Loading [MathJax]/extensions/tex2jax.js
The Gaudi Framework  v31r0 (aeb156f0)
•All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Properties Friends Macros Modules Pages
Classes | Macros | Functions
pfm_gen_analysis.cpp File Reference
#include <ctype.h>
#include <cxxabi.h>
#include <fcntl.h>
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#include <zlib.h>
#include <algorithm>
#include <iostream>
#include <list>
#include <map>
#include <sstream>
#include <string>
#include <vector>
#include <dirent.h>
#include <errno.h>
Include dependency graph for pfm_gen_analysis.cpp:

Go to the source code of this file.

Classes

class  PipeReader
 
class  FileInfo
 
struct  FileInfo::CacheItem
 
struct  FileInfo::CacheItemComparator
 
class  S_module
 

Macros

#define CORE_L2_MISS_CYCLES   200
 
#define CORE_L2_HIT_CYCLES   14.5
 
#define CORE_L1_DTLB_MISS_CYCLES   10
 
#define CORE_LCP_STALL_CYCLES   6
 
#define CORE_UNKNOWN_ADDR_STORE_CYCLES   5
 
#define CORE_OVERLAPPING_CYCLES   6
 
#define CORE_SPAN_ACROSS_CACHE_LINE_CYCLES   20
 
#define I7_L1_DTLB_WALK_COMPLETED_CYCLES   35
 
#define I7_L1_ITLB_WALK_COMPLETED_CYCLES   35
 
#define I7_L2_HIT_CYCLES   6
 
#define I7_L3_UNSHARED_HIT_CYCLES   35
 
#define I7_OTHER_CORE_L2_HIT_CYCLES   60
 
#define I7_OTHER_CORE_L2_HITM_CYCLES   75
 
#define I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES   225
 
#define I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES   360
 
#define I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES   180
 
#define I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT   200
 
#define I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT   350
 
#define I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP   35
 
#define I7_IFETCH_L2_MISS_L3_HIT_SNOOP   60
 
#define I7_IFETCH_L2_MISS_L3_HITM   75
 
#define I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD   180
 
#define MAX_MODULES   1000
 
#define EXPECTED_CPI   0.25
 
#define MAX_FILENAME_LENGTH   1024
 
#define MAX_SAMPLE_INDEX_LENGTH   10000
 
#define MAX_SYM_LENGTH   15000
 
#define MAX_SYM_MOD_LENGTH   20000
 
#define MAX_LIB_LENGTH   5000
 
#define MAX_LIB_MOD_LENGTH   7000
 
#define MAX_SIMPLE_SYM_LENGTH   300
 
#define MAX_SIMPLE_SYM_MOD_LENGTH   500
 
#define MAX_SIMPLE_LIB_LENGTH   300
 
#define MAX_SIMPLE_LIB_MOD_LENGTH   500
 
#define MAX_LINE_LENGTH   20000
 
#define MAX_EVENT_NAME_LENGTH   150
 
#define MAX_MODULE_NAME_LENGTH   250
 
#define MAX_VALUE_STRING_LENGTH   250
 
#define MAX_ARCH_NAME_LENGTH   20
 
#define MAX_CMASK_STR_LENGTH   5
 
#define MAX_INV_STR_LENGTH   5
 
#define MAX_SP_STR_LENGTH   50
 
#define PIPE_BUFFER_LENGTH   1000
 

Functions

bool skipWhitespaces (const char *srcbuffer, const char **destbuffer)
 
bool skipString (const char *strptr, const char *srcbuffer, const char **dstbuffer)
 
void init_core_caa_events ()
 
void init_nhm_caa_events ()
 
bool check_for_core_caa_events ()
 
bool check_for_nhm_caa_events ()
 
void init_core_caa_events_displ ()
 
void calc_core_deriv_values (double totalCycles)
 
void init_nhm_caa_events_displ ()
 
void calc_nhm_deriv_values (double totalCycles)
 
void html_special_chars (const char *s, char *s_mod)
 
const char * func_name (const char *demangled_symbol)
 
void put_S_module (S_module *cur_module, const char *dir)
 
int read_S_file (const char *dir, const char *filename)
 
int read_S_events (const char *dir, const char *filename)
 
int finalize_S_html_pages (const char *dir)
 
int read_C_file (const char *dir, const char *filename)
 
void put_C_header (FILE *fp, std::vector< std::string > &columns)
 
void put_C_modules (FILE *fp, std::vector< std::string > &columns)
 
void put_C_footer (FILE *fp)
 
void put_C_header_csv (FILE *fp, std::vector< std::string > &columns)
 
void put_C_modules_csv (FILE *fp, std::vector< std::string > &columns)
 
double normalize (std::string field, double value, double normalizeTo)
 
void calc_post_deriv_values ()
 
double getTotalCycles ()
 
int main (int argc, char *argv[])
 

