621]527
[628]528 DspinSignals<dspin_rsp_width>** signal_dspin_v_rsp_inc =
[937]529 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_v_rsp_inc", XMAX, YMAX-1);
[628]530 DspinSignals<dspin_rsp_width>** signal_dspin_v_rsp_dec =
[937]531 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_v_rsp_dec", XMAX, YMAX-1);
[621]532
[628]533 DspinSignals<dspin_cmd_width>** signal_dspin_v_m2p_inc =
[937]534 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_m2p_inc", XMAX, YMAX-1);
[628]535 DspinSignals<dspin_cmd_width>** signal_dspin_v_m2p_dec =
[937]536 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_m2p_dec", XMAX, YMAX-1);
[621]537
[628]538 DspinSignals<dspin_rsp_width>** signal_dspin_v_p2m_inc =
[937]539 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_v_p2m_inc", XMAX, YMAX-1);
[628]540 DspinSignals<dspin_rsp_width>** signal_dspin_v_p2m_dec =
[937]541 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_v_p2m_dec", XMAX, YMAX-1);
[628]542
543 DspinSignals<dspin_cmd_width>** signal_dspin_v_cla_inc =
[937]544 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_cla_inc", XMAX, YMAX-1);
[628]545 DspinSignals<dspin_cmd_width>** signal_dspin_v_cla_dec =
[937]546 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_v_cla_dec", XMAX, YMAX-1);
[628]547
548 // Mesh boundaries DSPIN signals (Most of those signals are not used...)
549 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_cmd_in =
[937]550 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_cmd_in" , XMAX, YMAX, 4);
[628]551 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_cmd_out =
[937]552 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_cmd_out", XMAX, YMAX, 4);
[628]553
554 DspinSignals<dspin_rsp_width>*** signal_dspin_bound_rsp_in =
[937]555 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_bound_rsp_in" , XMAX, YMAX, 4);
[628]556 DspinSignals<dspin_rsp_width>*** signal_dspin_bound_rsp_out =
[937]557 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_bound_rsp_out", XMAX, YMAX, 4);
[628]558
559 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_m2p_in =
[937]560 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_m2p_in" , XMAX, YMAX, 4);
[628]561 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_m2p_out =
[937]562 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_m2p_out", XMAX, YMAX, 4);
[628]563
564 DspinSignals<dspin_rsp_width>*** signal_dspin_bound_p2m_in =
[937]565 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_bound_p2m_in" , XMAX, YMAX, 4);
[628]566 DspinSignals<dspin_rsp_width>*** signal_dspin_bound_p2m_out =
[937]567 alloc_elems<DspinSignals<dspin_rsp_width> >("signal_dspin_bound_p2m_out", XMAX, YMAX, 4);
[628]568
569 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_cla_in =
[937]570 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_cla_in" , XMAX, YMAX, 4);
[628]571 DspinSignals<dspin_cmd_width>*** signal_dspin_bound_cla_out =
[937]572 alloc_elems<DspinSignals<dspin_cmd_width> >("signal_dspin_bound_cla_out", XMAX, YMAX, 4);
[628]573
574 // VCI signals for iobus and peripherals
575 VciSignals<vci_param_int> signal_vci_ini_bdev("signal_vci_ini_bdev");
576 VciSignals<vci_param_int> signal_vci_ini_cdma("signal_vci_ini_cdma");
577 VciSignals<vci_param_int> signal_vci_ini_iopi("signal_vci_ini_iopi");
578
[792]579 VciSignals<vci_param_int>* signal_vci_ini_proc =
[628]580 alloc_elems<VciSignals<vci_param_int> >("signal_vci_ini_proc", NB_PROCS_MAX );
581
582 VciSignals<vci_param_int> signal_vci_tgt_memc("signal_vci_tgt_memc");
583 VciSignals<vci_param_int> signal_vci_tgt_xicu("signal_vci_tgt_xicu");
584 VciSignals<vci_param_int> signal_vci_tgt_bdev("signal_vci_tgt_bdev");
585 VciSignals<vci_param_int> signal_vci_tgt_mtty("signal_vci_tgt_mtty");
586 VciSignals<vci_param_int> signal_vci_tgt_fbuf("signal_vci_tgt_fbuf");
