From 9cafb850bdc8ba404add817f90c25faef6ecf9e3 Mon Sep 17 00:00:00 2001 From: Nick Date: Fri, 13 Sep 2019 07:48:05 -0400 Subject: Revert "add /src/gpuwattch formatting" This reverts commit b7776785a18d3f4e6229cd625f92c11b55894b75. --- src/gpuwattch/processor.h | 511 +++++++++++++++++++++------------------------- 1 file changed, 238 insertions(+), 273 deletions(-) (limited to 'src/gpuwattch/processor.h') diff --git a/src/gpuwattch/processor.h b/src/gpuwattch/processor.h index 326206c..7202460 100644 --- a/src/gpuwattch/processor.h +++ b/src/gpuwattch/processor.h @@ -29,8 +29,7 @@ * ***************************************************************************/ /******************************************************************** -* Modified by: -** +* Modified by: * * Jingwen Leng, Univeristy of Texas, Austin * * Syed Gilani, University of Wisconsin–Madison * * Tayler Hetherington, University of British Columbia * @@ -39,287 +38,253 @@ #ifndef PROCESSOR_H_ #define PROCESSOR_H_ -#include -#include "../gpgpu-sim/visualizer.h" #include "XML_Parse.h" -#include "array.h" -#include "basic_components.h" -#include "cacti/arbiter.h" #include "cacti/area.h" #include "cacti/decoder.h" #include "cacti/parameter.h" -#include "cacti/router.h" +#include "array.h" +#include "cacti/arbiter.h" +#include +#include "basic_components.h" #include "core.h" -#include "iocontrollers.h" #include "memoryctrl.h" -#include "noc.h" +#include "cacti/router.h" #include "sharedcache.h" +#include "noc.h" +#include "iocontrollers.h" +#include "../gpgpu-sim/visualizer.h" + +class Processor : public Component +{ + public: + ParseXML *XML; + vector cores; + vector l2array; + vector l3array; + vector l1dirarray; + vector l2dirarray; + vector nocs; + MemoryController * mc; + NIUController * niu; + PCIeController * pcie; + FlashController * flashcontroller; + InputParameter interface_ip; + double exClockRate; + ProcParam procdynp; + //for debugging nonlinear model + double dyn_power_before_scaling; + + //wire globalInterconnect; + //clock_network globalClock; + Component core, l2, l3, l1dir, l2dir, noc, mcs, cc, nius, pcies,flashcontrollers; + int numCore, numL2, numL3, numNOC, numL1Dir, numL2Dir; + Processor(ParseXML *XML_interface); + void compute(); + void set_proc_param(); + void visualizer_print( gzFile visualizer_file ); + void displayEnergy(uint32_t indent = 0,int plevel = 100, bool is_tdp_parm=true); + void displayDeviceType(int device_type_, uint32_t indent = 0); + void displayInterconnectType(int interconnect_type_, uint32_t indent = 0); + double l2_power; + double idle_core_power; + + double get_const_dynamic_power() + { + + double constpart=0; + constpart+=(mc->frontend->power.readOp.dynamic*0.1*mc->frontend->mcp.clockRate*mc->frontend->mcp.num_mcs*mc->frontend->mcp.executionTime); + constpart+=(mc->transecEngine->power.readOp.dynamic*0.1*mc->transecEngine->mcp.clockRate*mc->transecEngine->mcp.num_mcs*mc->transecEngine->mcp.executionTime); + constpart+=(mc->PHY->power.readOp.dynamic*0.1*mc->PHY->mcp.clockRate*mc->PHY->mcp.num_mcs*mc->PHY->mcp.executionTime); + constpart+=(cores[0]->exu->exeu->base_energy/cores[0]->exu->exeu->clockRate)*(cores[0]->exu->rf_fu_clockRate/cores[0]->exu->clockRate); + constpart+=(cores[0]->exu->mul->base_energy/cores[0]->exu->mul->clockRate); + constpart+=(cores[0]->exu->fp_u->base_energy/cores[0]->exu->fp_u->clockRate); + return