Macro Definition Documentation

#define CORE_L1_DTLB_MISS_CYCLES   10

Definition at line 41 of file pfm_gen_analysis.cpp.

#define CORE_L2_HIT_CYCLES   14.5

Definition at line 40 of file pfm_gen_analysis.cpp.

#define CORE_L2_MISS_CYCLES   200

Definition at line 39 of file pfm_gen_analysis.cpp.

#define CORE_LCP_STALL_CYCLES   6

Definition at line 42 of file pfm_gen_analysis.cpp.

#define CORE_OVERLAPPING_CYCLES   6

Definition at line 44 of file pfm_gen_analysis.cpp.

#define CORE_SPAN_ACROSS_CACHE_LINE_CYCLES   20

Definition at line 45 of file pfm_gen_analysis.cpp.

#define CORE_UNKNOWN_ADDR_STORE_CYCLES   5

Definition at line 43 of file pfm_gen_analysis.cpp.

#define EXPECTED_CPI   0.25

Definition at line 66 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP   35

Definition at line 59 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L2_MISS_L3_HIT_SNOOP   60

Definition at line 60 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L2_MISS_L3_HITM   75

Definition at line 61 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT   200

Definition at line 57 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD   180

Definition at line 62 of file pfm_gen_analysis.cpp.

#define I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT   350

Definition at line 58 of file pfm_gen_analysis.cpp.

#define I7_L1_DTLB_WALK_COMPLETED_CYCLES   35

Definition at line 48 of file pfm_gen_analysis.cpp.

#define I7_L1_ITLB_WALK_COMPLETED_CYCLES   35

Definition at line 49 of file pfm_gen_analysis.cpp.

#define I7_L2_HIT_CYCLES   6

Definition at line 50 of file pfm_gen_analysis.cpp.

#define I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES   225

Definition at line 54 of file pfm_gen_analysis.cpp.

#define I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES   180

Definition at line 56 of file pfm_gen_analysis.cpp.

#define I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES   360

Definition at line 55 of file pfm_gen_analysis.cpp.

#define I7_L3_UNSHARED_HIT_CYCLES   35

Definition at line 51 of file pfm_gen_analysis.cpp.

#define I7_OTHER_CORE_L2_HIT_CYCLES   60

Definition at line 52 of file pfm_gen_analysis.cpp.

#define I7_OTHER_CORE_L2_HITM_CYCLES   75

Definition at line 53 of file pfm_gen_analysis.cpp.

#define MAX_ARCH_NAME_LENGTH   20

Definition at line 82 of file pfm_gen_analysis.cpp.

#define MAX_CMASK_STR_LENGTH   5

Definition at line 83 of file pfm_gen_analysis.cpp.

#define MAX_EVENT_NAME_LENGTH   150

Definition at line 79 of file pfm_gen_analysis.cpp.

#define MAX_FILENAME_LENGTH   1024

Definition at line 68 of file pfm_gen_analysis.cpp.

#define MAX_INV_STR_LENGTH   5

Definition at line 84 of file pfm_gen_analysis.cpp.

#define MAX_LIB_LENGTH   5000

Definition at line 72 of file pfm_gen_analysis.cpp.

#define MAX_LIB_MOD_LENGTH   7000

Definition at line 73 of file pfm_gen_analysis.cpp.

#define MAX_LINE_LENGTH   20000

Definition at line 78 of file pfm_gen_analysis.cpp.

#define MAX_MODULE_NAME_LENGTH   250

Definition at line 80 of file pfm_gen_analysis.cpp.

#define MAX_MODULES   1000

Definition at line 64 of file pfm_gen_analysis.cpp.

#define MAX_SAMPLE_INDEX_LENGTH   10000

Definition at line 69 of file pfm_gen_analysis.cpp.