587 VciSignals<vci_param_int> signal_vci_tgt_mnic("signal_vci_tgt_mnic");
588 VciSignals<vci_param_int> signal_vci_tgt_cdma("signal_vci_tgt_cdma");
589 VciSignals<vci_param_int> signal_vci_tgt_iopi("signal_vci_tgt_iopi");
590
591 VciSignals<vci_param_int> signal_vci_cmd_to_noc("signal_vci_cmd_to_noc");
592 VciSignals<vci_param_int> signal_vci_cmd_from_noc("signal_vci_cmd_from_noc");
[792]593
[621]594 ////////////////////////////
[792]595 // Loader
[621]596 ////////////////////////////
597
[796]598#if USE_IOC_RDK
[795]599 std::ostringstream ramdisk_name;
600 ramdisk_name << disk_name << "@" << std::hex << SEG_RDK_BASE << ":";
601 soclib::common::Loader loader( soft_name, ramdisk_name.str().c_str() );
[664]602#else
[692]603 soclib::common::Loader loader( soft_name );
[664]604#endif
[937]605
[939]606 loader.memory_default(0x55);
[621]607
608 typedef soclib::common::GdbServer<soclib::common::Mips32ElIss> proc_iss;
[630]609 proc_iss::set_loader( loader );
[621]610
[664]611 //////////////////////////////////////////////////////////////
[937]612 // mesh construction: XMAX * YMAX clusters
[664]613 //////////////////////////////////////////////////////////////
[621]614
615 TsarLetiCluster<dspin_cmd_width,
616 dspin_rsp_width,
617 vci_param_int,
[937]618 vci_param_ext>* clusters[XMAX][YMAX];
[621]619
620#if USE_OPENMP
621#pragma omp parallel
622 {
623#pragma omp for
624#endif
[937]625 for (size_t i = 0; i < (XMAX * YMAX); i++)
[621]626 {
[937]627 size_t x = i / (YMAX);
628 size_t y = i % (YMAX);
[621]629
630#if USE_OPENMP
631#pragma omp critical
632 {
633#endif
634 std::cout << std::endl;
[792]635 std::cout << "Cluster_" << std::dec << x << "_" << y
[628]636 << " with cluster_xy = " << std::hex << cluster(x,y) << std::endl;
[621]637 std::cout << std::endl;
638
[664]639 std::ostringstream cluster_name;
640 cluster_name << "cluster_" << std::dec << x << "_" << y;
641
[621]642 clusters[x][y] = new TsarLetiCluster<dspin_cmd_width,
643 dspin_rsp_width,
644 vci_param_int,
645 vci_param_ext>
646 (
[664]647 cluster_name.str().c_str(),
[621]648 NB_PROCS_MAX,
649 x,
650 y,
651 cluster(x,y),
652 maptabd,
653 maptabx,
654 RESET_ADDRESS,
655 X_WIDTH,
656 Y_WIDTH,
657 vci_srcid_width - X_WIDTH - Y_WIDTH, // l_id width,
[803]658 P_WIDTH,
[621]659 MEMC_TGTID,
660 XICU_TGTID,
661 MTTY_TGTID,
662 BDEV_TGTID,
[692]663 disk_name,
[621]664 MEMC_WAYS,
665 MEMC_SETS,
666 L1_IWAYS,
667 L1_ISETS,
668 L1_DWAYS,
669 L1_DSETS,
670 XRAM_LATENCY,
671 loader,
672 frozen_cycles,
673 trace_from,
[792]674 trace_proc_ok,
[621]675 trace_proc_id,
[792]676 trace_memc_ok,
[621]677 trace_memc_id
678 );
679
680#if USE_OPENMP
681 } // end critical
682#endif
683 } // end for
684#if USE_OPENMP
685 }
686#endif
687
[937]688#if USE_PIC
[826]689
[628]690 //////////////////////////////////////////////////////////////////
[937]691 // IO bus and external peripherals in cluster[X_SIZE-1][Y_SIZE-1]
[664]692 // - 6 local targets : FBF, TTY, CMA, NIC, PIC, IOC
693 // - 3 local initiators : IOC, CMA, PIC
694 // There is no PROC, no MEMC and no XICU in this cluster,
695 // but the crossbar has (NB_PROCS_MAX + 3) intiators and
696 // 8 targets, in order to use the same SRCID and TGTID space
[792]697 // (same mapping table for the internal components,
698 // and for the external peripherals)
[628]699 //////////////////////////////////////////////////////////////////
[621]700
[664]701 std::cout << std::endl;
702 std::cout << " Building IO cluster (external peripherals)" << std::endl;
703 std::cout << std::endl;
[792]704
[937]705 size_t cluster_io = cluster(XMAX-1, YMAX);
[621]706
[792]707 //////////// vci_local_crossbar
[628]708 VciLocalCrossbar<vci_param_int>*
709 iobus = new VciLocalCrossbar<vci_param_int>(
710 "iobus",
711 maptabd, // mapping table
712 cluster_io, // cluster_xy
713 NB_PROCS_MAX + 3, // number of local initiators
714 8, // number of local targets
