From 0deef376d96bfe0a3a2496714ac22471d9ee818a Mon Sep 17 00:00:00 2001 From: Yasuko Eckert Date: Tue, 3 Jun 2014 13:32:59 -0700 Subject: ext: McPAT interface changes and fixes This patch includes software engineering changes and some generic bug fixes Joel Hestness and Yasuko Eckert made to McPAT 0.8. There are still known issues/concernts we did not have a chance to address in this patch. High-level changes in this patch include: 1) Making XML parsing modular and hierarchical: - Shift parsing responsibility into the components - Read XML in a (mostly) context-free recursive manner so that McPAT input files can contain arbitrary component hierarchies 2) Making power, energy, and area calculations a hierarchical and recursive process - Components track their subcomponents and recursively call compute functions in stages - Make C++ object hierarchy reflect inheritance of classes of components with similar structures - Simplify computeArea() and computeEnergy() functions to eliminate successive calls to calculate separate TDP vs. runtime energy - Remove Processor component (now unnecessary) and introduce a more abstract System component 3) Standardizing McPAT output across all components - Use a single, common data structure for storing and printing McPAT output - Recursively call print functions through component hierarchy 4) For caches, allow splitting data array and tag array reads and writes for better accuracy 5) Improving the usability of CACTI by printing more helpful warning and error messages 6) Minor: Impose more rigorous code style for clarity (more work still to be done) Overall, these changes greatly reduce the amount of replicated code, and they improve McPAT runtime and decrease memory footprint. --- ext/mcpat/memoryctrl.cc | 1125 +++++++++++++++++++++-------------------------- 1 file changed, 501 insertions(+), 624 deletions(-) (limited to 'ext/mcpat/memoryctrl.cc') diff --git a/ext/mcpat/memoryctrl.cc b/ext/mcpat/memoryctrl.cc index ae3bc75ec..dec24512e 100644 --- a/ext/mcpat/memoryctrl.cc +++ b/ext/mcpat/memoryctrl.cc @@ -2,6 +2,7 @@ * McPAT * SOFTWARE LICENSE AGREEMENT * Copyright 2012 Hewlett-Packard Development Company, L.P. + * Copyright (c) 2010-2013 Advanced Micro Devices, Inc. * All Rights Reserved * * Redistribution and use in source and binary forms, with or without @@ -25,18 +26,19 @@ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * 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.” + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * ***************************************************************************/ + #include #include #include #include #include -#include "XML_Parse.h" #include "basic_circuit.h" #include "basic_components.h" +#include "common.h" #include "const.h" #include "io.h" #include "logic.h" @@ -69,668 +71,543 @@ * */ -MCBackend::MCBackend(InputParameter* interface_ip_, const MCParam & mcp_, enum MemoryCtrl_type mc_type_) -:l_ip(*interface_ip_), - mc_type(mc_type_), - mcp(mcp_) -{ - - local_result = init_interface(&l_ip); - compute(); - +MCBackend::MCBackend(XMLNode* _xml_data, InputParameter* interface_ip_, + const MCParameters & mcp_, const MCStatistics & mcs_) + : McPATComponent(_xml_data), l_ip(*interface_ip_), mcp(mcp_), mcs(mcs_) { + name = "Transaction Engine"; + local_result = init_interface(&l_ip, name); + + // Set up stats for the power calculations + tdp_stats.reset(); + tdp_stats.readAc.access = 0.5 * mcp.num_channels * mcp.clockRate; + tdp_stats.writeAc.access = 0.5 * mcp.num_channels * mcp.clockRate; + rtp_stats.reset(); + rtp_stats.readAc.access = mcs.reads; + rtp_stats.writeAc.access = mcs.writes; } +void MCBackend::computeArea() { + // The area is in nm^2 + if (mcp.mc_type == MC) { + if (mcp.type == 0) { + output_data.area = (2.7927 * log(mcp.peak_transfer_rate * 2) - + 19.862) / 2.0 * mcp.dataBusWidth / 128.0 * + (l_ip.F_sz_um / 0.09) * mcp.num_channels; + } else { + output_data.area = 0.15 * mcp.dataBusWidth / 72.0 * + (l_ip.F_sz_um / 0.065) * (l_ip.F_sz_um / 0.065) * + mcp.num_channels; + } + } else { + //skip old model + cout << "Unknown memory controllers" << endl; + exit(0); + //area based on Cadence ChipEstimator for 8bit bus + output_data.area = 0.243 * mcp.dataBusWidth / 8; + } +} -void MCBackend::compute() -{ - //double max_row_addr_width = 20.0;//Current address 12~18bits - double C_MCB, mc_power, backend_dyn, backend_gates;//, refresh_period,refresh_freq;//Equivalent per bit Cap for backend, - double pmos_to_nmos_sizing_r = pmos_to_nmos_sz_ratio(); - double NMOS_sizing, PMOS_sizing; - - if (mc_type == MC) - { - if (mcp.type == 0) - { - //area = (2.2927*log(peakDataTransferRate)-14.504)*memDataWidth/144.0*(l_ip.F_sz_um/0.09); - area.set_area((2.7927*log(mcp.peakDataTransferRate*2)-19.862)/2.0*mcp.dataBusWidth/128.0*(l_ip.F_sz_um/0.09)*mcp.num_channels*1e6);//um^2 - //assuming the approximately same scaling factor as seen in processors. - //C_MCB=0.2/1.3/1.3/266/64/0.09*g_ip.F_sz_um;//based on AMD Geode processor which has a very basic mc on chip. - //C_MCB = 1.6/200/1e6/144/1.2/1.2*g_ip.F_sz_um/0.19;//Based on Niagara power numbers.The base power (W) is divided by device frequency and vdd and scale to target process. - //mc_power = 0.0291*2;//29.1mW@200MHz @130nm From Power Analysis of SystemLevel OnChip Communication Architectures by Lahiri et - mc_power = 4.32*0.1;//4.32W@1GhzMHz @65nm Cadence ChipEstimator 10% for backend - C_MCB = mc_power/1e9/72/1.1/1.1*l_ip.F_sz_um/0.065; - power_t.readOp.dynamic = C_MCB*g_tp.peri_global.Vdd*g_tp.peri_global.Vdd*(mcp.dataBusWidth/*+mcp.addressBusWidth*/);//per access energy in memory controller - power_t.readOp.leakage = area.get_area()/2 *(g_tp.scaling_factor.core_tx_density)*cmos_Isub_leakage(g_tp.min_w_nmos_, g_tp.min_w_nmos_*pmos_to_nmos_sizing_r, 1, inv)*g_tp.peri_global.Vdd;//unit W - power_t.readOp.gate_leakage = area.get_area()/2 *(g_tp.scaling_factor.core_tx_density)*cmos_Ig_leakage(g_tp.min_w_nmos_, g_tp.min_w_nmos_*pmos_to_nmos_sizing_r, 1, inv)*g_tp.peri_global.Vdd;//unit W +void MCBackend::computeEnergy() { + double C_MCB, mc_power; + double backend_dyn; + double backend_gates; + double pmos_to_nmos_sizing_r = pmos_to_nmos_sz_ratio(); + double NMOS_sizing = g_tp.min_w_nmos_; + double PMOS_sizing = g_tp.min_w_nmos_ * pmos_to_nmos_sizing_r; + double area_um2 = output_data.area * 1e6; + + if (mcp.mc_type == MC) { + if (mcp.type == 0) { + //assuming the approximately same scaling factor as seen in processors. + //C_MCB = 1.6/200/1e6/144/1.2/1.2*g_ip.F_sz_um/0.19;//Based on Niagara power numbers.The base power (W) is divided by device frequency and vdd and scale to target process. + //mc_power = 0.0291*2;//29.1mW@200MHz @130nm From Power Analysis of SystemLevel OnChip Communication Architectures by Lahiri et + mc_power = 4.32*0.1;//4.32W@1GhzMHz @65nm Cadence ChipEstimator 10% for backend + C_MCB = mc_power/1e9/72/1.1/1.1*l_ip.F_sz_um/0.065; + //per access energy in memory controller + power.readOp.dynamic = C_MCB * g_tp.peri_global.Vdd * + g_tp.peri_global.Vdd * + (mcp.dataBusWidth/*+mcp.addressBusWidth*/); + power.readOp.leakage = area_um2 / 2 * + (g_tp.scaling_factor.core_tx_density) * + cmos_Isub_leakage(NMOS_sizing, PMOS_sizing, 1, inv) * + g_tp.peri_global.Vdd;//unit W + power.readOp.gate_leakage = area_um2 / 2 * + (g_tp.scaling_factor.core_tx_density) * + cmos_Ig_leakage(NMOS_sizing, PMOS_sizing, 1, inv) * + g_tp.peri_global.Vdd;//unit W + } else { + //Average on DDR2/3 protocol controller and DDRC 1600/800A in + //Cadence ChipEstimate + backend_dyn = 0.9e-9 / 800e6 * mcp.clockRate / 12800 * + mcp.peak_transfer_rate* mcp.dataBusWidth / 72.0 * + g_tp.peri_global.Vdd / 1.1 * g_tp.peri_global.Vdd / 1.1 * + (l_ip.F_sz_nm/65.0); + //Scaling to technology and DIMM feature. The base IP support + //DDR3-1600(PC3 12800) + //5000 is from Cadence ChipEstimator + backend_gates = 50000 * mcp.dataBusWidth / 64.0; + + power.readOp.dynamic = backend_dyn; + power.readOp.leakage = (backend_gates) * + cmos_Isub_leakage(NMOS_sizing, PMOS_sizing, 2, nand) * + g_tp.peri_global.Vdd;//unit W + power.readOp.gate_leakage = (backend_gates) * + cmos_Ig_leakage(NMOS_sizing, PMOS_sizing, 2, nand) * + g_tp.peri_global.Vdd;//unit W } - else - { NMOS_sizing = g_tp.min_w_nmos_; - PMOS_sizing = g_tp.min_w_nmos_*pmos_to_nmos_sizing_r; - area.set_area(0.15*mcp.dataBusWidth/72.0*(l_ip.F_sz_um/0.065)* (l_ip.F_sz_um/0.065)*mcp.num_channels*1e6);//um^2 - backend_dyn = 0.9e-9/800e6*mcp.clockRate/12800*mcp.peakDataTransferRate*mcp.dataBusWidth/72.0*g_tp.peri_global.Vdd/1.1*g_tp.peri_global.Vdd/1.1*(l_ip.F_sz_nm/65.0);//Average on DDR2/3 protocol controller and DDRC 1600/800A in Cadence ChipEstimate - //Scaling to technology and DIMM feature. The base IP support DDR3-1600(PC3 12800) - backend_gates = 50000*mcp.dataBusWidth/64.0;//5000 is from Cadence ChipEstimator - - power_t.readOp.dynamic = backend_dyn; - power_t.readOp.leakage = (backend_gates)*cmos_Isub_leakage(NMOS_sizing, PMOS_sizing, 2, nand)*g_tp.peri_global.Vdd;//unit W - power_t.readOp.gate_leakage = (backend_gates)*cmos_Ig_leakage(NMOS_sizing, PMOS_sizing, 2, nand)*g_tp.peri_global.Vdd;//unit W + } else { + //skip old model + cout<<"Unknown memory controllers"< 0; + interface_ip.tag_w = tag; + interface_ip.access_mode = Normal; + interface_ip.obj_func_dyn_energy = 0; + interface_ip.obj_func_dyn_power = 0; + interface_ip.obj_func_leak_power = 0; + interface_ip.obj_func_cycle_t = 1; + interface_ip.num_rw_ports = 0; + interface_ip.num_rd_ports = mcp.num_channels; + interface_ip.num_wr_ports = interface_ip.num_rd_ports; + interface_ip.num_se_rd_ports = 0; + interface_ip.num_search_ports = mcp.num_channels; + interface_ip.is_cache = true; + interface_ip.pure_cam = false; + interface_ip.pure_ram = false; + interface_ip.throughput = 1.0 / mcp.clockRate; + interface_ip.latency = 1.0 / mcp.clockRate; + frontendBuffer = new CacheArray(xml_data, &interface_ip, "Reorder Buffer", + Uncore_device, mcp.clockRate); + children.push_back(frontendBuffer); + + frontendBuffer->tdp_stats.reset(); + frontendBuffer->tdp_stats.readAc.access = + frontendBuffer->l_ip.num_search_ports + + frontendBuffer->l_ip.num_wr_ports; + frontendBuffer->tdp_stats.writeAc.access = + frontendBuffer->l_ip.num_search_ports; + frontendBuffer->tdp_stats.searchAc.access = + frontendBuffer->l_ip.num_wr_ports; + frontendBuffer->rtp_stats.reset(); + // TODO: These stats assume that access power is calculated per buffer + // bit, which requires the stats to take into account the number of + // bits for each buffer slot. This should be revised... + //For each channel, each memory word need to check the address data to + //achieve best scheduling results. + //and this need to be done on all physical DIMMs in each logical memory + //DIMM *mcp.dataBusWidth/72 + frontendBuffer->rtp_stats.readAc.access = mcs.reads * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth * mcp.dataBusWidth / 72; + frontendBuffer->rtp_stats.writeAc.access = mcs.writes * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth * mcp.dataBusWidth / 72; + frontendBuffer->rtp_stats.searchAc.access = + frontendBuffer->rtp_stats.readAc.access + + frontendBuffer->rtp_stats.writeAc.access; + + // Read Buffers + //Support key words first operation + data = (int)ceil(mcp.dataBusWidth / BITS_PER_BYTE); + + interface_ip.cache_sz = data * mcp.IO_buffer_size_per_channel; + interface_ip.line_sz = data; + interface_ip.assoc = mcp.read_buffer_assoc; + interface_ip.nbanks = mcp.read_buffer_nbanks; + interface_ip.out_w = interface_ip.line_sz * BITS_PER_BYTE; + interface_ip.specific_tag = mcp.read_buffer_tag_width > 0; + interface_ip.tag_w = mcp.read_buffer_tag_width; + interface_ip.access_mode = Sequential; + interface_ip.obj_func_dyn_energy = 0; + interface_ip.obj_func_dyn_power = 0; + interface_ip.obj_func_leak_power = 0; + interface_ip.obj_func_cycle_t = 1; + interface_ip.num_rw_ports = 0; + interface_ip.num_rd_ports = mcp.num_channels; + interface_ip.num_wr_ports = interface_ip.num_rd_ports; + interface_ip.num_se_rd_ports = 0; + interface_ip.num_search_ports = 0; + interface_ip.is_cache = false; + interface_ip.pure_cam = false; + interface_ip.pure_ram = true; + interface_ip.throughput = 1.0 / mcp.clockRate; + interface_ip.latency = 1.0 / mcp.clockRate; + readBuffer = new CacheArray(xml_data, &interface_ip, "Read Buffer", + Uncore_device, mcp.clockRate); + children.push_back(readBuffer); + + readBuffer->tdp_stats.reset(); + readBuffer->tdp_stats.readAc.access = readBuffer->l_ip.num_rd_ports * + mcs.duty_cycle; + readBuffer->tdp_stats.writeAc.access = readBuffer->l_ip.num_wr_ports * + mcs.duty_cycle; + readBuffer->rtp_stats.reset(); + readBuffer->rtp_stats.readAc.access = mcs.reads * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth; + readBuffer->rtp_stats.writeAc.access = mcs.reads * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth; + + // Write Buffer + //Support key words first operation + data = (int)ceil(mcp.dataBusWidth / BITS_PER_BYTE); + + interface_ip.cache_sz = data * mcp.IO_buffer_size_per_channel; + interface_ip.line_sz = data; + interface_ip.assoc = mcp.write_buffer_assoc; + interface_ip.nbanks = mcp.write_buffer_nbanks; + interface_ip.out_w = interface_ip.line_sz * BITS_PER_BYTE; + interface_ip.specific_tag = mcp.write_buffer_tag_width > 0; + interface_ip.tag_w = mcp.write_buffer_tag_width; + interface_ip.access_mode = Normal; + interface_ip.obj_func_dyn_energy = 0; + interface_ip.obj_func_dyn_power = 0; + interface_ip.obj_func_leak_power = 0; + interface_ip.obj_func_cycle_t = 1; + interface_ip.num_rw_ports = 0; + interface_ip.num_rd_ports = mcp.num_channels; + interface_ip.num_wr_ports = interface_ip.num_rd_ports; + interface_ip.num_se_rd_ports = 0; + interface_ip.num_search_ports = 0; + interface_ip.is_cache = false; + interface_ip.pure_cam = false; + interface_ip.pure_ram = true; + interface_ip.throughput = 1.0 / mcp.clockRate; + interface_ip.latency = 1.0 / mcp.clockRate; + writeBuffer = new CacheArray(xml_data, &interface_ip, "Write Buffer", + Uncore_device, mcp.clockRate); + children.push_back(writeBuffer); + + writeBuffer->tdp_stats.reset(); + writeBuffer->tdp_stats.readAc.access = writeBuffer->l_ip.num_rd_ports * + mcs.duty_cycle; + writeBuffer->tdp_stats.writeAc.access = writeBuffer->l_ip.num_wr_ports * + mcs.duty_cycle; + writeBuffer->rtp_stats.reset(); + writeBuffer->rtp_stats.readAc.access = mcs.reads * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth; + writeBuffer->rtp_stats.writeAc.access = mcs.writes * mcp.llcBlockSize * + BITS_PER_BYTE / mcp.dataBusWidth; + + // TODO: Set up selection logic as a leaf node in tree + //selection and arbitration logic + MC_arb = + new selection_logic(xml_data, is_default, + mcp.req_window_size_per_channel, 1, &interface_ip, + "Arbitration Logic", (mcs.reads + mcs.writes), + mcp.clockRate, Uncore_device); + // MC_arb is not included in the roll-up due to the uninitialized area + //children.push_back(MC_arb); +} -void MCPHY::computeEnergy(bool is_tdp) -{ - if (is_tdp) - { - //init stats for Peak - stats_t.readAc.access = 0.5*mcp.num_channels; //time share on buses - stats_t.writeAc.access = 0.5*mcp.num_channels; - tdp_stats = stats_t; - } - else - { - //init stats for runtime power (RTP) - stats_t.readAc.access = mcp.reads; - stats_t.writeAc.access = mcp.writes; - tdp_stats = stats_t; - } - - if (is_tdp) - { - double data_transfer_unit = (mc_type == MC)? 72:16;/*DIMM data width*/ - power = power_t; - power.readOp.dynamic = power.readOp.dynamic * (mcp.peakDataTransferRate*8*1e6/1e9/*change to Gbs*/)*mcp.dataBusWidth/data_transfer_unit*mcp.num_channels/mcp.clockRate; - // divide by clock rate is for match the final computation where *clock is used - //(stats_t.readAc.access*power_t.readOp.dynamic+ -// stats_t.writeAc.access*power_t.readOp.dynamic); - - } - else - { - rt_power = power_t; -// rt_power.readOp.dynamic = (stats_t.readAc.access*power_t.readOp.dynamic+ -// stats_t.writeAc.access*power_t.readOp.dynamic); - - rt_power.readOp.dynamic=power_t.readOp.dynamic*(stats_t.readAc.access + stats_t.writeAc.access)*(mcp.llcBlockSize)*8/1e9/mcp.executionTime*(mcp.executionTime); - rt_power.readOp.dynamic = rt_power.readOp.dynamic + power.readOp.dynamic*0.1*mcp.clockRate*mcp.num_mcs*mcp.executionTime; +MemoryController::MemoryController(XMLNode* _xml_data, + InputParameter* interface_ip_) + : McPATComponent(_xml_data), interface_ip(*interface_ip_) { + name = "Memory Controller"; + set_mc_param(); + // TODO: Pass params and stats as pointers + children.push_back(new MCFrontEnd(xml_data, &interface_ip, mcp, mcs)); + children.push_back(new MCBackend(xml_data, &interface_ip, mcp, mcs)); + + if (mcp.type==0 || (mcp.type == 1 && mcp.withPHY)) { + children.push_back(new MCPHY(xml_data, &interface_ip, mcp, mcs)); } } -MCFrontEnd::MCFrontEnd(ParseXML *XML_interface,InputParameter* interface_ip_, const MCParam & mcp_, enum MemoryCtrl_type mc_type_) -:XML(XML_interface), - interface_ip(*interface_ip_), - mc_type(mc_type_), - mcp(mcp_), - MC_arb(0), - frontendBuffer(0), - readBuffer(0), - writeBuffer(0) -{ - /* All computations are for a single MC - * - */ - - int tag, data; - bool is_default =true;//indication for default setup - - /* MC frontend engine channels share the same engines but logically partitioned - * For all hardware inside MC. different channels do not share resources. - * TODO: add docodeing/mux stage to steer memory requests to different channels. - */ - - //memory request reorder buffer - tag = mcp.addressBusWidth + EXTRA_TAG_BITS + mcp.opcodeW; - data = int(ceil((XML->sys.physical_address_width + mcp.opcodeW)/8.0)); - interface_ip.cache_sz = data*XML->sys.mc.req_window_size_per_channel; - interface_ip.line_sz = data; - interface_ip.assoc = 0; - interface_ip.nbanks = 1; - interface_ip.out_w = interface_ip.line_sz*8; - interface_ip.specific_tag = 1; - interface_ip.tag_w = tag; - interface_ip.access_mode = 0; - interface_ip.throughput = 1.0/mcp.clockRate; - interface_ip.latency = 1.0/mcp.clockRate; - interface_ip.is_cache = true; - interface_ip.pure_cam = false; - interface_ip.pure_ram = false; - interface_ip.obj_func_dyn_energy = 0; - interface_ip.obj_func_dyn_power = 0; - interface_ip.obj_func_leak_power = 0; - interface_ip.obj_func_cycle_t = 1; - interface_ip.num_rw_ports = 0; - interface_ip.num_rd_ports = XML->sys.mc.memory_channels_per_mc; - interface_ip.num_wr_ports = interface_ip.num_rd_ports; - interface_ip.num_se_rd_ports = 0; - interface_ip.num_search_ports = XML->sys.mc.memory_channels_per_mc; - frontendBuffer = new ArrayST(&interface_ip, "MC ReorderBuffer", Uncore_device); - frontendBuffer->area.set_area(frontendBuffer->area.get_area()+ frontendBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); - area.set_area(area.get_area()+ frontendBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); - - //selection and arbitration logic - MC_arb = new selection_logic(is_default, XML->sys.mc.req_window_size_per_channel,1,&interface_ip, Uncore_device); - - //read buffers. - data = (int)ceil(mcp.dataBusWidth/8.0);//Support key words first operation //8 means converting bit to Byte - interface_ip.cache_sz = data*XML->sys.mc.IO_buffer_size_per_channel;//*llcBlockSize; - interface_ip.line_sz = data; - interface_ip.assoc = 1; - interface_ip.nbanks = 1; - interface_ip.out_w = interface_ip.line_sz*8; - interface_ip.access_mode = 1; - interface_ip.throughput = 1.0/mcp.clockRate; - interface_ip.latency = 1.0/mcp.clockRate; - interface_ip.is_cache = false; - interface_ip.pure_cam = false; - interface_ip.pure_ram = true; - interface_ip.obj_func_dyn_energy = 0; - interface_ip.obj_func_dyn_power = 0; - interface_ip.obj_func_leak_power = 0; - interface_ip.obj_func_cycle_t = 1; - interface_ip.num_rw_ports = 0;//XML->sys.mc.memory_channels_per_mc*2>2?2:XML->sys.mc.memory_channels_per_mc*2; - interface_ip.num_rd_ports = XML->sys.mc.memory_channels_per_mc; - interface_ip.num_wr_ports = interface_ip.num_rd_ports; - interface_ip.num_se_rd_ports = 0; - readBuffer = new ArrayST(&interface_ip, "MC ReadBuffer", Uncore_device); - readBuffer->area.set_area(readBuffer->area.get_area()+ readBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); - area.set_area(area.get_area()+ readBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); - - //write buffer - data = (int)ceil(mcp.dataBusWidth/8.0);//Support key words first operation //8 means converting bit to Byte - interface_ip.cache_sz = data*XML->sys.mc.IO_buffer_size_per_channel;//*llcBlockSize; - interface_ip.line_sz = data; - interface_ip.assoc = 1; - interface_ip.nbanks = 1; - interface_ip.out_w = interface_ip.line_sz*8; - interface_ip.access_mode = 0; - interface_ip.throughput = 1.0/mcp.clockRate; - interface_ip.latency = 1.0/mcp.clockRate; - interface_ip.obj_func_dyn_energy = 0; - interface_ip.obj_func_dyn_power = 0; - interface_ip.obj_func_leak_power = 0; - interface_ip.obj_func_cycle_t = 1; - interface_ip.num_rw_ports = 0; - interface_ip.num_rd_ports = XML->sys.mc.memory_channels_per_mc; - interface_ip.num_wr_ports = interface_ip.num_rd_ports; - interface_ip.num_se_rd_ports = 0; - writeBuffer = new ArrayST(&interface_ip, "MC writeBuffer", Uncore_device); - writeBuffer->area.set_area(writeBuffer->area.get_area()+ writeBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); - area.set_area(area.get_area()+ writeBuffer->local_result.area*XML->sys.mc.memory_channels_per_mc); +void MemoryController::initialize_params() { + memset(&mcp, 0, sizeof(MCParameters)); } -void MCFrontEnd::computeEnergy(bool is_tdp) -{ - if (is_tdp) - { - //init stats for Peak - frontendBuffer->stats_t.readAc.access = frontendBuffer->l_ip.num_search_ports; - frontendBuffer->stats_t.writeAc.access = frontendBuffer->l_ip.num_wr_ports; - frontendBuffer->tdp_stats = frontendBuffer->stats_t; - - readBuffer->stats_t.readAc.access = readBuffer->l_ip.num_rd_ports*mcp.frontend_duty_cycle; - readBuffer->stats_t.writeAc.access = readBuffer->l_ip.num_wr_ports*mcp.frontend_duty_cycle; - readBuffer->tdp_stats = readBuffer->stats_t; - - writeBuffer->stats_t.readAc.access = writeBuffer->l_ip.num_rd_ports*mcp.frontend_duty_cycle; - writeBuffer->stats_t.writeAc.access = writeBuffer->l_ip.num_wr_ports*mcp.frontend_duty_cycle; - writeBuffer->tdp_stats = writeBuffer->stats_t; - - } - else - { - //init stats for runtime power (RTP) - frontendBuffer->stats_t.readAc.access = XML->sys.mc.memory_reads *mcp.llcBlockSize*8.0/mcp.dataBusWidth*mcp.dataBusWidth/72; - //For each channel, each memory word need to check the address data to achieve best scheduling results. - //and this