#define MAX_SIMPLE_LIB_LENGTH   300

Definition at line 76 of file pfm_gen_analysis.cpp.

#define MAX_SIMPLE_LIB_MOD_LENGTH   500

Definition at line 77 of file pfm_gen_analysis.cpp.

#define MAX_SIMPLE_SYM_LENGTH   300

Definition at line 74 of file pfm_gen_analysis.cpp.

#define MAX_SIMPLE_SYM_MOD_LENGTH   500

Definition at line 75 of file pfm_gen_analysis.cpp.

#define MAX_SP_STR_LENGTH   50

Definition at line 85 of file pfm_gen_analysis.cpp.

#define MAX_SYM_LENGTH   15000

Definition at line 70 of file pfm_gen_analysis.cpp.

#define MAX_SYM_MOD_LENGTH   20000

Definition at line 71 of file pfm_gen_analysis.cpp.

#define MAX_VALUE_STRING_LENGTH   250

Definition at line 81 of file pfm_gen_analysis.cpp.

#define PIPE_BUFFER_LENGTH   1000

Definition at line 87 of file pfm_gen_analysis.cpp.

Function Documentation

void calc_core_deriv_values ( double  totalCycles)

Definition at line 411 of file pfm_gen_analysis.cpp.

411  {
412  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
413  ++it ) {
414  ( it->second )["Total Cycles"] = ( it->second )["UNHALTED_CORE_CYCLES"];
415  ( it->second )["Stalled Cycles"] = ( it->second )["RS_UOPS_DISPATCHED CMASK=1 INV=1"];
416  ( it->second )["L2 Miss Impact"] = ( it->second )["MEM_LOAD_RETIRED:L2_LINE_MISS"] * CORE_L2_MISS_CYCLES;
417  ( it->second )["L2 Hit Impact"] =
418  ( ( it->second )["MEM_LOAD_RETIRED:L1D_LINE_MISS"] - ( it->second )["MEM_LOAD_RETIRED:L2_LINE_MISS"] ) *
419  CORE_L2_HIT_CYCLES;
420  ( it->second )["L1 DTLB Miss Impact"] = ( it->second )["MEM_LOAD_RETIRED:DTLB_MISS"] * CORE_L1_DTLB_MISS_CYCLES;
421  ( it->second )["LCP Stalls Impact"] = ( it->second )["ILD_STALL"] * CORE_LCP_STALL_CYCLES;
422  ( it->second )["Loads Blocked by Unknown Address Store Impact"] =
423  ( it->second )["LOAD_BLOCK:STA"] * CORE_UNKNOWN_ADDR_STORE_CYCLES;
424  ( it->second )["Loads Overlapped with Stores Impact"] =
425  ( it->second )["LOAD_BLOCK:OVERLAP_STORE"] * CORE_OVERLAPPING_CYCLES;
426  ( it->second )["Loads Spanning across Cache Lines Impact"] =
427  ( it->second )["LOAD_BLOCK:UNTIL_RETIRE"] * CORE_SPAN_ACROSS_CACHE_LINE_CYCLES;
428  ( it->second )["Store-Fwd Stalls Impact"] = ( it->second )["Loads Blocked by Unknown Address Store Impact"] +
429  ( it->second )["Loads Overlapped with Stores Impact"] +
430  ( it->second )["Loads Spanning across Cache Lines Impact"];
431  ( it->second )["Counted Stalled Cycles"] =
432  ( it->second )["L2 Miss Impact"] + ( it->second )["L2 Hit Impact"] + ( it->second )["LCP Stalls Impact"] +
433  ( it->second )["L1 DTLB Miss Impact"] + ( it->second )["Store-Fwd Stalls Impact"];
434  ( it->second )["Instructions Retired"] = ( it->second )["INSTRUCTIONS_RETIRED"];
435  ( it->second )["ITLB Miss Rate in %"] =
436  ( ( it->second )["ITLB_MISS_RETIRED"] / ( it->second )["INSTRUCTIONS_RETIRED"] ) * 100;
437  ( it->second )["Branch Instructions"] = ( it->second )["BRANCH_INSTRUCTIONS_RETIRED"];
438  ( it->second )["Load Instructions"] = ( it->second )["INST_RETIRED:LOADS"];
439  ( it->second )["Store Instructions"] = ( it->second )["INST_RETIRED:STORES"];
440  ( it->second )["Other Instructions"] = ( it->second )["INST_RETIRED:OTHER"] -
441  ( it->second )["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"] -
442  ( it->second )["BRANCH_INSTRUCTIONS_RETIRED"];
443  ( it->second )["% of Mispredicted Branches"] =
444  ( ( it->second )["MISPREDICTED_BRANCH_RETIRED"] / ( it->second )["BRANCH_INSTRUCTIONS_RETIRED"] ) * 100;
445  ( it->second )["Packed SIMD Computational Instructions"] =
446  ( it->second )["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"];
447  ( it->second )["Counted Instructions Retired"] =
448  ( it->second )["Branch Instructions"] + ( it->second )["Load Instructions"] +