715 BDEV_TGTID ); // default target index
[621]716
[792]717 //////////// vci_framebuffer
[628]718 VciFrameBuffer<vci_param_int>*
719 fbuf = new VciFrameBuffer<vci_param_int>(
720 "fbuf",
721 IntTab(cluster_io, FBUF_TGTID),
722 maptabd,
723 FBUF_X_SIZE, FBUF_Y_SIZE );
[621]724
[792]725 //////////// vci_block_device
[628]726 VciBlockDeviceTsar<vci_param_int>*
727 bdev = new VciBlockDeviceTsar<vci_param_int>(
728 "bdev",
729 maptabd,
730 IntTab(cluster_io, BDEV_SRCID),
731 IntTab(cluster_io, BDEV_TGTID),
[692]732 disk_name,
[628]733 512, // block size
734 64 ); // burst size
735
[792]736 //////////// vci_multi_nic
[628]737 VciMultiNic<vci_param_int>*
738 mnic = new VciMultiNic<vci_param_int>(
[681]739 "mnic",
[628]740 IntTab(cluster_io, MNIC_TGTID),
741 maptabd,
742 NB_NIC_CHANNELS,
[937]743 0, // default MAC_4 address
744 0, // default MAC_2 address
745 1 ); // NIC_MODE_SYNTHESIS
[628]746
[792]747 ///////////// vci_chbuf_dma
[628]748 VciChbufDma<vci_param_int>*
749 cdma = new VciChbufDma<vci_param_int>(
750 "cdma",
751 maptabd,
752 IntTab(cluster_io, CDMA_SRCID),
753 IntTab(cluster_io, CDMA_TGTID),
[937]754 64, // burst size
[792]755 NB_CMA_CHANNELS );
[628]756
757 ////////////// vci_multi_tty
758 std::vector<std::string> vect_names;
759 for (size_t id = 0; id < NB_TTY_CHANNELS; id++)
760 {
761 std::ostringstream term_name;
762 term_name << "ext_" << id;
763 vect_names.push_back(term_name.str().c_str());
764 }
765
[792]766 VciMultiTty<vci_param_int>*
[628]767 mtty = new VciMultiTty<vci_param_int>(
768 "mtty",
769 IntTab(cluster_io, MTTY_TGTID),
770 maptabd,
771 vect_names );
772
773 ///////////// vci_iopic
774 VciIopic<vci_param_int>*
775 iopic = new VciIopic<vci_param_int>(
776 "iopic",
777 maptabd,
778 IntTab(cluster_io, IOPI_SRCID),
779 IntTab(cluster_io, IOPI_TGTID),
[792]780 32 );
[628]781
[792]782 ////////////// vci_dspin wrappers
[628]783 VciDspinTargetWrapper<vci_param_int, dspin_cmd_width, dspin_rsp_width>*
784 wt_iobus = new VciDspinTargetWrapper<vci_param_int, dspin_cmd_width, dspin_rsp_width>(
[937]785 "wt_iobus",
[628]786 vci_srcid_width );
787
788 VciDspinInitiatorWrapper<vci_param_int, dspin_cmd_width, dspin_rsp_width>*
789 wi_iobus = new VciDspinInitiatorWrapper<vci_param_int, dspin_cmd_width, dspin_rsp_width>(
[937]790 "wi_iobus",
[628]791 vci_srcid_width );
792
793 ///////////////////////////////////////////////////////////////
[937]794 // IObus Net-list
[628]795 ///////////////////////////////////////////////////////////////
796
[792]797 // iobus
798 iobus->p_clk (signal_clk);
[628]799 iobus->p_resetn (signal_resetn);
800
801 iobus->p_target_to_up (signal_vci_cmd_from_noc);
802 iobus->p_initiator_to_up (signal_vci_cmd_to_noc);
803
804 iobus->p_to_target[MEMC_TGTID] (signal_vci_tgt_memc);
805 iobus->p_to_target[XICU_TGTID] (signal_vci_tgt_xicu);
806 iobus->p_to_target[MTTY_TGTID] (signal_vci_tgt_mtty);
807 iobus->p_to_target[FBUF_TGTID] (signal_vci_tgt_fbuf);
808 iobus->p_to_target[MNIC_TGTID] (signal_vci_tgt_mnic);
809 iobus->p_to_target[BDEV_TGTID] (signal_vci_tgt_bdev);
810 iobus->p_to_target[CDMA_TGTID] (signal_vci_tgt_cdma);
811 iobus->p_to_target[IOPI_TGTID] (signal_vci_tgt_iopi);
812
813 for( size_t p=0 ; p<NB_PROCS_MAX ; p++ )
814 {
815 iobus->p_to_initiator[p] (signal_vci_ini_proc[p]);
816 }
817 iobus->p_to_initiator[BDEV_SRCID] (signal_vci_ini_bdev);
818 iobus->p_to_initiator[CDMA_SRCID] (signal_vci_ini_cdma);
819 iobus->p_to_initiator[IOPI_SRCID] (signal_vci_ini_iopi);
820
821 std::cout << " - IOBUS connected" << std::endl;
822
823 // block_device
824 bdev->p_clk (signal_clk);
825 bdev->p_resetn (signal_resetn);
826 bdev->p_vci_target (signal_vci_tgt_bdev);
827 bdev->p_vci_initiator (signal_vci_ini_bdev);
828 bdev->p_irq (signal_irq_bdev);
829
830 std::cout << " - BDEV connected" << std::endl;
831
832 // frame_buffer
833 fbuf->p_clk (signal_clk);
