mem-ruby,mem-garnet: Multiple networks per RubySystem
[gem5.git] / configs / ruby / MOESI_hammer.py
index da8b004ee32e3754c5a628f5f14b3aa0c82c3301..5aac38b3e956f28195928c5079fd6509b47b2893 100644 (file)
 # THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-#
-# Authors: Brad Beckmann
 
 import math
 import m5
 from m5.objects import *
 from m5.defines import buildEnv
-from Ruby import create_topology
-
-#
-# Note: the L1 Cache latency is only used by the sequencer on fast path hits
-#
-class L1Cache(RubyCache):
-    latency = 2
+from Ruby import create_topology, create_directories
+from Ruby import send_evicts
+from common import FileSystemConfig
 
 #
-# Note: the L2 Cache latency is not currently used
+# Declare caches used by the protocol
 #
-class L2Cache(RubyCache):
-    latency = 10
-
+class L1Cache(RubyCache): pass
+class L2Cache(RubyCache): pass
 #
-# Probe filter is a cache, latency is not used
+# Probe filter is a cache
 #
-class ProbeFilter(RubyCache):
-    latency = 1
+class ProbeFilter(RubyCache): pass
 
 def define_options(parser):
     parser.add_option("--allow-atomic-migration", action="store_true",
@@ -59,20 +51,20 @@ def define_options(parser):
     parser.add_option("--dir-on", action="store_true",
           help="Hammer: enable Full-bit Directory")
 
-def create_system(options, system, piobus, dma_ports, ruby_system):
+def create_system(options, full_system, system, dma_ports, bootmem,
+                  ruby_system):
 
     if buildEnv['PROTOCOL'] != 'MOESI_hammer':
         panic("This script requires the MOESI_hammer protocol to be built.")
 
     cpu_sequencers = []
-    
+
     #
     # The ruby network creation expects the list of nodes in the system to be
     # consistent with the NetDest list.  Therefore the l1 controller nodes must be
     # listed before the directory nodes and directory nodes before dma nodes, etc.
     #
     l1_cntrl_nodes = []
-    dir_cntrl_nodes = []
     dma_cntrl_nodes = []
 
     #
@@ -81,9 +73,7 @@ def create_system(options, system, piobus, dma_ports, ruby_system):
     #
     block_size_bits = int(math.log(options.cacheline_size, 2))
 
-    cntrl_count = 0
-    
-    for i in xrange(options.num_cpus):
+    for i in range(options.num_cpus):
         #
         # First create the Ruby objects associated with this cpu
         #
@@ -98,43 +88,59 @@ def create_system(options, system, piobus, dma_ports, ruby_system):
                            assoc = options.l2_assoc,
                            start_index_bit = block_size_bits)
 
-        l1_cntrl = L1Cache_Controller(version = i,
-                                      cntrl_id = cntrl_count,
-                                      L1Icache = l1i_cache,
-                                      L1Dcache = l1d_cache,
-                                      L2cache = l2_cache,
-                                      no_mig_atomic = not \
-                                        options.allow_atomic_migration,
-                                      send_evictions = (
-                                          options.cpu_type == "detailed"),
-                                      ruby_system = ruby_system)
-
-        cpu_seq = RubySequencer(version = i,
-                                icache = l1i_cache,
-                                dcache = l1d_cache,
-                                ruby_system = ruby_system)
+        # the ruby random tester reuses num_cpus to specify the
+        # number of cpu ports connected to the tester object, which
+        # is stored in system.cpu. because there is only ever one
+        # tester object, num_cpus is not necessarily equal to the
+        # size of system.cpu; therefore if len(system.cpu) == 1
+        # we use system.cpu[0] to set the clk_domain, thereby ensuring
+        # we don't index off the end of the cpu list.
+        if len(system.cpu) == 1:
+            clk_domain = system.cpu[0].clk_domain
+        else:
+            clk_domain = system.cpu[i].clk_domain
 
-        l1_cntrl.sequencer = cpu_seq
+        l1_cntrl = L1Cache_Controller(version=i, L1Icache=l1i_cache,
+                                      L1Dcache=l1d_cache, L2cache=l2_cache,
+                                      no_mig_atomic=not \
+                                      options.allow_atomic_migration,
+                                      send_evictions=send_evicts(options),
+                                      transitions_per_cycle=options.ports,
+                                      clk_domain=clk_domain,
+                                      ruby_system=ruby_system)
 
