From 3ba131f4d50e17170531ea69bd1d3733f498e381 Mon Sep 17 00:00:00 2001
From: Andreas Hansson <andreas.hansson@arm.com>
Date: Fri, 1 Mar 2013 13:20:22 -0500
Subject: [PATCH] mem: Add support for multi-channel DRAM configurations

This patch adds support for multi-channel instances of the DRAM
controller model by stripping away the channel bits in the address
decoding. The patch relies on the availiability of address
interleaving and, at this time, it is up to the user to configure the
interleaving appropriately. At the moment it is assumed that the
channel interleaving bits are immediately following the column bits
(smallest sensible interleaving). Convenience methods for building
multi-channel configurations will be added later.
---
 src/mem/SimpleDRAM.py  |  4 ++++
 src/mem/simple_dram.cc | 29 ++++++++++++++++++++++++++++-
 src/mem/simple_dram.hh |  1 +
 3 files changed, 33 insertions(+), 1 deletion(-)

diff --git a/src/mem/SimpleDRAM.py b/src/mem/SimpleDRAM.py
index 41bad9356..0e43a6a39 100644
--- a/src/mem/SimpleDRAM.py
+++ b/src/mem/SimpleDRAM.py
@@ -80,6 +80,10 @@ class SimpleDRAM(AbstractMemory):
     lines_per_rowbuffer = Param.Unsigned("Row buffer size in cache lines")
     ranks_per_channel = Param.Unsigned("Number of ranks per channel")
     banks_per_rank = Param.Unsigned("Number of banks per rank")
+    # only used for the address mapping as the controller by
+    # construction is a single channel and multiple controllers have
+    # to be instantiated for a multi-channel configuration
+    channels = Param.Unsigned(1, "Number of channels")
 
     # timing behaviour and constraints - all in nanoseconds
 
diff --git a/src/mem/simple_dram.cc b/src/mem/simple_dram.cc
index d822fbeff..ba5345c3f 100644
--- a/src/mem/simple_dram.cc
+++ b/src/mem/simple_dram.cc
@@ -57,7 +57,7 @@ SimpleDRAM::SimpleDRAM(const SimpleDRAMParams* p) :
     bytesPerCacheLine(0),
     linesPerRowBuffer(p->lines_per_rowbuffer),
     ranksPerChannel(p->ranks_per_channel),
-    banksPerRank(p->banks_per_rank), rowsPerBank(0),
+    banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0),
     readBufferSize(p->read_buffer_size),
     writeBufferSize(p->write_buffer_size),
     writeThresholdPerc(p->write_thresh_perc),
@@ -115,6 +115,23 @@ SimpleDRAM::init()
     rowsPerBank = capacity / (bytesPerCacheLine * linesPerRowBuffer *
                               banksPerRank * ranksPerChannel);
 
+    if (range.interleaved()) {
+        if (channels != range.stripes())
+            panic("%s has %d interleaved address stripes but %d channel(s)\n",
+                  name(), range.stripes(), channels);
+
+        if (addrMapping == Enums::openmap) {
+            if (bytesPerCacheLine * linesPerRowBuffer !=
+                range.granularity()) {
+                panic("Interleaving of %s doesn't match open address map\n",
+                      name());
+            }
+        } else if (addrMapping == Enums::closemap) {
+            if (bytesPerCacheLine != range.granularity())
+                panic("Interleaving of %s doesn't match closed address map\n",
+                      name());
+        }
+    }
 }
 
 void
@@ -190,6 +207,11 @@ SimpleDRAM::decodeAddr(PacketPtr pkt)
         // sequential cache lines occupy the same row
         addr = addr / linesPerRowBuffer;
 
+        // take out the channel part of the address, note that this has
+        // to match with how accesses are interleaved between the
+        // controllers in the address mapping
+        addr = addr / channels;
+
         // after the column bits, we get the bank bits to interleave
         // over the banks
         bank = addr % banksPerRank;
@@ -207,6 +229,11 @@ SimpleDRAM::decodeAddr(PacketPtr pkt)
         // optimise for closed page mode and utilise maximum
         // parallelism of the DRAM (at the cost of power)
 
+        // take out the channel part of the address, not that this has
+        // to match with how accesses are interleaved between the
+        // controllers in the address mapping
+        addr = addr / channels;
+
         // start with the bank bits, as this provides the maximum
         // opportunity for parallelism between requests
         bank = addr % banksPerRank;
diff --git a/src/mem/simple_dram.hh b/src/mem/simple_dram.hh
index d8f51a745..1f6e1a837 100644
--- a/src/mem/simple_dram.hh
+++ b/src/mem/simple_dram.hh
@@ -378,6 +378,7 @@ class SimpleDRAM : public AbstractMemory
     const uint32_t linesPerRowBuffer;
     const uint32_t ranksPerChannel;
     const uint32_t banksPerRank;
+    const uint32_t channels;
     uint32_t rowsPerBank;
     const uint32_t readBufferSize;
     const uint32_t writeBufferSize;