834 fbuf->p_resetn (signal_resetn);
835 fbuf->p_vci (signal_vci_tgt_fbuf);
836
837 std::cout << " - FBUF connected" << std::endl;
838
839 // multi_nic
840 mnic->p_clk (signal_clk);
841 mnic->p_resetn (signal_resetn);
842 mnic->p_vci (signal_vci_tgt_mnic);
843 for ( size_t i=0 ; i<NB_NIC_CHANNELS ; i++ )
844 {
845 mnic->p_rx_irq[i] (signal_irq_mnic_rx[i]);
846 mnic->p_tx_irq[i] (signal_irq_mnic_tx[i]);
847 }
848
849 std::cout << " - MNIC connected" << std::endl;
850
851 // chbuf_dma
852 cdma->p_clk (signal_clk);
853 cdma->p_resetn (signal_resetn);
854 cdma->p_vci_target (signal_vci_tgt_cdma);
855 cdma->p_vci_initiator (signal_vci_ini_cdma);
856 for ( size_t i=0 ; i<NB_CMA_CHANNELS ; i++)
857 {
858 cdma->p_irq[i] (signal_irq_cdma[i]);
859 }
860
861 std::cout << " - CDMA connected" << std::endl;
862
863 // multi_tty
864 mtty->p_clk (signal_clk);
865 mtty->p_resetn (signal_resetn);
866 mtty->p_vci (signal_vci_tgt_mtty);
867 for ( size_t i=0 ; i<NB_TTY_CHANNELS ; i++ )
868 {
[664]869 mtty->p_irq[i] (signal_irq_mtty_rx[i]);
[628]870 }
871
872 std::cout << " - MTTY connected" << std::endl;
873
874 // iopic
[664]875 // NB_NIC_CHANNELS <= 2
876 // NB_CMA_CHANNELS <= 4
[937]877 // NB_TTY_CHANNELS <= 16
[628]878 iopic->p_clk (signal_clk);
879 iopic->p_resetn (signal_resetn);
880 iopic->p_vci_target (signal_vci_tgt_iopi);
881 iopic->p_vci_initiator (signal_vci_ini_iopi);
[664]882 for ( size_t i=0 ; i<32 ; i++)
[628]883 {
884 if (i < NB_NIC_CHANNELS) iopic->p_hwi[i] (signal_irq_mnic_rx[i]);
885 else if(i < 2 ) iopic->p_hwi[i] (signal_irq_false);
886 else if(i < 2+NB_NIC_CHANNELS) iopic->p_hwi[i] (signal_irq_mnic_tx[i-2]);
887 else if(i < 4 ) iopic->p_hwi[i] (signal_irq_false);
888 else if(i < 4+NB_CMA_CHANNELS) iopic->p_hwi[i] (signal_irq_cdma[i-4]);
[664]889 else if(i < 8) iopic->p_hwi[i] (signal_irq_false);
890 else if(i == 8) iopic->p_hwi[i] (signal_irq_bdev);
891 else if(i < 16) iopic->p_hwi[i] (signal_irq_false);
892 else if(i < 16+NB_TTY_CHANNELS) iopic->p_hwi[i] (signal_irq_mtty_rx[i-16]);
[628]893 else iopic->p_hwi[i] (signal_irq_false);
894 }
895
[664]896 std::cout << " - IOPIC connected" << std::endl;
897
[628]898 // vci/dspin wrappers
899 wi_iobus->p_clk (signal_clk);
900 wi_iobus->p_resetn (signal_resetn);
901 wi_iobus->p_vci (signal_vci_cmd_to_noc);
[937]902 wi_iobus->p_dspin_cmd (signal_dspin_bound_cmd_in[XMAX-1][YMAX-1][NORTH]);
903 wi_iobus->p_dspin_rsp (signal_dspin_bound_rsp_out[XMAX-1][YMAX-1][NORTH]);
[628]904
905 // vci/dspin wrappers
906 wt_iobus->p_clk (signal_clk);
907 wt_iobus->p_resetn (signal_resetn);
908 wt_iobus->p_vci (signal_vci_cmd_from_noc);
[937]909 wt_iobus->p_dspin_cmd (signal_dspin_bound_cmd_out[XMAX-1][YMAX-1][NORTH]);
910 wt_iobus->p_dspin_rsp (signal_dspin_bound_rsp_in[XMAX-1][YMAX-1][NORTH]);
[628]911
[937]912#endif // USE_PIC
[826]913
[628]914 // Clock & RESET for clusters
[937]915 for (size_t x = 0; x < (XMAX); x++)
[628]916 {
[937]917 for (size_t y = 0; y < (YMAX); y++)
[628]918 {
919 clusters[x][y]->p_clk (signal_clk);
920 clusters[x][y]->p_resetn (signal_resetn);
921 }
922 }
923
924 // Inter Clusters horizontal connections
[937]925 if (XMAX > 1)
[628]926 {
[937]927 for (size_t x = 0; x < (XMAX-1); x++)
[628]928 {
[937]929 for (size_t y = 0; y < (YMAX); y++)
[628]930 {
931 clusters[x][y]->p_cmd_out[EAST] (signal_dspin_h_cmd_inc[x][y]);
932 clusters[x+1][y]->p_cmd_in[WEST] (signal_dspin_h_cmd_inc[x][y]);
933 clusters[x][y]->p_cmd_in[EAST] (signal_dspin_h_cmd_dec[x][y]);
934 clusters[x+1][y]->p_cmd_out[WEST] (signal_dspin_h_cmd_dec[x][y]);
935
936 clusters[x][y]->p_rsp_out[EAST] (signal_dspin_h_rsp_inc[x][y]);
937 clusters[x+1][y]->p_rsp_in[WEST] (signal_dspin_h_rsp_inc[x][y]);
938 clusters[x][y]->p_rsp_in[EAST] (signal_dspin_h_rsp_dec[x][y]);
939 clusters[x+1][y]->p_rsp_out[WEST] (signal_dspin_h_rsp_dec[x][y]);
940