constpart; + } +#define COALESCE_SCALE 1 + double get_coefficient_readcoalescing() + { + double value=0; + double perAccessCoalescingEnergy=COALESCE_SCALE * ((0.443e-3)*(0.5e-9)*g_tp.peri_global.Vdd*g_tp.peri_global.Vdd)/(1*1); + value+=mc->frontend->PRT->local_result.power.readOp.dynamic; + value+=mc->frontend->threadMasks->local_result.power.readOp.dynamic; + value+=mc->frontend->PRC->local_result.power.readOp.dynamic; + value+=perAccessCoalescingEnergy; + return value; -class Processor : public Component { - public: - ParseXML *XML; - vector cores; - vector l2array; - vector l3array; - vector l1dirarray; - vector l2dirarray; - vector nocs; - MemoryController *mc; - NIUController *niu; - PCIeController *pcie; - FlashController *flashcontroller; - InputParameter interface_ip; - double exClockRate; - ProcParam procdynp; - // for debugging nonlinear model - double dyn_power_before_scaling; - - // wire globalInterconnect; - // clock_network globalClock; - Component core, l2, l3, l1dir, l2dir, noc, mcs, cc, nius, pcies, - flashcontrollers; - int numCore, numL2, numL3, numNOC, numL1Dir, numL2Dir; - Processor(ParseXML *XML_interface); - void compute(); - void set_proc_param(); - void visualizer_print(gzFile visualizer_file); - void displayEnergy(uint32_t indent = 0, int plevel = 100, - bool is_tdp_parm = true); - void displayDeviceType(int device_type_, uint32_t indent = 0); - void displayInterconnectType(int interconnect_type_, uint32_t indent = 0); - double l2_power; - double idle_core_power; - - double get_const_dynamic_power() { - double constpart = 0; - constpart += (mc->frontend->power.readOp.dynamic * 0.1 * - mc->frontend->mcp.clockRate * mc->frontend->mcp.num_mcs * - mc->frontend->mcp.executionTime); - constpart += - (mc->transecEngine->power.readOp.dynamic * 0.1 * - mc->transecEngine->mcp.clockRate * mc->transecEngine->mcp.num_mcs * - mc->transecEngine->mcp.executionTime); - constpart += (mc->PHY->power.readOp.dynamic * 0.1 * mc->PHY->mcp.clockRate * - mc->PHY->mcp.num_mcs * mc->PHY->mcp.executionTime); - constpart += - (cores[0]->exu->exeu->base_energy / cores[0]->exu->exeu->clockRate) * - (cores[0]->exu->rf_fu_clockRate / cores[0]->exu->clockRate); - constpart += - (cores[0]->exu->mul->base_energy / cores[0]->exu->mul->clockRate); - constpart += - (cores[0]->exu->fp_u->base_energy / cores[0]->exu->fp_u->clockRate); - return constpart; - } -#define COALESCE_SCALE 1 - double get_coefficient_readcoalescing() { - double value = 0; - double perAccessCoalescingEnergy = - COALESCE_SCALE * - ((0.443e-3) * (0.5e-9) * g_tp.peri_global.Vdd * g_tp.peri_global.Vdd) / - (1 * 1); - value += mc->frontend->PRT->local_result.power.readOp.dynamic; - value += mc->frontend->threadMasks->local_result.power.readOp.dynamic; - value += mc->frontend->PRC->local_result.power.readOp.dynamic; - value += perAccessCoalescingEnergy; - return value; - } - double get_coefficient_writecoalescing() { - double value = 0; - double perAccessCoalescingEnergy = - COALESCE_SCALE * - ((0.443e-3) * (0.5e-9) * g_tp.peri_global.Vdd * g_tp.peri_global.Vdd) / - (1 * 1); - value += (mc->frontend->PRT->local_result.power.writeOp.dynamic); - value += mc->frontend->threadMasks->local_result.power.writeOp.dynamic; - value += mc->frontend->PRC->local_result.power.writeOp.dynamic; - value += perAccessCoalescingEnergy; - return value; - } - - double get_coefficient_noc_accesses() { - double read_coef = 0; - // the 32/4 is applied to the NoC access counters (32/4*L2 cache access) - read_coef += nocs[0]->router->buffer.power.readOp.dynamic; - read_coef += nocs[0]->router->buffer.power.writeOp.dynamic; - read_coef += nocs[0]->router->crossbar.power.readOp.dynamic; - read_coef += nocs[0]->router->arbiter.power.readOp.dynamic; - return read_coef; - } - - double get_coefficient_l2_read_hits() { - double read_coef = 0; - if (XML->sys.number_of_L2s > 0) - read_coef = - l2array[0]->unicache.caches->local_result.power.readOp.dynamic; - return read_coef; - } - - double get_coefficient_l2_read_misses() { - double read_coef = 0; - if (XML->sys.number_of_L2s > 0) - read_coef = - l2array[0] - ->unicache.caches->local_result.tag_array2->power.readOp.dynamic; - return read_coef; - } - - double get_coefficient_l2_write_hits() { - double read_coef = 0; - if (XML->sys.number_of_L2s > 0) - read_coef = - l2array[0]->unicache.caches->local_result.power.writeOp.dynamic; - return read_coef; - } - double get_coefficient_l2_write_misses() { - double read_coef = 0; - if (XML->sys.number_of_L2s > 0) { - read_coef = l2array[0] - ->unicache.caches->local_result.tag_array2->power.writeOp - .dynamic; //*(32/4); // removed by Jingwen, the scaling - //of 32/4 is not used in the mcpat - read_coef += - l2array[0]->unicache.caches->local_result.power.writeOp.dynamic; - read_coef += - l2array[0]->unicache.missb->local_result.power.searchOp.dynamic; - read_coef += - l2array[0]->unicache.missb->local_result.power.writeOp.dynamic; - read_coef += - l2array[0]->unicache.ifb->local_result.power.searchOp.dynamic; - read_coef += l2array[0]->unicache.ifb->local_result.power.writeOp.dynamic; - read_coef += - l2array[0]->unicache.prefetchb->local_result.power.searchOp.dynamic; - read_coef += - l2array[0]->unicache.prefetchb->local_result.power.writeOp.dynamic; - read_coef += - l2array[0]->unicache.wbb->local_result.power.searchOp.dynamic; - read_coef += l2array[0]->unicache.wbb->local_result.power.writeOp.dynamic; } + double get_coefficient_writecoalescing() + { + double value=0; + double perAccessCoalescingEnergy=COALESCE_SCALE *((0.443e-3)*(0.5e-9)*g_tp.peri_global.Vdd*g_tp.peri_global.Vdd)/(1*1); + value+=(mc->frontend->PRT->local_result.power.writeOp.dynamic); + value+=mc->frontend->threadMasks->local_result.power.writeOp.dynamic; + value+=mc->frontend->PRC->local_result.power.writeOp.dynamic; + value+=perAccessCoalescingEnergy; + return value; + } + + + double get_coefficient_noc_accesses() { + double read_coef=0; + // the 32/4 is applied to the NoC access counters (32/4*L2 cache access) + read_coef+= nocs[0]->router->buffer.power.readOp.dynamic; + read_coef+= nocs[0]->router->buffer.power.writeOp.dynamic; + read_coef+= nocs[0]->router->crossbar.power.readOp.dynamic; + read_coef+= nocs[0]->router->arbiter.power.readOp.dynamic; + return read_coef; + } + + double get_coefficient_l2_read_hits(){ + double read_coef=0; + if(XML->sys.number_of_L2s>0) + read_coef = l2array[0]->unicache.caches->local_result.power.readOp.dynamic; + return read_coef; + } + + double get_coefficient_l2_read_misses(){ + double read_coef=0; + if(XML->sys.number_of_L2s>0) + read_coef = l2array[0]->unicache.caches->local_result.tag_array2->power.readOp.dynamic; + return read_coef; + + } + + double get_coefficient_l2_write_hits(){ + double read_coef=0; + if(XML->sys.number_of_L2s>0) + read_coef = l2array[0]->unicache.caches->local_result.power.writeOp.dynamic; + return read_coef; + + } + double get_coefficient_l2_write_misses(){ + double read_coef=0; + if(XML->sys.number_of_L2s>0){ + read_coef = l2array[0]->unicache.caches->local_result.tag_array2->power.writeOp.dynamic;//*(32/4); // removed by Jingwen, the scaling of 32/4 is not used in the mcpat + read_coef +=l2array[0]->unicache.caches->local_result.power.writeOp.dynamic; + read_coef += l2array[0]->unicache.missb->local_result.power.searchOp.dynamic; + read_coef += l2array[0]->unicache.missb->local_result.power.writeOp.dynamic; + read_coef += l2array[0]->unicache.ifb->local_result.power.searchOp.dynamic; + read_coef += l2array[0]->unicache.ifb->local_result.power.writeOp.dynamic; + read_coef += l2array[0]->unicache.prefetchb->local_result.power.searchOp.dynamic; + read_coef += l2array[0]->unicache.prefetchb->local_result.power.writeOp.dynamic; + read_coef += l2array[0]->unicache.wbb->local_result.power.searchOp.dynamic; + read_coef += l2array[0]->unicache.wbb->local_result.power.writeOp.dynamic; + } + + return read_coef; + + } + + + double get_coefficient_mem_reads() + { + double value=0; + value+= (mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* + (mc->frontend->frontendBuffer->local_result.power.searchOp.dynamic); + + value+= (mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* + (mc->frontend->frontendBuffer->local_result.power.readOp.dynamic); + +//TODO: Jingwen this should only compute for one time? + //value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72) + //*(mc->frontend->frontendBuffer->local_result.power.readOp.dynamic); + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->mcp.dataBusWidth)* + (mc->frontend->readBuffer->local_result.power.readOp.dynamic); + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->mcp.dataBusWidth)* + (mc->frontend->readBuffer->local_result.power.writeOp.dynamic); + + value+=mc->dram->dramp.rd_coeff; + /* + value+=mc->frontend->PRT->local_result.power.readOp.dynamic; + value+=mc->frontend->threadMasks->local_result.power.readOp.dynamic; + value+=mc->frontend->PRC->local_result.power.readOp.dynamic; + value+=perAccessCoalescingEnergy; + */ + value+=(mc->transecEngine->mcp.llcBlockSize*8.0/mc->transecEngine->mcp.dataBusWidth*mc->transecEngine->power_t.readOp.dynamic); + + //if mcp.type ==1 TODO: add this check here + value += (mc->PHY->power_t.readOp.dynamic)*(mc->PHY->mcp.llcBlockSize)*8/1e9/mc->PHY->mcp.executionTime*(mc->PHY->mcp.executionTime); + //printf("MC PHY read power coeff: %f\n",(mc->PHY->power_t.readOp.dynamic)*(mc->PHY->mcp.llcBlockSize)*8/1e9/mc->PHY->mcp.executionTime*(mc->PHY->mcp.executionTime)); + //printf("MC trans read power coeff: %f\n",(mc->transecEngine->mcp.llcBlockSize*8.0/mc->transecEngine->mcp.dataBusWidth*mc->transecEngine->power_t.readOp.dynamic)); + +//TODO: Jingwen nocs stats should not be here +// value+= nocs[0]->router->buffer.power.readOp.dynamic*(32/4); +// value+= nocs[0]->router->buffer.power.writeOp.dynamic*(32/4); +// value+= nocs[0]->router->crossbar.power.readOp.dynamic*(32/4); +// value+= nocs[0]->router->arbiter.power.readOp.dynamic*(32/4); + + //return 