449  ( it->second )["Store Instructions"] + ( it->second )["Other Instructions"] +
450  ( it->second )["Packed SIMD Computational Instructions"];
451  ( it->second )["CPI"] = ( it->second )["UNHALTED_CORE_CYCLES"] / ( it->second )["INSTRUCTIONS_RETIRED"];
452 
453  double localPerformanceImprovement = ( it->second )["CPI"] / EXPECTED_CPI;
454  double cyclesAfterImprovement = ( it->second )["UNHALTED_CORE_CYCLES"] / localPerformanceImprovement;
455  double totalCyclesAfterImprovement = totalCycles - ( it->second )["UNHALTED_CORE_CYCLES"] + cyclesAfterImprovement;
456  ( it->second )["iMargin"] = 100 - ( totalCyclesAfterImprovement / totalCycles ) * 100;
457 
458  ( it->second )["% of Total Cycles"] =
459  ( it->second )["RS_UOPS_DISPATCHED CMASK=1 INV=1"] * 100 / ( it->second )["UNHALTED_CORE_CYCLES"];
460  ( it->second )["L2 Miss % of counted Stalled Cycles"] =
461  ( it->second )["L2 Miss Impact"] * 100 / ( it->second )["Counted Stalled Cycles"];
462  ( it->second )["L2 Hit % of counted Stalled Cycles"] =
463  ( it->second )["L2 Hit Impact"] * 100 / ( it->second )["Counted Stalled Cycles"];
464  ( it->second )["L1 DTLB Miss % of counted Stalled Cycles"] =
465  ( it->second )["L1 DTLB Miss Impact"] * 100 / ( it->second )["Counted Stalled Cycles"];
466  ( it->second )["LCP Stalls % of counted Stalled Cycles"] =
467  ( it->second )["LCP Stalls Impact"] * 100 / ( it->second )["Counted Stalled Cycles"];
468  ( it->second )["Store-Fwd Stalls % of counted Stalled Cycles"] =
469  ( it->second )["Store-Fwd Stalls Impact"] * 100 / ( it->second )["Counted Stalled Cycles"];
470  ( it->second )["Loads Blocked % of Store-Fwd Stalls Cycles"] =
471  ( it->second )["Loads Blocked by Unknown Address Store Impact"] * 100 /
472  ( it->second )["Store-Fwd Stalls Impact"];
473  ( it->second )["Loads Overlapped % of Store-Fwd Stalls Cycles"] =
474  ( it->second )["Loads Overlapped with Stores Impact"] * 100 / ( it->second )["Store-Fwd Stalls Impact"];
475  ( it->second )["Loads Spanning % of Store-Fwd Stalls Cycles"] =
476  ( it->second )["Loads Spanning across Cache Lines Impact"] * 100 / ( it->second )["Store-Fwd Stalls Impact"];
477 
478  ( it->second )["Load % of all Instructions"] =
479  ( it->second )["INST_RETIRED:LOADS"] * 100 / ( it->second )["Counted Instructions Retired"];
480  ( it->second )["Store % of all Instructions"] =
481  ( it->second )["INST_RETIRED:STORES"] * 100 / ( it->second )["Counted Instructions Retired"];
482  ( it->second )["Branch % of all Instructions"] =
483  ( it->second )["BRANCH_INSTRUCTIONS_RETIRED"] * 100 / ( it->second )["Counted Instructions Retired"];
484  ( it->second )["Packed SIMD % of all Instructions"] =
485  ( it->second )["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"] * 100 /
486  ( it->second )["Counted Instructions Retired"];
487  ( it->second )["Other % of all Instructions"] =
488  ( it->second )["Other Instructions"] * 100 / ( it->second )["Counted Instructions Retired"];
489  }
490 }
CORE_UNKNOWN_ADDR_STORE_CYCLES
#define CORE_UNKNOWN_ADDR_STORE_CYCLES
Definition: pfm_gen_analysis.cpp:43
CORE_LCP_STALL_CYCLES
#define CORE_LCP_STALL_CYCLES
Definition: pfm_gen_analysis.cpp:42
EXPECTED_CPI
#define EXPECTED_CPI
Definition: pfm_gen_analysis.cpp:66
CORE_L1_DTLB_MISS_CYCLES
#define CORE_L1_DTLB_MISS_CYCLES
Definition: pfm_gen_analysis.cpp:41
CORE_OVERLAPPING_CYCLES
#define CORE_OVERLAPPING_CYCLES
Definition: pfm_gen_analysis.cpp:44
std::map
STL class.
std::string
STL class.
CORE_SPAN_ACROSS_CACHE_LINE_CYCLES
#define CORE_SPAN_ACROSS_CACHE_LINE_CYCLES
Definition: pfm_gen_analysis.cpp:45
CORE_L2_MISS_CYCLES
#define CORE_L2_MISS_CYCLES
Definition: pfm_gen_analysis.cpp:39
CORE_L2_HIT_CYCLES
#define CORE_L2_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:40
void calc_nhm_deriv_values ( double  totalCycles)