941 clusters[x][y]->p_m2p_out[EAST] (signal_dspin_h_m2p_inc[x][y]);
942 clusters[x+1][y]->p_m2p_in[WEST] (signal_dspin_h_m2p_inc[x][y]);
943 clusters[x][y]->p_m2p_in[EAST] (signal_dspin_h_m2p_dec[x][y]);
944 clusters[x+1][y]->p_m2p_out[WEST] (signal_dspin_h_m2p_dec[x][y]);
945
946 clusters[x][y]->p_p2m_out[EAST] (signal_dspin_h_p2m_inc[x][y]);
947 clusters[x+1][y]->p_p2m_in[WEST] (signal_dspin_h_p2m_inc[x][y]);
948 clusters[x][y]->p_p2m_in[EAST] (signal_dspin_h_p2m_dec[x][y]);
949 clusters[x+1][y]->p_p2m_out[WEST] (signal_dspin_h_p2m_dec[x][y]);
950
951 clusters[x][y]->p_cla_out[EAST] (signal_dspin_h_cla_inc[x][y]);
952 clusters[x+1][y]->p_cla_in[WEST] (signal_dspin_h_cla_inc[x][y]);
953 clusters[x][y]->p_cla_in[EAST] (signal_dspin_h_cla_dec[x][y]);
954 clusters[x+1][y]->p_cla_out[WEST] (signal_dspin_h_cla_dec[x][y]);
[621]955 }
[628]956 }
957 }
[792]958 std::cout << std::endl << "Horizontal connections done" << std::endl;
[621]959
[628]960 // Inter Clusters vertical connections
[937]961 if (YMAX > 1)
[628]962 {
[937]963 for (size_t y = 0; y < (YMAX-1); y++)
[628]964 {
[937]965 for (size_t x = 0; x < XMAX; x++)
[628]966 {
967 clusters[x][y]->p_cmd_out[NORTH] (signal_dspin_v_cmd_inc[x][y]);
968 clusters[x][y+1]->p_cmd_in[SOUTH] (signal_dspin_v_cmd_inc[x][y]);
969 clusters[x][y]->p_cmd_in[NORTH] (signal_dspin_v_cmd_dec[x][y]);
970 clusters[x][y+1]->p_cmd_out[SOUTH] (signal_dspin_v_cmd_dec[x][y]);
971
972 clusters[x][y]->p_rsp_out[NORTH] (signal_dspin_v_rsp_inc[x][y]);
973 clusters[x][y+1]->p_rsp_in[SOUTH] (signal_dspin_v_rsp_inc[x][y]);
974 clusters[x][y]->p_rsp_in[NORTH] (signal_dspin_v_rsp_dec[x][y]);
975 clusters[x][y+1]->p_rsp_out[SOUTH] (signal_dspin_v_rsp_dec[x][y]);
976
977 clusters[x][y]->p_m2p_out[NORTH] (signal_dspin_v_m2p_inc[x][y]);
978 clusters[x][y+1]->p_m2p_in[SOUTH] (signal_dspin_v_m2p_inc[x][y]);
979 clusters[x][y]->p_m2p_in[NORTH] (signal_dspin_v_m2p_dec[x][y]);
980 clusters[x][y+1]->p_m2p_out[SOUTH] (signal_dspin_v_m2p_dec[x][y]);
981
982 clusters[x][y]->p_p2m_out[NORTH] (signal_dspin_v_p2m_inc[x][y]);
983 clusters[x][y+1]->p_p2m_in[SOUTH] (signal_dspin_v_p2m_inc[x][y]);
984 clusters[x][y]->p_p2m_in[NORTH] (signal_dspin_v_p2m_dec[x][y]);
985 clusters[x][y+1]->p_p2m_out[SOUTH] (signal_dspin_v_p2m_dec[x][y]);
986
987 clusters[x][y]->p_cla_out[NORTH] (signal_dspin_v_cla_inc[x][y]);
988 clusters[x][y+1]->p_cla_in[SOUTH] (signal_dspin_v_cla_inc[x][y]);
989 clusters[x][y]->p_cla_in[NORTH] (signal_dspin_v_cla_dec[x][y]);
990 clusters[x][y+1]->p_cla_out[SOUTH] (signal_dspin_v_cla_dec[x][y]);
[621]991 }
[628]992 }
993 }
994 std::cout << std::endl << "Vertical connections done" << std::endl;
[621]995
[628]996 // East & West boundary cluster connections
[937]997 for (size_t y = 0; y < (YMAX); y++)
[628]998 {
999 clusters[0][y]->p_cmd_in[WEST] (signal_dspin_bound_cmd_in[0][y][WEST]);
1000 clusters[0][y]->p_cmd_out[WEST] (signal_dspin_bound_cmd_out[0][y][WEST]);
[937]1001 clusters[XMAX-1][y]->p_cmd_in[EAST] (signal_dspin_bound_cmd_in[XMAX-1][y][EAST]);
1002 clusters[XMAX-1][y]->p_cmd_out[EAST] (signal_dspin_bound_cmd_out[XMAX-1][y][EAST]);
[621]1003
[628]1004 clusters[0][y]->p_rsp_in[WEST] (signal_dspin_bound_rsp_in[0][y][WEST]);
1005 clusters[0][y]->p_rsp_out[WEST] (signal_dspin_bound_rsp_out[0][y][WEST]);
[937]1006 clusters[XMAX-1][y]->p_rsp_in[EAST] (signal_dspin_bound_rsp_in[XMAX-1][y][EAST]);
1007 clusters[XMAX-1][y]->p_rsp_out[EAST] (signal_dspin_bound_rsp_out[XMAX-1][y][EAST]);
[621]1008
[628]1009 clusters[0][y]->p_m2p_in[WEST] (signal_dspin_bound_m2p_in[0][y][WEST]);
1010 clusters[0][y]->p_m2p_out[WEST] (signal_dspin_bound_m2p_out[0][y][WEST]);
[937]1011 clusters[XMAX-1][y]->p_m2p_in[EAST] (signal_dspin_bound_m2p_in[XMAX-1][y][EAST]);
1012 clusters[XMAX-1][y]->p_m2p_out[EAST] (signal_dspin_bound_m2p_out[XMAX-1][y][EAST]);
[621]1013
[628]1014 clusters[0][y]->p_p2m_in[WEST] (signal_dspin_bound_p2m_in[0][y][WEST]);