-        if piobus != None:
-            cpu_seq.pio_port = piobus.slave
+        cpu_seq = RubySequencer(version=i, icache=l1i_cache,
+                                dcache=l1d_cache,clk_domain=clk_domain,
+                                ruby_system=ruby_system)
 
+        l1_cntrl.sequencer = cpu_seq
         if options.recycle_latency:
             l1_cntrl.recycle_latency = options.recycle_latency
 
         exec("ruby_system.l1_cntrl%d = l1_cntrl" % i)
-        #
+
         # Add controllers and sequencers to the appropriate lists
-        #
         cpu_sequencers.append(cpu_seq)
         l1_cntrl_nodes.append(l1_cntrl)
 
-        cntrl_count += 1
+        # Connect the L1 controller and the network
+        # Connect the buffers from the controller to network
+        l1_cntrl.requestFromCache = MessageBuffer()
+        l1_cntrl.requestFromCache.master = ruby_system.network.slave
+        l1_cntrl.responseFromCache = MessageBuffer()
+        l1_cntrl.responseFromCache.master = ruby_system.network.slave
+        l1_cntrl.unblockFromCache = MessageBuffer()
+        l1_cntrl.unblockFromCache.master = ruby_system.network.slave
+
+        l1_cntrl.triggerQueue = MessageBuffer()
+
+        # Connect the buffers from the network to the controller
+        l1_cntrl.mandatoryQueue = MessageBuffer()
+        l1_cntrl.forwardToCache = MessageBuffer()
+        l1_cntrl.forwardToCache.slave = ruby_system.network.master
+        l1_cntrl.responseToCache = MessageBuffer()
+        l1_cntrl.responseToCache.slave = ruby_system.network.master
 
-    phys_mem_size = sum(map(lambda mem: mem.range.size(),
-                            system.memories.unproxy(system)))
-    assert(phys_mem_size % options.num_dirs == 0)
-    mem_module_size = phys_mem_size / options.num_dirs
 
     #
     # determine size and index bits for probe filter
@@ -166,70 +172,117 @@ def create_system(options, system, piobus, dma_ports, ruby_system):
                                           clk_domain=ruby_system.clk_domain,
                                           clk_divider=3)
 
-    for i in xrange(options.num_dirs):
-        #
-        # Create the Ruby objects associated with the directory controller
-        #
-
-        mem_cntrl = RubyMemoryControl(
-                              clk_domain = ruby_system.memctrl_clk_domain,
-                              version = i,
-                              ruby_system = ruby_system)
-
-        dir_size = MemorySize('0B')
-        dir_size.value = mem_module_size
-
+    mem_dir_cntrl_nodes, rom_dir_cntrl_node = create_directories(
+        options, bootmem, ruby_system, system)
+    dir_cntrl_nodes = mem_dir_cntrl_nodes[:]
+    if rom_dir_cntrl_node is not None:
+        dir_cntrl_nodes.append(rom_dir_cntrl_node)
+    for dir_cntrl in dir_cntrl_nodes:
         pf = ProbeFilter(size = pf_size, assoc = 4,
                          start_index_bit = pf_start_bit)
 
-        dir_cntrl = Directory_Controller(version = i,
-                                         cntrl_id = cntrl_count,
-                                         directory = \
-                                         RubyDirectoryMemory( \
-                                                    version = i,
-                                                    size = dir_size,
-                                                    use_map = options.use_map,
-                                                    map_levels = \
-                                                    options.map_levels,
-                                                    numa_high_bit = \
-                                                      options.numa_high_bit),
-                                         probeFilter = pf,
-                                         memBuffer = mem_cntrl,
-                                         probe_filter_enabled = options.pf_on,
-                                         full_bit_dir_enabled = options.dir_on,
-                                         ruby_system = ruby_system)
+        dir_cntrl.probeFilter = pf
+        dir_cntrl.probe_filter_enabled = options.pf_on
+        dir_cntrl.full_bit_dir_enabled = options.dir_on
 
         if options.recycle_latency:
             dir_cntrl.recycle_latency = options.recycle_latency
 