Definition at line 597 of file pfm_gen_analysis.cpp.

597  {
598  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
599  ++it ) {
600  ( it->second )["Total Cycles"] = ( it->second )["CPU_CLK_UNHALTED:THREAD_P"];
601 
602  ( it->second )["L2 Hit Impact"] = ( it->second )["MEM_LOAD_RETIRED:L2_HIT"] * I7_L2_HIT_CYCLES;
603  ( it->second )["L3 Unshared Hit Impact"] =
604  ( it->second )["MEM_LOAD_RETIRED:L3_UNSHARED_HIT"] * I7_L3_UNSHARED_HIT_CYCLES;
605  if ( ( it->second )["MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM"] >
606  ( it->second )["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"] ) {
607  ( it->second )["L2 Other Core Hit Impact"] = ( ( it->second )["MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM"] -
608  ( it->second )["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"] ) *
609  I7_OTHER_CORE_L2_HIT_CYCLES;
610  } else {
611  ( it->second )["L2 Other Core Hit Impact"] = 0.0;
612  }
613  ( it->second )["L2 Other Core Hit Modified Impact"] =
614  ( it->second )["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"] * I7_OTHER_CORE_L2_HITM_CYCLES;
615  ( it->second )["L3 Miss -> Local DRAM Hit Impact"] =
616  ( it->second )["MEM_UNCORE_RETIRED:LOCAL_DRAM"] * I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES;
617  ( it->second )["L3 Miss -> Remote DRAM Hit Impact"] =
618  ( it->second )["MEM_UNCORE_RETIRED:REMOTE_DRAM"] * I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES;
619  ( it->second )["L3 Miss -> Remote Cache Hit Impact"] =
620  ( it->second )["MEM_UNCORE_RETIRED:REMOTE_CACHE_LOCAL_HOME_HIT"] * I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES;
621  ( it->second )["L3 Miss -> Total Impact"] = ( it->second )["L3 Miss -> Local DRAM Hit Impact"] +
622  ( it->second )["L3 Miss -> Remote DRAM Hit Impact"] +
623  ( it->second )["L3 Miss -> Remote Cache Hit Impact"];
624  ( it->second )["L1 DTLB Miss Impact"] =
625  ( it->second )["DTLB_LOAD_MISSES:WALK_COMPLETED"] * I7_L1_DTLB_WALK_COMPLETED_CYCLES;
626  ( it->second )["Counted Stalled Cycles due to Load Ops"] =
627  ( it->second )["L3 Miss -> Total Impact"] + ( it->second )["L2 Hit Impact"] +
628  ( it->second )["L1 DTLB Miss Impact"] + ( it->second )["L3 Unshared Hit Impact"] +
629  ( it->second )["L2 Other Core Hit Modified Impact"] + ( it->second )["L2 Other Core Hit Impact"];
630  ( it->second )["Cycles spent during DIV & SQRT Ops"] = ( it->second )["ARITH:CYCLES_DIV_BUSY"];
631  ( it->second )["Total Counted Stalled Cycles"] =
632  ( it->second )["Counted Stalled Cycles due to Load Ops"] + ( it->second )["Cycles spent during DIV & SQRT Ops"];
633  ( it->second )["Stalled Cycles"] =
634  ( it->second )["Total Counted Stalled Cycles"]; // TO BE FIXED when UOPS_EXECUTED:0x3f is fixed!!
635  ( it->second )["% of Total Cycles"] =
636  ( it->second )["Stalled Cycles"] * 100 / ( it->second )["CPU_CLK_UNHALTED:THREAD_P"]; // TO BE FIXED!! see above
637  ( it->second )["L3 Miss % of Load Stalls"] =
638  ( it->second )["L3 Miss -> Total Impact"] * 100 / ( it->second )["Counted Stalled Cycles due to Load Ops"];
639  ( it->second )["L2 Hit % of Load Stalls"] =
640  ( it->second )["L2 Hit Impact"] * 100 / ( it->second )["Counted Stalled Cycles due to Load Ops"];