1015 clusters[0][y]->p_p2m_out[WEST] (signal_dspin_bound_p2m_out[0][y][WEST]);
[937]1016 clusters[XMAX-1][y]->p_p2m_in[EAST] (signal_dspin_bound_p2m_in[XMAX-1][y][EAST]);
1017 clusters[XMAX-1][y]->p_p2m_out[EAST] (signal_dspin_bound_p2m_out[XMAX-1][y][EAST]);
[621]1018
[628]1019 clusters[0][y]->p_cla_in[WEST] (signal_dspin_bound_cla_in[0][y][WEST]);
1020 clusters[0][y]->p_cla_out[WEST] (signal_dspin_bound_cla_out[0][y][WEST]);
[937]1021 clusters[XMAX-1][y]->p_cla_in[EAST] (signal_dspin_bound_cla_in[XMAX-1][y][EAST]);
1022 clusters[XMAX-1][y]->p_cla_out[EAST] (signal_dspin_bound_cla_out[XMAX-1][y][EAST]);
[628]1023 }
[621]1024
[664]1025 std::cout << std::endl << "West & East boundaries connections done" << std::endl;
1026
[628]1027 // North & South boundary clusters connections
[937]1028 for (size_t x = 0; x < XMAX; x++)
[628]1029 {
1030 clusters[x][0]->p_cmd_in[SOUTH] (signal_dspin_bound_cmd_in[x][0][SOUTH]);
1031 clusters[x][0]->p_cmd_out[SOUTH] (signal_dspin_bound_cmd_out[x][0][SOUTH]);
[937]1032 clusters[x][YMAX-1]->p_cmd_in[NORTH] (signal_dspin_bound_cmd_in[x][YMAX-1][NORTH]);
1033 clusters[x][YMAX-1]->p_cmd_out[NORTH] (signal_dspin_bound_cmd_out[x][YMAX-1][NORTH]);
[621]1034
[628]1035 clusters[x][0]->p_rsp_in[SOUTH] (signal_dspin_bound_rsp_in[x][0][SOUTH]);
1036 clusters[x][0]->p_rsp_out[SOUTH] (signal_dspin_bound_rsp_out[x][0][SOUTH]);
[937]1037 clusters[x][YMAX-1]->p_rsp_in[NORTH] (signal_dspin_bound_rsp_in[x][YMAX-1][NORTH]);
1038 clusters[x][YMAX-1]->p_rsp_out[NORTH] (signal_dspin_bound_rsp_out[x][YMAX-1][NORTH]);
[621]1039
[628]1040 clusters[x][0]->p_m2p_in[SOUTH] (signal_dspin_bound_m2p_in[x][0][SOUTH]);
1041 clusters[x][0]->p_m2p_out[SOUTH] (signal_dspin_bound_m2p_out[x][0][SOUTH]);
[937]1042 clusters[x][YMAX-1]->p_m2p_in[NORTH] (signal_dspin_bound_m2p_in[x][YMAX-1][NORTH]);
1043 clusters[x][YMAX-1]->p_m2p_out[NORTH] (signal_dspin_bound_m2p_out[x][YMAX-1][NORTH]);
[621]1044
[628]1045 clusters[x][0]->p_p2m_in[SOUTH] (signal_dspin_bound_p2m_in[x][0][SOUTH]);
1046 clusters[x][0]->p_p2m_out[SOUTH] (signal_dspin_bound_p2m_out[x][0][SOUTH]);
[937]1047 clusters[x][YMAX-1]->p_p2m_in[NORTH] (signal_dspin_bound_p2m_in[x][YMAX-1][NORTH]);
1048 clusters[x][YMAX-1]->p_p2m_out[NORTH] (signal_dspin_bound_p2m_out[x][YMAX-1][NORTH]);
[628]1049
1050 clusters[x][0]->p_cla_in[SOUTH] (signal_dspin_bound_cla_in[x][0][SOUTH]);
1051 clusters[x][0]->p_cla_out[SOUTH] (signal_dspin_bound_cla_out[x][0][SOUTH]);
[937]1052 clusters[x][YMAX-1]->p_cla_in[NORTH] (signal_dspin_bound_cla_in[x][YMAX-1][NORTH]);
1053 clusters[x][YMAX-1]->p_cla_out[NORTH] (signal_dspin_bound_cla_out[x][YMAX-1][NORTH]);
[628]1054 }
1055
[664]1056 std::cout << std::endl << "North & South boundaries connections done" << std::endl;
1057
[628]1058 std::cout << std::endl;
1059
1060 ////////////////////////////////////////////////////////
1061 // Simulation
1062 ///////////////////////////////////////////////////////
1063
1064 sc_start(sc_core::sc_time(0, SC_NS));
1065 signal_resetn = false;
1066 signal_irq_false = false;
1067
1068 // set network boundaries signals default values
1069 // for all boundary clusters but the IO cluster
[937]1070 for (size_t x = 0; x < XMAX ; x++)
[628]1071 {
[937]1072 for (size_t y = 0; y < YMAX ; y++)
[628]1073 {
1074 for (size_t face = 0; face < 4; face++)
1075 {
[937]1076 if ( (x != XMAX-1) or (y != YMAX-1) or (face != NORTH) )
[628]1077 {
1078 signal_dspin_bound_cmd_in [x][y][face].write = false;
1079 signal_dspin_bound_cmd_in [x][y][face].read = true;
1080 signal_dspin_bound_cmd_out[x][y][face].write = false;
1081 signal_dspin_bound_cmd_out[x][y][face].read = true;
1082
1083 signal_dspin_bound_rsp_in [x][y][face].write = false;
1084 signal_dspin_bound_rsp_in [x][y][face].read = true;
1085 signal_dspin_bound_rsp_out[x][y][face].write = false;
