portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-05 19:55:56 +01:00
Code Releases Activity

Namespace, format

Browse Source
This commit is contained in:
paboyle 2018-01-12 23:55:38 +00:00
parent 7e70f4ed9c
commit c01a1e02fe
4 changed files with 109 additions and 107 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

23
lib/perfmon/PerfCount.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -25,8 +25,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_PERFCOUNT_H
#define GRID_PERFCOUNT_H
@ -47,7 +47,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <x86intrin.h>
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
#ifdef __linux__
static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid,
@ -84,9 +84,9 @@ inline uint64_t cyclecount(void){
#ifdef __bgq__
inline uint64_t cyclecount(void){
uint64_t tmp;
asm volatile ("mfspr %0,0x10C" : "=&r" (tmp) );
return tmp;
uint64_t tmp;
asm volatile ("mfspr %0,0x10C" : "=&r" (tmp) );
return tmp;
}
#elif defined __x86_64__
inline uint64_t cyclecount(void){
@ -97,7 +97,7 @@ inline uint64_t cyclecount(void){
#else
inline uint64_t cyclecount(void){
return 0;
return 0;
}
#endif
@ -225,8 +225,8 @@ public:
int N = PerformanceCounterConfigs[PCT].normalisation;
const char * sn = PerformanceCounterConfigs[N].name ;
const char * sc = PerformanceCounterConfigs[PCT].name;
std::printf("tsc = %llu %s = %llu %s = %20llu\n (%s/%s) rate = %lf\n", elapsed,sn ,cycles,
sc, count, sc,sn, (double)count/(double)cycles);
std::printf("tsc = %llu %s = %llu %s = %20llu\n (%s/%s) rate = %lf\n", elapsed,sn ,cycles,
sc, count, sc,sn, (double)count/(double)cycles);
#else
std::printf("%llu cycles \n", elapsed );
#endif
@ -241,5 +241,6 @@ public:
};
}
NAMESPACE_END(Grid);
#endif

67
lib/perfmon/Stat.cc
View File

@ -2,7 +2,7 @@
#include <Grid/perfmon/PerfCount.h>
#include <Grid/perfmon/Stat.h>
namespace Grid {
NAMESPACE_BEGIN(Grid);
bool PmuStat::pmu_initialized=false;
@ -175,39 +175,39 @@ void PmuStat::KNLevsetup(const char *ename, int &fd, int event, int umask)
}
void PmuStat::KNLsetup(void){
void PmuStat::KNLsetup(void){
int ret;
char fname[1024];
int ret;
char fname[1024];
// MC RPQ inserts and WPQ inserts (reads & writes)
for (int mc = 0; mc < NMC; ++mc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_imc_%d",mc);
// RPQ Inserts
KNLevsetup(fname, gbl.mc_rd[mc], 0x1, 0x1);
// WPQ Inserts
KNLevsetup(fname, gbl.mc_wr[mc], 0x2, 0x1);
}
// EDC RPQ inserts and WPQ inserts
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_eclk_%d",edc);
// RPQ inserts
KNLevsetup(fname, gbl.edc_rd[edc], 0x1, 0x1);
// WPQ inserts
KNLevsetup(fname, gbl.edc_wr[edc], 0x2, 0x1);
}
// EDC HitE, HitM, MissE, MissM
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_uclk_%d", edc);
KNLevsetup(fname, gbl.edc_hite[edc], 0x2, 0x1);
KNLevsetup(fname, gbl.edc_hitm[edc], 0x2, 0x2);
KNLevsetup(fname, gbl.edc_misse[edc], 0x2, 0x4);
KNLevsetup(fname, gbl.edc_missm[edc], 0x2, 0x8);
}
}
// MC RPQ inserts and WPQ inserts (reads & writes)
for (int mc = 0; mc < NMC; ++mc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_imc_%d",mc);
// RPQ Inserts
KNLevsetup(fname, gbl.mc_rd[mc], 0x1, 0x1);
// WPQ Inserts
KNLevsetup(fname, gbl.mc_wr[mc], 0x2, 0x1);
}
// EDC RPQ inserts and WPQ inserts
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_eclk_%d",edc);
// RPQ inserts
KNLevsetup(fname, gbl.edc_rd[edc], 0x1, 0x1);
// WPQ inserts
KNLevsetup(fname, gbl.edc_wr[edc], 0x2, 0x1);
}
// EDC HitE, HitM, MissE, MissM
for (int edc=0; edc < NEDC; ++edc)
{
::snprintf(fname, sizeof(fname), "/sys/devices/uncore_edc_uclk_%d", edc);
KNLevsetup(fname, gbl.edc_hite[edc], 0x2, 0x1);
KNLevsetup(fname, gbl.edc_hitm[edc], 0x2, 0x2);
KNLevsetup(fname, gbl.edc_misse[edc], 0x2, 0x4);
KNLevsetup(fname, gbl.edc_missm[edc], 0x2, 0x8);
}
}
uint64_t PmuStat::KNLreadctr(int fd)
{
@ -242,4 +242,5 @@ void PmuStat::KNLreadctrs(ctrs &c)
}
#endif
}
NAMESPACE_END(Grid);