641  ( it->second )["L1 DTLB Miss % of Load Stalls"] =
642  ( it->second )["L1 DTLB Miss Impact"] * 100 / ( it->second )["Counted Stalled Cycles due to Load Ops"];
643  ( it->second )["L3 Unshared Hit % of Load Stalls"] =
644  ( it->second )["L3 Unshared Hit Impact"] * 100 / ( it->second )["Counted Stalled Cycles due to Load Ops"];
645  ( it->second )["L2 Other Core Hit % of Load Stalls"] =
646  ( it->second )["L2 Other Core Hit Impact"] * 100 / ( it->second )["Counted Stalled Cycles due to Load Ops"];
647  ( it->second )["L2 Other Core Hit Modified % of Load Stalls"] =
648  ( it->second )["L2 Other Core Hit Modified Impact"] * 100 /
649  ( it->second )["Counted Stalled Cycles due to Load Ops"];
650  ( it->second )["DIV & SQRT Ops % of counted Stalled Cycles"] =
651  ( it->second )["Cycles spent during DIV & SQRT Ops"] * 100 / ( it->second )["Total Counted Stalled Cycles"];
652 
653  ( it->second )["Cycles IFETCH served by Local DRAM"] =
654  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM"] * I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT;
655  ( it->second )["Cycles IFETCH served by L3 (Modified)"] =
656  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM"] * I7_IFETCH_L2_MISS_L3_HITM;
657  ( it->second )["Cycles IFETCH served by L3 (Clean Snoop)"] =
658  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP"] * I7_IFETCH_L2_MISS_L3_HIT_SNOOP;
659  ( it->second )["Cycles IFETCH served by Remote L2"] =
660  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD"] * I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD;
661  ( it->second )["Cycles IFETCH served by Remote DRAM"] =
662  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM"] * I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT;
663  ( it->second )["Cycles IFETCH served by L3 (No Snoop)"] =
664  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT"] * I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP;
665  ( it->second )["Total L2 IFETCH miss Impact"] =
666  ( it->second )["Cycles IFETCH served by Local DRAM"] + ( it->second )["Cycles IFETCH served by L3 (Modified)"] +
667  ( it->second )["Cycles IFETCH served by L3 (Clean Snoop)"] +
668  ( it->second )["Cycles IFETCH served by Remote L2"] + ( it->second )["Cycles IFETCH served by Remote DRAM"] +
669  ( it->second )["Cycles IFETCH served by L3 (No Snoop)"];
670  ( it->second )["Local DRAM IFECTHes % Impact"] =
671  ( it->second )["Cycles IFETCH served by Local DRAM"] * 100 / ( it->second )["Total L2 IFETCH miss Impact"];
672  ( it->second )["L3 (Modified) IFECTHes % Impact"] =
673  ( it->second )["Cycles IFETCH served by L3 (Modified)"] * 100 / ( it->second )["Total L2 IFETCH miss Impact"];
674  ( it->second )["L3 (Clean Snoop) IFECTHes % Impact"] = ( it->second )["Cycles IFETCH served by L3 (Clean Snoop)"] *
675  100 / ( it->second )["Total L2 IFETCH miss Impact"];
676  ( it->second )["Remote L2 IFECTHes % Impact"] =
677  ( it->second )["Cycles IFETCH served by Remote L2"] * 100 / ( it->second )["Total L2 IFETCH miss Impact"];
678  ( it->second )["Remote DRAM IFECTHes % Impact"] =
679  ( it->second )["Cycles IFETCH served by Remote DRAM"] * 100 / ( it->second )["Total L2 IFETCH miss Impact"];