1086 signal_dspin_bound_rsp_out[x][y][face].read = true;
1087 }
1088
1089 signal_dspin_bound_m2p_in [x][y][face].write = false;
1090 signal_dspin_bound_m2p_in [x][y][face].read = true;
1091 signal_dspin_bound_m2p_out[x][y][face].write = false;
1092 signal_dspin_bound_m2p_out[x][y][face].read = true;
1093
1094 signal_dspin_bound_p2m_in [x][y][face].write = false;
1095 signal_dspin_bound_p2m_in [x][y][face].read = true;
1096 signal_dspin_bound_p2m_out[x][y][face].write = false;
1097 signal_dspin_bound_p2m_out[x][y][face].read = true;
1098
1099 signal_dspin_bound_cla_in [x][y][face].write = false;
1100 signal_dspin_bound_cla_in [x][y][face].read = true;
1101 signal_dspin_bound_cla_out[x][y][face].write = false;
1102 signal_dspin_bound_cla_out[x][y][face].read = true;
[621]1103 }
[628]1104 }
1105 }
[621]1106
[826]1107#if USE_PIC == 0
[937]1108 signal_dspin_bound_cmd_in[XMAX-1][YMAX-1][NORTH].write = false;
1109 signal_dspin_bound_rsp_out[XMAX-1][YMAX-1][NORTH].read = true;
1110 signal_dspin_bound_cmd_out[XMAX-1][YMAX-1][NORTH].read = true;
1111 signal_dspin_bound_rsp_in[XMAX-1][YMAX-1][NORTH].write = false;
[826]1112#endif
1113
[792]1114 // set default values for VCI signals connected to unused ports on iobus
[664]1115 signal_vci_tgt_memc.rspval = false;
1116 signal_vci_tgt_xicu.rspval = false;
1117 for ( size_t p = 0 ; p < NB_PROCS_MAX ; p++ ) signal_vci_ini_proc[p].cmdval = false;
1118
[628]1119 sc_start(sc_core::sc_time(1, SC_NS));
1120 signal_resetn = true;
[621]1121
[792]1122 if (gettimeofday(&t1, NULL) != 0)
[628]1123 {
1124 perror("gettimeofday");
1125 return EXIT_FAILURE;
1126 }
[621]1127
[937]1128 // simulation loop
[664]1129 for (uint64_t n = 1; n < ncycles && !stop_called; n++)
1130 {
[937]1131 // Monitor a specific address for L1 cache
[664]1132 // clusters[0][0]->proc[0]->cache_monitor(0x110002C078ULL);
[621]1133
[937]1134 // Monitor a specific address for L2 cache
[939]1135 // clusters[0][0]->memc->cache_monitor( 0x0000201E00ULL );
[937]1136
1137 // Monitor a specific address for one XRAM
[939]1138 // clusters[0][0]->xram->start_monitor( 0x0000201E00ULL , 64);
[937]1139
[664]1140 // stats display
1141 if( (n % 5000000) == 0)
1142 {
[621]1143
[792]1144 if (gettimeofday(&t2, NULL) != 0)
[664]1145 {
1146 perror("gettimeofday");
1147 return EXIT_FAILURE;
1148 }
[621]1149
[664]1150 ms1 = (uint64_t) t1.tv_sec * 1000ULL + (uint64_t) t1.tv_usec / 1000;
1151 ms2 = (uint64_t) t2.tv_sec * 1000ULL + (uint64_t) t2.tv_usec / 1000;
[792]1152 std::cerr << "platform clock frequency "
[664]1153 << (double) 5000000 / (double) (ms2 - ms1) << "Khz" << std::endl;
[621]1154
[792]1155 if (gettimeofday(&t1, NULL) != 0)
[664]1156 {
1157 perror("gettimeofday");
1158 return EXIT_FAILURE;
1159 }
1160 }
[621]1161
[664]1162 // trace display
1163 if ( trace_ok and (n > trace_from) )
1164 {
1165 std::cout << "****************** cycle " << std::dec << n ;
[937]1166 std::cout << " ********************************************" << std::endl;
[621]1167
[664]1168 size_t l = 0;
1169 size_t x = 0;
1170 size_t y = 0;
[621]1171
[664]1172 if ( trace_proc_ok )
1173 {
[803]1174 l = trace_proc_id & ((1<<P_WIDTH)-1) ;
1175 x = (trace_proc_id >> P_WIDTH) >> Y_WIDTH ;
1176 y = (trace_proc_id >> P_WIDTH) & ((1<<Y_WIDTH) - 1);
[621]1177
[664]1178 std::ostringstream proc_signame;
1179 proc_signame << "[SIG]PROC_" << x << "_" << y << "_" << l ;
[681]1180 clusters[x][y]->proc[l]->print_trace(1);
[664]1181 clusters[x][y]->signal_vci_ini_proc[l].print_trace(proc_signame.str());
[628]1182
[664]1183 std::ostringstream xicu_signame;
1184 xicu_signame << "[SIG]XICU_" << x << "_" << y ;
1185 clusters[x][y]->xicu->print_trace(0);
1186 clusters[x][y]->signal_vci_tgt_xicu.print_trace(xicu_signame.str());
[937]1187
1188 if ( clusters[x][y]->signal_proc_irq[0] )
1189 std::cout << "### IRQ_PROC_" << x << "_" << y << "_0" << std::endl;