0.4*value; + return value; + } + + + double get_coefficient_mem_writes() + { + double value=0; + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* + (mc->frontend->frontendBuffer->local_result.power.searchOp.dynamic); + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* + (mc->frontend->frontendBuffer->local_result.power.writeOp.dynamic); + + //value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* + // (mc->frontend->frontendBuffer->local_result.power.writeOp.dynamic); + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth)* + (mc->frontend->writeBuffer->local_result.power.readOp.dynamic); + + value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth)* + (mc->frontend->writeBuffer->local_result.power.writeOp.dynamic); + + value+=mc->dram->dramp.wr_coeff; + /* + value+=(mc->frontend->PRT->local_result.power.writeOp.dynamic); + + value+=mc->frontend->threadMasks->local_result.power.writeOp.dynamic; + + value+=mc->frontend->PRC->local_result.power.writeOp.dynamic; + + value+=perAccessCoalescingEnergy; + */ + + value+=(mc->transecEngine->mcp.llcBlockSize*8.0/mc->transecEngine->mcp.dataBusWidth*mc->transecEngine->power_t.readOp.dynamic); + + + //if mcp.type ==1 TODO: add this check here + value += (mc->PHY->power_t.readOp.dynamic)*(mc->PHY->mcp.llcBlockSize)*8/1e9/mc->PHY->mcp.executionTime*(mc->PHY->mcp.executionTime); + + +//TODO: Jingwen nocs stats should not be here +// value+= nocs[0]->router->buffer.power.readOp.dynamic*(32/4); +// +// value+= nocs[0]->router->buffer.power.writeOp.dynamic*(32/4); +// +// value+= nocs[0]->router->crossbar.power.readOp.dynamic*(32/4); +// +// value+= nocs[0]->router->arbiter.power.readOp.dynamic*(32/4); +// + //return 0.4*value; + return value; + } + + double get_coefficient_mem_pre() + { + double value=0; + value+=mc->dram->dramp.pre_coeff; + //return 0.4*value; + return value; + + } + + //nonlinear scale + void nonlinear_scale(int, double, int); + void coefficient_scale(); + void iterative_lse(double *, double* ); - return read_coef; - } - - double get_coefficient_mem_reads() { - double value = 0; - value += - (mc->frontend->mcp.llcBlockSize * 8.0 / mc->frontend->mcp.dataBusWidth * - mc->frontend->mcp.dataBusWidth / 72) * - (mc->frontend->frontendBuffer->local_result.power.searchOp.dynamic); - - value += - (mc->frontend->mcp.llcBlockSize * 8.0 / mc->frontend->mcp.dataBusWidth * - mc->frontend->mcp.dataBusWidth / 72) * - (mc->frontend->frontendBuffer->local_result.power.readOp.dynamic); - - // TODO: Jingwen this should only compute for one time? - // value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72) - //*(mc->frontend->frontendBuffer->local_result.power.readOp.dynamic); - - value += (mc->frontend->mcp.llcBlockSize * 8.0 / mc->mcp.dataBusWidth) * - (mc->frontend->readBuffer->local_result.power.readOp.dynamic); - - value += (mc->frontend->mcp.llcBlockSize * 8.0 / mc->mcp.dataBusWidth) * - (mc->frontend->readBuffer->local_result.power.writeOp.dynamic); - - value += mc->dram->dramp.rd_coeff; - /* - value+=mc->frontend->PRT->local_result.power.readOp.dynamic; - value+=mc->frontend->threadMasks->local_result.power.readOp.dynamic; - value+=mc->frontend->PRC->local_result.power.readOp.dynamic; - value+=perAccessCoalescingEnergy; - */ - value += (mc->transecEngine->mcp.llcBlockSize * 