diff -r b41f18136264 -r 08c7971d5f0a src/mem/dram_ctrl.hh --- a/src/mem/dram_ctrl.hh Wed Feb 05 16:17:41 2014 -0600 +++ b/src/mem/dram_ctrl.hh Mon Aug 18 13:32:55 2014 +0100 @@ -453,6 +453,7 @@ const uint32_t burstSize; const uint32_t rowBufferSize; const uint32_t columnsPerRowBuffer; + const uint32_t columnsPerStripe; const uint32_t ranksPerChannel; const uint32_t banksPerRank; const uint32_t channels; @@ -469,7 +470,7 @@ * Basic memory timing parameters initialized based on parameter * values. */ - const Tick tCK; + const Tick M5_CLASS_VAR_USED tCK; const Tick tWTR; const Tick tRTW; const Tick tBURST; diff -r b41f18136264 -r 08c7971d5f0a src/mem/dram_ctrl.cc --- a/src/mem/dram_ctrl.cc Wed Feb 05 16:17:41 2014 -0600 +++ b/src/mem/dram_ctrl.cc Mon Aug 18 13:32:55 2014 +0100 @@ -67,6 +67,7 @@ burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8), rowBufferSize(devicesPerRank * deviceRowBufferSize), columnsPerRowBuffer(rowBufferSize / burstSize), + columnsPerStripe(range.granularity() / burstSize), ranksPerChannel(p->ranks_per_channel), banksPerRank(p->banks_per_rank), channels(p->channels), rowsPerBank(0), readBufferSize(p->read_buffer_size), @@ -122,6 +123,7 @@ rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel); + // a bit of sanity checks on the interleaving if (range.interleaved()) { if (channels != range.stripes()) fatal("%s has %d interleaved address stripes but %d channel(s)\n", @@ -132,15 +134,31 @@ fatal("Interleaving of %s doesn't match RoRaBaChCo " "address map\n", name()); } - } else if (addrMapping == Enums::RoRaBaCoCh) { - if (system()->cacheLineSize() != range.granularity()) { - fatal("Interleaving of %s doesn't match RoRaBaCoCh " - "address map\n", name()); + } else if (addrMapping == Enums::RoRaBaCoCh || + addrMapping == Enums::RoCoRaBaCh) { + // for the interleavings with channel bits in the bottom, + // if the system uses a striping granularity that is + // larger than the DRAM burst size, then map the + // sequential accesses within a stripe to a number of + // columns in the DRAM, effectively placing some of the + // lower-order column bits as the least-significant bits + // of the address (above the ones denoting the burst size) + assert(columnsPerStripe >= 1); + + // striping has to be done at a granularity that is equal + // or larger to a cache line + if (system()->cacheLineSize() > range.granularity()) { + fatal("Interleaving of %s must be at least as large as the " + "cache line size\n", name()); } - } else if (addrMapping == Enums::RoCoRaBaCh) { - if (system()->cacheLineSize() != range.granularity()) - fatal("Interleaving of %s doesn't match RoCoRaBaCh " - "address map\n", name()); + + // ...and equal or smaller than the row-buffer size + if (rowBufferSize < range.granularity()) { + fatal("Interleaving of %s must be at most as large as the " + "row-buffer size\n", name()); + } + // this is essentially the check above, so just to be sure + assert(columnsPerStripe <= columnsPerRowBuffer); } } @@ -228,7 +246,8 @@ // always the top bits, and check before creating the DRAMPacket uint64_t row; - // truncate the address to the access granularity + // truncate the address to a DRAM burst, which makes it unique to + // a specific column, row, bank, rank and channel Addr addr = dramPktAddr / burstSize; // we have removed the lowest order address bits that denote the @@ -255,11 +274,14 @@ row = addr % rowsPerBank; addr = addr / rowsPerBank; } else if (addrMapping == Enums::RoRaBaCoCh) { + // take out the lower-order column bits + addr = addr / columnsPerStripe; + // take out the channel part of the address addr = addr / channels; - // next, the column - addr = addr / columnsPerRowBuffer; + // next, the higher-order column bites + addr = addr / (columnsPerRowBuffer / columnsPerStripe); // after the column bits, we get the bank bits to interleave // over the banks @@ -278,6 +300,9 @@ // optimise for closed page mode and utilise maximum // parallelism of the DRAM (at the cost of power) + // take out the lower-order column bits + addr = addr / columnsPerStripe; + // 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 @@ -292,9 +317,8 @@ rank = addr % ranksPerChannel; addr = addr / ranksPerChannel; - // next the column bits which we do not need to keep track of - // and simply skip past - addr = addr / columnsPerRowBuffer; + // next, the higher-order column bites + addr = addr / (columnsPerRowBuffer / columnsPerStripe); // lastly, get the row bits row = addr % rowsPerBank;