need to be done on all physical DIMMs in each logical memory DIMM *mcp.dataBusWidth/72 - frontendBuffer->stats_t.writeAc.access = XML->sys.mc.memory_writes*mcp.llcBlockSize*8.0/mcp.dataBusWidth*mcp.dataBusWidth/72; - frontendBuffer->rtp_stats = frontendBuffer->stats_t; - - readBuffer->stats_t.readAc.access = XML->sys.mc.memory_reads*mcp.llcBlockSize*8.0/mcp.dataBusWidth;//support key word first - readBuffer->stats_t.writeAc.access = XML->sys.mc.memory_reads*mcp.llcBlockSize*8.0/mcp.dataBusWidth;//support key word first - readBuffer->rtp_stats = readBuffer->stats_t; - - writeBuffer->stats_t.readAc.access = XML->sys.mc.memory_writes*mcp.llcBlockSize*8.0/mcp.dataBusWidth; - writeBuffer->stats_t.writeAc.access = XML->sys.mc.memory_writes*mcp.llcBlockSize*8.0/mcp.dataBusWidth; - writeBuffer->rtp_stats = writeBuffer->stats_t; - } - - frontendBuffer->power_t.reset(); - readBuffer->power_t.reset(); - writeBuffer->power_t.reset(); - -// frontendBuffer->power_t.readOp.dynamic += (frontendBuffer->stats_t.readAc.access* -// (frontendBuffer->local_result.power.searchOp.dynamic+frontendBuffer->local_result.power.readOp.dynamic)+ -// frontendBuffer->stats_t.writeAc.access*frontendBuffer->local_result.power.writeOp.dynamic); - - frontendBuffer->power_t.readOp.dynamic += (frontendBuffer->stats_t.readAc.access + - frontendBuffer->stats_t.writeAc.access)*frontendBuffer->local_result.power.searchOp.dynamic - + frontendBuffer->stats_t.readAc.access * frontendBuffer->local_result.power.readOp.dynamic - + frontendBuffer->stats_t.writeAc.access*frontendBuffer->local_result.power.writeOp.dynamic; - - readBuffer->power_t.readOp.dynamic += (readBuffer->stats_t.readAc.access* - readBuffer->local_result.power.readOp.dynamic+ - readBuffer->stats_t.writeAc.access*readBuffer->local_result.power.writeOp.dynamic); - writeBuffer->power_t.readOp.dynamic += (writeBuffer->stats_t.readAc.access* - writeBuffer->local_result.power.readOp.dynamic+ - writeBuffer->stats_t.writeAc.access*writeBuffer->local_result.power.writeOp.dynamic); - - if (is_tdp) - { - power = power + frontendBuffer->power_t + readBuffer->power_t + writeBuffer->power_t + - (frontendBuffer->local_result.power + - readBuffer->local_result.power + - writeBuffer->local_result.power)*pppm_lkg; - +void MemoryController::set_mc_param() { + initialize_params(); + + int num_children = xml_data->nChildNode("param"); + int tech_type; + int mat_type; + int i; + for (i = 0; i < num_children; i++) { + XMLNode* paramNode = xml_data->getChildNodePtr("param", &i); + XMLCSTR node_name = paramNode->getAttribute("name"); + XMLCSTR value = paramNode->getAttribute("value"); + + if (!node_name) + warnMissingParamName(paramNode->getAttribute("id")); + + ASSIGN_FP_IF("mc_clock", mcp.clockRate); + ASSIGN_INT_IF("tech_type", tech_type); + ASSIGN_ENUM_IF("mc_type", mcp.mc_type, MemoryCtrl_type); + ASSIGN_FP_IF("num_mcs", mcp.num_mcs); + ASSIGN_INT_IF("llc_line_length", mcp.llc_line_length); + ASSIGN_INT_IF("databus_width", mcp.databus_width); + ASSIGN_INT_IF("memory_channels_per_mc", mcp.num_channels); + ASSIGN_INT_IF("req_window_size_per_channel", + mcp.req_window_size_per_channel); + ASSIGN_INT_IF("IO_buffer_size_per_channel", + mcp.IO_buffer_size_per_channel); + ASSIGN_INT_IF("addressbus_width", mcp.addressbus_width); + ASSIGN_INT_IF("opcode_width", mcp.opcodeW); + ASSIGN_INT_IF("type", mcp.type); + ASSIGN_ENUM_IF("LVDS", mcp.LVDS, bool); + ASSIGN_ENUM_IF("withPHY", mcp.withPHY, bool); + ASSIGN_INT_IF("peak_transfer_rate", mcp.peak_transfer_rate); + ASSIGN_INT_IF("number_ranks", mcp.number_ranks); + ASSIGN_INT_IF("reorder_buffer_assoc", mcp.reorder_buffer_assoc); + ASSIGN_INT_IF("reorder_buffer_nbanks", mcp.reorder_buffer_nbanks); + ASSIGN_INT_IF("read_buffer_assoc", mcp.read_buffer_assoc); + ASSIGN_INT_IF("read_buffer_nbanks", mcp.read_buffer_nbanks); + ASSIGN_INT_IF("read_buffer_tag_width", mcp.read_buffer_tag_width); + ASSIGN_INT_IF("write_buffer_assoc", mcp.write_buffer_assoc); + ASSIGN_INT_IF("write_buffer_nbanks", mcp.write_buffer_nbanks); + ASSIGN_INT_IF("write_buffer_tag_width", mcp.write_buffer_tag_width); + ASSIGN_INT_IF("wire_mat_type", mat_type); + ASSIGN_ENUM_IF("wire_type", interface_ip.wt, Wire_type); + + else { + warnUnrecognizedParam(node_name); + } } - else - { - rt_power = rt_power + frontendBuffer->power_t + readBuffer->power_t + writeBuffer->power_t + - (frontendBuffer->local_result.power + - readBuffer->local_result.power + - writeBuffer->local_result.power)*pppm_lkg; - rt_power.readOp.dynamic = rt_power.readOp.dynamic + power.readOp.dynamic*0.1*mcp.clockRate*mcp.num_mcs*mcp.executionTime; + + if (mcp.mc_type != MC) { + cout << "Unknown memory controller type: Only DRAM controller is " + << "supported for now" << endl; + exit(0); } -} -void MCFrontEnd::displayEnergy(uint32_t indent,int plevel,bool is_tdp) -{ - string indent_str(indent, ' '); - string indent_str_next(indent+2, ' '); - - if (is_tdp) - { - cout << indent_str << "Front End ROB:" << endl; - cout << indent_str_next << "Area = " << frontendBuffer->area.get_area()*1e-6<< " mm^2" << endl; - cout << indent_str_next << "Peak Dynamic = " << frontendBuffer->power.readOp.dynamic*mcp.clockRate << " W" << endl; - cout << indent_str_next << "Subthreshold Leakage = " << frontendBuffer->power.readOp.leakage <<" W" << endl; - cout << indent_str_next << "Gate Leakage = " << frontendBuffer->power.readOp.gate_leakage << " W" << endl; - cout << indent_str_next << "Runtime Dynamic = " << frontendBuffer->rt_power.readOp.dynamic/mcp.executionTime << " W" << endl; - - cout <nChildNode("stat"); + for (i = 0; i < num_children; i++) { + XMLNode* statNode = xml_data->getChildNodePtr("stat", &i); + XMLCSTR node_name = statNode->getAttribute("name"); + XMLCSTR value = statNode->getAttribute("value"); -MemoryController::MemoryController(ParseXML *XML_interface,InputParameter* interface_ip_, enum MemoryCtrl_type mc_type_) -:XML(XML_interface), - interface_ip(*interface_ip_), - mc_type(mc_type_), - frontend(0), - transecEngine(0), - PHY(0), - pipeLogic(0) -{ - /* All computations are for a single MC - * - */ - interface_ip.wire_is_mat_type = 2; - interface_ip.wire_os_mat_type = 2; - interface_ip.wt =Global; - set_mc_param(); - frontend = new MCFrontEnd(XML, &interface_ip, mcp, mc_type); - area.set_area(area.get_area()+ frontend->area.get_area()); - transecEngine = new MCBackend(&interface_ip, mcp, mc_type); - area.set_area(area.get_area()+ transecEngine->area.get_area()); - if (mcp.type==0 || (mcp.type==1&&mcp.withPHY)) - { - PHY = new MCPHY(&interface_ip, mcp, mc_type); - area.set_area(area.get_area()+ PHY->area.get_area()); - } - //+++++++++Transaction engine +++++++++++++++++ ////TODO needs better numbers, Run the RTL code from OpenSparc. -// transecEngine.initialize(&interface_ip); -// transecEngine.peakDataTransferRate = XML->sys.mem.peak_transfer_rate; -// transecEngine.memDataWidth = dataBusWidth; -// transecEngine.memRank = XML->sys.mem.number_ranks; -// //transecEngine.memAccesses=XML->sys.mc.memory_accesses; -// //transecEngine.llcBlocksize=llcBlockSize; -// transecEngine.compute(); -// transecEngine.area.set_area(XML->sys.mc.memory_channels_per_mc*transecEngine.area.get_area()) ; -// area.set_area(area.get_area()+ transecEngine.area.get_area()); -// ///cout<<"area="<sys.mem.peak_transfer_rate; -// PHY.memDataWidth = dataBusWidth; -// //PHY.memAccesses=PHY.peakDataTransferRate;//this is the max power -// //PHY.llcBlocksize=llcBlockSize; -// PHY.compute(); -// PHY.area.set_area(XML->sys.mc.memory_channels_per_mc*PHY.area.get_area()) ; -// area.set_area(area.get_area()+ PHY.area.get_area()); - ///cout<<"area="<sys.core[0].opcode_width + dataBusWidth; -// pipeLogic = new pipeline(is_default, &interface_ip); -// //pipeLogic.init_pipeline(is_default, &interface_ip); -// pipeLogic->compute_pipeline(); -// area.set_area(area.get_area()+ pipeLogic->area.get_area()*1e-6); -// area.set_area((area.get_area()+mc_area*1e-6)*1.1);//placement and routing overhead -// -// -//// //clock -//// clockNetwork.init_wire_external(is_default, &interface_ip); -//// clockNetwork.clk_area =area*1.1;//10% of placement overhead. rule of thumb -//// clockNetwork.end_wiring_level =5;//toplevel metal -//// clockNetwork.start_wiring_level =5;//toplevel metal -//// clockNetwork.num_regs = pipeLogic.tot_stage_vector; -//// clockNetwork.optimize_wire(); + if (!node_name) + warnMissingStatName(statNode->getAttribute("id")); + ASSIGN_FP_IF("duty_cycle", mcs.duty_cycle); + ASSIGN_FP_IF("perc_load", mcs.perc_load); + ASSIGN_FP_IF("memory_reads", mcs.reads); + ASSIGN_INT_IF("memory_writes", mcs.writes); -} -void MemoryController::computeEnergy(bool is_tdp) -{ - - frontend->computeEnergy(is_tdp); - transecEngine->computeEnergy(is_tdp); - if (mcp.type==0 || (mcp.type==1&&mcp.withPHY)) - { - PHY->computeEnergy(is_tdp); - } - if (is_tdp) - { - power = power + frontend->power + transecEngine->power; - if (mcp.type==0 || (mcp.type==1&&mcp.withPHY)) - { - power = power + PHY->power; - } - } - else - { - rt_power = rt_power + frontend->rt_power + transecEngine->rt_power; - if (mcp.type==0 || (mcp.type==1&&mcp.withPHY)) - { - rt_power = rt_power + PHY->rt_power; - } - } -} - -void MemoryController::displayEnergy(uint32_t indent,int plevel,bool is_tdp) -{ - string indent_str(indent, ' '); - string indent_str_next(indent+2, ' '); - bool long_channel = XML->sys.longer_channel_device; - - if (is_tdp) - { - cout << "Memory Controller:" << endl; - cout << indent_str<< "Area = " << area.get_area()*1e-6<< " mm^2" << endl; - cout << indent_str << "Peak Dynamic = " << power.readOp.dynamic*mcp.clockRate << " W" << endl; - cout << indent_str<< "Subthreshold Leakage = " - << (long_channel? power.readOp.longer_channel_leakage:power.readOp.leakage) <<" W" << endl; - //cout << indent_str<< "Subthreshold Leakage = " << power.readOp.longer_channel_leakage <<" W" << endl; - cout << indent_str<< "Gate Leakage = " << power.readOp.gate_leakage << " W" << endl; - cout << indent_str << "Runtime Dynamic = " << rt_power.readOp.dynamic/mcp.executionTime << " W" << endl; - cout<2){ - frontend->displayEnergy(indent+4,is_tdp); - } - cout << indent_str << "Transaction Engine:" << endl; - cout << indent_str_next << "Area = " << transecEngine->area.get_area()*1e-6<< " mm^2" << endl; - cout << indent_str_next << "Peak Dynamic = " << transecEngine->power.readOp.dynamic*mcp.clockRate << " W" << endl; - cout << indent_str_next << "Subthreshold Leakage = " - << (long_channel? transecEngine->power.readOp.longer_channel_leakage:transecEngine->power.readOp.leakage) <<" W" << endl; - cout << indent_str_next << "Gate Leakage = " << transecEngine->power.readOp.gate_leakage << " W" << endl; - cout << indent_str_next << "Runtime Dynamic = " << transecEngine->rt_power.readOp.dynamic/mcp.executionTime << " W" << endl; - cout <sys.mc.mc_clock*2;//DDR double pumped - mcp.clockRate *= 1e6; - mcp.executionTime = XML->sys.total_cycles/(XML->sys.target_core_clockrate*1e6); - - mcp.llcBlockSize =int(ceil(XML->sys.mc.llc_line_length/8.0))+XML->sys.mc.llc_line_length;//ecc overhead - mcp.dataBusWidth =int(ceil(XML->sys.mc.databus_width/8.0)) + XML->sys.mc.databus_width; - mcp.addressBusWidth =int(ceil(XML->sys.mc.addressbus_width));//XML->sys.physical_address_width; - mcp.opcodeW =16; - mcp.num_mcs = XML->sys.mc.number_mcs; - mcp.num_channels = XML->sys.mc.memory_channels_per_mc; - mcp.reads = XML->sys.mc.memory_reads; - mcp.writes = XML->sys.mc.memory_writes; - //+++++++++Transaction engine +++++++++++++++++ ////TODO needs better numbers, Run the RTL code from OpenSparc. - mcp.peakDataTransferRate = XML->sys.mc.peak_transfer_rate; - mcp.memRank = XML->sys.mc.number_ranks; - //++++++++++++++PHY ++++++++++++++++++++++++++ //TODO needs better numbers - //PHY.memAccesses=PHY.peakDataTransferRate;//this is the max power - //PHY.llcBlocksize=llcBlockSize; - mcp.frontend_duty_cycle = 0.5;//for max power, the actual off-chip links is bidirectional but time shared - mcp.LVDS = XML->sys.mc.LVDS; - mcp.type = XML->sys.mc.type; - mcp.withPHY = XML->sys.mc.withPHY; - } -// else if (mc_type==FLASHC) -// { -// mcp.clockRate =XML->sys.flashc.mc_clock*2;//DDR double pumped -// mcp.clockRate *= 1e6; -// mcp.executionTime = XML->sys.total_cycles/(XML->sys.target_core_clockrate*1e6); -// -// mcp.llcBlockSize =int(ceil(XML->sys.flashc.llc_line_length/8.0))+XML->sys.flashc.llc_line_length;//ecc overhead -// mcp.dataBusWidth =int(ceil(XML->sys.flashc.databus_width/8.0)) + XML->sys.flashc.databus_width; -// mcp.addressBusWidth =int(ceil(XML->sys.flashc.addressbus_width));//XML->sys.physical_address_width; -// mcp.opcodeW =16; -// mcp.num_mcs = XML->sys.flashc.number_mcs; -// mcp.num_channels = XML->sys.flashc.memory_channels_per_mc; -// mcp.reads = XML->sys.flashc.memory_reads; -// mcp.writes = XML->sys.flashc.memory_writes; -// //+++++++++Transaction engine +++++++++++++++++ ////TODO needs better numbers, Run the RTL code from OpenSparc. -// mcp.peakDataTransferRate = XML->sys.flashc.peak_transfer_rate; -// mcp.memRank = XML->sys.flashc.number_ranks; -// //++++++++++++++PHY ++++++++++++++++++++++++++ //TODO needs better numbers -// //PHY.memAccesses=PHY.peakDataTransferRate;//this is the max power -// //PHY.llcBlocksize=llcBlockSize; -// mcp.frontend_duty_cycle = 0.5;//for max power, the actual off-chip links is bidirectional but time shared -// mcp.LVDS = XML->sys.flashc.LVDS; -// mcp.type = XML->sys.flashc.type; -// } - else - { - cout<<"Unknown memory controller type: neither DRAM controller nor Flash controller" <