1190 if ( clusters[x][y]->signal_proc_irq[4] )
1191 std::cout << "### IRQ_PROC_" << x << "_" << y << "_1" << std::endl;
1192 if ( clusters[x][y]->signal_proc_irq[8] )
1193 std::cout << "### IRQ_PROC_" << x << "_" << y << "_2" << std::endl;
1194 if ( clusters[x][y]->signal_proc_irq[12] )
1195 std::cout << "### IRQ_PROC_" << x << "_" << y << "_3" << std::endl;
[664]1196 }
[628]1197
[664]1198 if ( trace_memc_ok )
1199 {
1200 x = trace_memc_id >> Y_WIDTH;
1201 y = trace_memc_id & ((1<<Y_WIDTH) - 1);
1202
1203 std::ostringstream smemc;
1204 smemc << "[SIG]MEMC_" << x << "_" << y;
1205 std::ostringstream sxram;
1206 sxram << "[SIG]XRAM_" << x << "_" << y;
1207
1208 clusters[x][y]->memc->print_trace();
1209 clusters[x][y]->signal_vci_tgt_memc.print_trace(smemc.str());
1210 clusters[x][y]->signal_vci_xram.print_trace(sxram.str());
[792]1211 }
[664]1212
1213 // trace coherence signals
1214 // clusters[0][0]->signal_dspin_m2p_proc[0].print_trace("[CC_M2P_0_0]");
1215 // clusters[0][1]->signal_dspin_m2p_proc[0].print_trace("[CC_M2P_0_1]");
1216 // clusters[1][0]->signal_dspin_m2p_proc[0].print_trace("[CC_M2P_1_0]");
1217 // clusters[1][1]->signal_dspin_m2p_proc[0].print_trace("[CC_M2P_1_1]");
1218
1219 // clusters[0][0]->signal_dspin_p2m_proc[0].print_trace("[CC_P2M_0_0]");
1220 // clusters[0][1]->signal_dspin_p2m_proc[0].print_trace("[CC_P2M_0_1]");
1221 // clusters[1][0]->signal_dspin_p2m_proc[0].print_trace("[CC_P2M_1_0]");
1222 // clusters[1][1]->signal_dspin_p2m_proc[0].print_trace("[CC_P2M_1_1]");
1223
1224 // trace xbar(s) m2p
1225 // clusters[0][0]->xbar_m2p->print_trace();
1226 // clusters[1][0]->xbar_m2p->print_trace();
1227 // clusters[0][1]->xbar_m2p->print_trace();
1228 // clusters[1][1]->xbar_m2p->print_trace();
[792]1229
[664]1230 // trace router(s) m2p
1231 // clusters[0][0]->router_m2p->print_trace();
1232 // clusters[1][0]->router_m2p->print_trace();
1233 // clusters[0][1]->router_m2p->print_trace();
1234 // clusters[1][1]->router_m2p->print_trace();
[792]1235
[826]1236#if USE_PIC
[664]1237 // trace external ioc
1238 bdev->print_trace();
1239 signal_vci_tgt_bdev.print_trace("[SIG]BDEV_TGT");
1240 signal_vci_ini_bdev.print_trace("[SIG]BDEV_INI");
[621]1241
[664]1242 // trace external iopic
1243 iopic->print_trace();
1244 signal_vci_tgt_iopi.print_trace("[SIG]IOPI_TGT");
1245 signal_vci_ini_iopi.print_trace("[SIG]IOPI_INI");
[937]1246
1247 // trace external interrupts
1248 if (signal_irq_bdev) std::cout << "### IRQ_BDEV" << std::endl;
[897]1249#else
1250 clusters[0][0]->bdev->print_trace();
1251 clusters[0][0]->signal_vci_tgt_bdev.print_trace("[SIG]BDEV_0_0");
1252 clusters[0][0]->signal_vci_ini_bdev.print_trace("[SIG]BDEV_0_0");
[826]1253#endif
[621]1254
[664]1255 } // end trace
[621]1256
[664]1257 sc_start(sc_core::sc_time(1, SC_NS));
1258 }
1259 // Free memory
[681]1260 for (size_t i = 0 ; i < (X_SIZE * Y_SIZE) ; i++)
[664]1261 {
[681]1262 size_t x = i / (Y_SIZE);
1263 size_t y = i % (Y_SIZE);
[664]1264 delete clusters[x][y];
1265 }
1266
1267 return EXIT_SUCCESS;
[621]1268}
1269
[792]1270void handler(int dummy = 0)
[628]1271{
[621]1272 stop_called = true;
1273 sc_stop();
1274}
1275
1276void voidhandler(int dummy = 0) {}
1277
1278int sc_main (int argc, char *argv[])
1279{
1280 signal(SIGINT, handler);
1281 signal(SIGPIPE, voidhandler);
1282
1283 try {
1284 return _main(argc, argv);
1285 } catch (std::exception &e) {
1286 std::cout << e.what() << std::endl;
1287 } catch (...) {
1288 std::cout << "Unknown exception occured" << std::endl;
1289 throw;
1290 }
1291 return 1;
1292}
1293
1294
1295// Local Variables:
1296// tab-width: 3
1297// c-basic-offset: 3
1298// c-file-offsets:((innamespace . 0)(inline-open . 0))
1299// indent-tabs-mode: nil
1300// End:
1301
1302// vim: filetype=cpp:expandtab:shiftwidth=3:tabstop=3:softtabstop=3
Note: See TracBrowser for help on using the repository browser.