8.0 / - mc->transecEngine->mcp.dataBusWidth * - mc->transecEngine->power_t.readOp.dynamic); - - // if mcp.type ==1 TODO: add this check here - value += (mc->PHY->power_t.readOp.dynamic) * (mc->PHY->mcp.llcBlockSize) * - 8 / 1e9 / mc->PHY->mcp.executionTime * - (mc->PHY->mcp.executionTime); - // printf("MC PHY read power coeff: - // %f\n",(mc->PHY->power_t.readOp.dynamic)*(mc->PHY->mcp.llcBlockSize)*8/1e9/mc->PHY->mcp.executionTime*(mc->PHY->mcp.executionTime)); - // printf("MC trans read power coeff: - // %f\n",(mc->transecEngine->mcp.llcBlockSize*8.0/mc->transecEngine->mcp.dataBusWidth*mc->transecEngine->power_t.readOp.dynamic)); - - // TODO: Jingwen nocs stats should not be here - // value+= nocs[0]->router->buffer.power.readOp.dynamic*(32/4); - // value+= nocs[0]->router->buffer.power.writeOp.dynamic*(32/4); - // value+= nocs[0]->router->crossbar.power.readOp.dynamic*(32/4); - // value+= nocs[0]->router->arbiter.power.readOp.dynamic*(32/4); - - // return 0.4*value; - return value; - } - - double get_coefficient_mem_writes() { - double value = 0; - - value += - (mc->frontend->mcp.llcBlockSize * 8.0 / mc->frontend->mcp.dataBusWidth * - mc->frontend->mcp.dataBusWidth / 72) * - (mc->frontend->frontendBuffer->local_result.power.searchOp.dynamic); - - value += - (mc->frontend->mcp.llcBlockSize * 8.0 / mc->frontend->mcp.dataBusWidth * - mc->frontend->mcp.dataBusWidth / 72) * - (mc->frontend->frontendBuffer->local_result.power.writeOp.dynamic); - - // value+=(mc->frontend->mcp.llcBlockSize*8.0/mc->frontend->mcp.dataBusWidth*mc->frontend->mcp.dataBusWidth/72)* - // (mc->frontend->frontendBuffer->local_result.power.writeOp.dynamic); - - value += (mc->frontend->mcp.llcBlockSize * 8.0 / - mc->frontend->mcp.dataBusWidth) * - (mc->frontend->writeBuffer->local_result.power.readOp.dynamic); - - value += (mc->frontend->mcp.llcBlockSize * 8.0 / - mc->frontend->mcp.dataBusWidth) * - (mc->frontend->writeBuffer->local_result.power.writeOp.dynamic); - - value += mc->dram->dramp.wr_coeff; - /* - value+=(mc->frontend->PRT->local_result.power.writeOp.dynamic); - - value+=mc->frontend->threadMasks->local_result.power.writeOp.dynamic; - - value+=mc->frontend->PRC->local_result.power.writeOp.dynamic; - - value+=perAccessCoalescingEnergy; - */ - - value += (mc->transecEngine->mcp.llcBlockSize * 8.0 / - mc->transecEngine->mcp.dataBusWidth * - mc->transecEngine->power_t.readOp.dynamic); - - // if mcp.type ==1 TODO: add this check here - value += (mc->PHY->power_t.readOp.dynamic) * (mc->PHY->mcp.llcBlockSize) * - 8 / 1e9 / mc->PHY->mcp.executionTime * - (mc->PHY->mcp.executionTime); - - // TODO: Jingwen nocs stats should not be here - // value+= nocs[0]->router->buffer.power.readOp.dynamic*(32/4); - // - // value+= nocs[0]->router->buffer.power.writeOp.dynamic*(32/4); - // - // value+= nocs[0]->router->crossbar.power.readOp.dynamic*(32/4); - // - // value+= nocs[0]->router->arbiter.power.readOp.dynamic*(32/4); - // - // return 0.4*value; - return value; - } - - double get_coefficient_mem_pre() { - double value = 0; - value += mc->dram->dramp.pre_coeff; - // return 0.4*value; - return value; - } - - // nonlinear scale - void nonlinear_scale(int, double, int); - void coefficient_scale(); - void iterative_lse(double *, double *); - - ~Processor(); + ~Processor(); }; #endif /* PROCESSOR_H_ */ -- cgit v1.3