-        exec("ruby_system.dir_cntrl%d = dir_cntrl" % i)
-        dir_cntrl_nodes.append(dir_cntrl)
+        # Connect the directory controller to the network
+        dir_cntrl.forwardFromDir = MessageBuffer()
+        dir_cntrl.forwardFromDir.master = ruby_system.network.slave
+        dir_cntrl.responseFromDir = MessageBuffer()
+        dir_cntrl.responseFromDir.master = ruby_system.network.slave
+        dir_cntrl.dmaResponseFromDir = MessageBuffer(ordered = True)
+        dir_cntrl.dmaResponseFromDir.master = ruby_system.network.slave
+
+        dir_cntrl.triggerQueue = MessageBuffer(ordered = True)
+
+        dir_cntrl.unblockToDir = MessageBuffer()
+        dir_cntrl.unblockToDir.slave = ruby_system.network.master
+        dir_cntrl.responseToDir = MessageBuffer()
+        dir_cntrl.responseToDir.slave = ruby_system.network.master
+        dir_cntrl.requestToDir = MessageBuffer()
+        dir_cntrl.requestToDir.slave = ruby_system.network.master
+        dir_cntrl.dmaRequestToDir = MessageBuffer(ordered = True)
+        dir_cntrl.dmaRequestToDir.slave = ruby_system.network.master
+        dir_cntrl.requestToMemory = MessageBuffer()
+        dir_cntrl.responseFromMemory = MessageBuffer()
 
-        cntrl_count += 1
 
     for i, dma_port in enumerate(dma_ports):
         #
         # Create the Ruby objects associated with the dma controller
         #
         dma_seq = DMASequencer(version = i,
-                               ruby_system = ruby_system)
-        
+                               ruby_system = ruby_system,
+                               slave = dma_port)
+
         dma_cntrl = DMA_Controller(version = i,
-                                   cntrl_id = cntrl_count,
                                    dma_sequencer = dma_seq,
+                                   transitions_per_cycle = options.ports,
                                    ruby_system = ruby_system)
 
         exec("ruby_system.dma_cntrl%d = dma_cntrl" % i)
-        exec("ruby_system.dma_cntrl%d.dma_sequencer.slave = dma_port" % i)
         dma_cntrl_nodes.append(dma_cntrl)
 
         if options.recycle_latency:
             dma_cntrl.recycle_latency = options.recycle_latency
 
-        cntrl_count += 1
+        # Connect the dma controller to the network
+        dma_cntrl.responseFromDir = MessageBuffer(ordered = True)
+        dma_cntrl.responseFromDir.slave = ruby_system.network.master
+        dma_cntrl.requestToDir = MessageBuffer()
+        dma_cntrl.requestToDir.master = ruby_system.network.slave
+        dma_cntrl.mandatoryQueue = MessageBuffer()
 
     all_cntrls = l1_cntrl_nodes + dir_cntrl_nodes + dma_cntrl_nodes
 
+    # Create the io controller and the sequencer
+    if full_system:
+        io_seq = DMASequencer(version=len(dma_ports), ruby_system=ruby_system)
+        ruby_system._io_port = io_seq
+        io_controller = DMA_Controller(version = len(dma_ports),
+                                       dma_sequencer = io_seq,
+                                       ruby_system = ruby_system)
+        ruby_system.io_controller = io_controller
+
+        # Connect the dma controller to the network
+        io_controller.responseFromDir = MessageBuffer(ordered = True)
+        io_controller.responseFromDir.slave = ruby_system.network.master
+        io_controller.requestToDir = MessageBuffer()
+        io_controller.requestToDir.master = ruby_system.network.slave
+        io_controller.mandatoryQueue = MessageBuffer()
+
+        all_cntrls = all_cntrls + [io_controller]
+    # Register configuration with filesystem
+    else:
+        for i in xrange(options.num_cpus):
+            FileSystemConfig.register_cpu(physical_package_id = 0,
+                                          core_siblings = [],
+                                          core_id = i,
+                                          thread_siblings = [])
+
+            FileSystemConfig.register_cache(level = 1,
+                                            idu_type = 'Instruction',
+                                            size = options.l1i_size,
+                                            line_size = options.cacheline_size,
+                                            assoc = options.l1i_assoc,
+                                            cpus = [i])
+            FileSystemConfig.register_cache(level = 1,
+                                            idu_type = 'Data',
+                                            size = options.l1d_size,
+                                            line_size = options.cacheline_size,
+                                            assoc = options.l1d_assoc,
+                                            cpus = [i])
+
+            FileSystemConfig.register_cache(level = 2,
+                                            idu_type = 'Unified',
+                                            size = options.l2_size,
+                                            line_size = options.cacheline_size,
+                                            assoc = options.l2_assoc,
+                                            cpus = [i])
+
+    ruby_system.network.number_of_virtual_networks = 6
     topology = create_topology(all_cntrls, options)
-
-    return (cpu_sequencers, dir_cntrl_nodes, topology)
+    return (cpu_sequencers, mem_dir_cntrl_nodes, topology)