111
lib/perfmon/Stat.h
View File

@ -5,7 +5,7 @@
#define _KNIGHTS_LANDING_ROOTONLY
#endif
namespace Grid {
NAMESPACE_BEGIN(Grid);
///////////////////////////////////////////////////////////////////////////////
// Extra KNL counters from MCDRAM
@ -15,14 +15,14 @@ namespace Grid {
#define NEDC 8
struct ctrs
{
uint64_t mcrd[NMC];
uint64_t mcwr[NMC];
uint64_t edcrd[NEDC];
uint64_t edcwr[NEDC];
uint64_t edchite[NEDC];
uint64_t edchitm[NEDC];
uint64_t edcmisse[NEDC];
uint64_t edcmissm[NEDC];
uint64_t mcrd[NMC];
uint64_t mcwr[NMC];
uint64_t edcrd[NEDC];
uint64_t edcwr[NEDC];
uint64_t edchite[NEDC];
uint64_t edchitm[NEDC];
uint64_t edcmisse[NEDC];
uint64_t edcmissm[NEDC];
};
// Peter/Azusa:
// Our modification of a code provided by Larry Meadows from Intel
@ -44,61 +44,62 @@ struct knl_gbl_
class PmuStat
{
uint64_t counters[8][256];
uint64_t counters[8][256];
#ifdef _KNIGHTS_LANDING_
static struct knl_gbl_ gbl;
static struct knl_gbl_ gbl;
#endif
const char *name;
const char *name;
uint64_t reads; // memory reads
uint64_t writes; // memory writes
uint64_t mrstart; // memory read counter at start of parallel region
uint64_t mrend; // memory read counter at end of parallel region
uint64_t mwstart; // memory write counter at start of parallel region
uint64_t mwend; // memory write counter at end of parallel region
uint64_t reads; // memory reads
uint64_t writes; // memory writes
uint64_t mrstart; // memory read counter at start of parallel region
uint64_t mrend; // memory read counter at end of parallel region
uint64_t mwstart; // memory write counter at start of parallel region
uint64_t mwend; // memory write counter at end of parallel region
// cumulative counters
uint64_t count; // number of invocations
uint64_t tregion; // total time in parallel region (from thread 0)
uint64_t tcycles; // total cycles inside parallel region
uint64_t inst, ref, cyc; // fixed counters
uint64_t pmc0, pmc1;// pmu
// add memory counters here
// temp variables
uint64_t tstart; // tsc at start of parallel region
uint64_t tend; // tsc at end of parallel region
// map for ctrs values
// 0 pmc0 start
// 1 pmc0 end
// 2 pmc1 start
// 3 pmc1 end
// 4 tsc start
// 5 tsc end
static bool pmu_initialized;
// cumulative counters
uint64_t count; // number of invocations
uint64_t tregion; // total time in parallel region (from thread 0)
uint64_t tcycles; // total cycles inside parallel region
uint64_t inst, ref, cyc; // fixed counters
uint64_t pmc0, pmc1;// pmu
// add memory counters here
// temp variables
uint64_t tstart; // tsc at start of parallel region
uint64_t tend; // tsc at end of parallel region
// map for ctrs values
// 0 pmc0 start
// 1 pmc0 end
// 2 pmc1 start
// 3 pmc1 end
// 4 tsc start
// 5 tsc end
static bool pmu_initialized;
public:
static bool is_init(void){ return pmu_initialized;}
static void pmu_init(void);
static void pmu_fini(void);
static void pmu_start(void);
static void pmu_stop(void);
void accum(int nthreads);
static void xmemctrs(uint64_t *mr, uint64_t *mw);
void start(void);
void enter(int t);
void exit(int t);
void print(void);
void init(const char *regname);
void clear(void);
static bool is_init(void){ return pmu_initialized;}
static void pmu_init(void);
static void pmu_fini(void);
static void pmu_start(void);
static void pmu_stop(void);
void accum(int nthreads);
static void xmemctrs(uint64_t *mr, uint64_t *mw);
void start(void);
void enter(int t);
void exit(int t);
void print(void);
void init(const char *regname);
void clear(void);
#ifdef _KNIGHTS_LANDING_
static void KNLsetup(void);
static uint64_t KNLreadctr(int fd);
static void KNLreadctrs(ctrs &c);
static void KNLevsetup(const char *ename, int &fd, int event, int umask);
static void KNLsetup(void);
static uint64_t KNLreadctr(int fd);
static void KNLreadctrs(ctrs &c);
static void KNLevsetup(const char *ename, int &fd, int event, int umask);
#endif
};
};
NAMESPACE_END(Grid);
}
#endif

15
lib/perfmon/Timer.h
View File

@ -1,4 +1,4 @@
/*************************************************************************************
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -24,8 +24,8 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
See the full license in the file "LICENSE" in the top level distribution directory
*************************************************************************************/
/* END LEGAL */
*************************************************************************************/
/* END LEGAL */
#ifndef GRID_TIME_H
#define GRID_TIME_H
@ -33,11 +33,9 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <ctime>
#include <chrono>
namespace Grid {
// Dress the output; use std::chrono
NAMESPACE_BEGIN(Grid)
// Dress the output; use std::chrono
// C++11 time facilities better?
inline double usecond(void) {
struct timeval tv;
@ -98,5 +96,6 @@ public:
}
};
}
NAMESPACE_END(Grid)
#endif