680  ( it->second )["L3 (No Snoop) IFECTHes % Impact"] =
681  ( it->second )["Cycles IFETCH served by L3 (No Snoop)"] * 100 / ( it->second )["Total L2 IFETCH miss Impact"];
682  ( it->second )["Total L2 IFETCH misses"] = ( it->second )["L2_RQSTS:IFETCH_MISS"];
683  ( it->second )["% of IFETCHes served by Local DRAM"] =
684  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM"] * 100 / ( it->second )["L2_RQSTS:IFETCH_MISS"];
685  ( it->second )["% of IFETCHes served by L3 (Modified)"] =
686  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM"] * 100 / ( it->second )["L2_RQSTS:IFETCH_MISS"];
687  ( it->second )["% of IFETCHes served by L3 (Clean Snoop)"] =
688  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP"] * 100 /
689  ( it->second )["L2_RQSTS:IFETCH_MISS"];
690  ( it->second )["% of IFETCHes served by Remote L2"] =
691  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD"] * 100 /
692  ( it->second )["L2_RQSTS:IFETCH_MISS"];
693  ( it->second )["% of IFETCHes served by Remote DRAM"] =
694  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM"] * 100 / ( it->second )["L2_RQSTS:IFETCH_MISS"];
695  ( it->second )["% of IFETCHes served by L3 (No Snoop)"] =
696  ( it->second )["OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT"] * 100 / ( it->second )["L2_RQSTS:IFETCH_MISS"];
697  ( it->second )["% of L2 IFETCH misses"] =
698  ( it->second )["L2_RQSTS:IFETCH_MISS"] * 100 /
699  ( ( it->second )["L2_RQSTS:IFETCH_MISS"] + ( it->second )["L2_RQSTS:IFETCH_HIT"] );
700  ( it->second )["L1 ITLB Miss Impact"] =
701  ( it->second )["ITLB_MISSES:WALK_COMPLETED"] * I7_L1_ITLB_WALK_COMPLETED_CYCLES;
702 
703  ( it->second )["Total Branch Instructions Executed"] = ( it->second )["BR_INST_EXEC:ANY"];
704  ( it->second )["% of Mispredicted Branches"] =
705  ( it->second )["BR_MISP_EXEC:ANY"] * 100 / ( it->second )["BR_INST_EXEC:ANY"];
706  ( it->second )["Direct Near Calls % of Total Branches Executed"] =
707  ( it->second )["BR_INST_EXEC:DIRECT_NEAR_CALL"] * 100 / ( it->second )["Total Branch Instructions Executed"];
708  ( it->second )["Indirect Near Calls % of Total Branches Executed"] =
709  ( it->second )["BR_INST_EXEC:INDIRECT_NEAR_CALL"] * 100 / ( it->second )["Total Branch Instructions Executed"];
710  ( it->second )["Indirect Near Non-Calls % of Total Branches Executed"] =
711  ( it->second )["BR_INST_EXEC:INDIRECT_NON_CALL"] * 100 / ( it->second )["Total Branch Instructions Executed"];
712  ( it->second )["All Near Calls % of Total Branches Executed"] =
713  ( it->second )["BR_INST_EXEC:NEAR_CALLS"] * 100 / ( it->second )["Total Branch Instructions Executed"];
714  ( it->second )["All Non Calls % of Total Branches Executed"] =
715  ( it->second )["BR_INST_EXEC:NON_CALLS"] * 100 / ( it->second )["Total Branch Instructions Executed"];
716  ( it->second )["All Returns % of Total Branches Executed"] =
717  ( it->second )["BR_INST_EXEC:RETURN_NEAR"] * 100 / ( it->second )["Total Branch Instructions Executed"];
718  ( it->second )["Total Branch Instructions Retired"] = ( it->second )["BR_INST_RETIRED:ALL_BRANCHES"];
719  ( it->second )["Conditionals % of Total Branches Retired"] =
720  ( it->second )["BR_INST_RETIRED:CONDITIONAL"] * 100 / ( it->second )["Total Branch Instructions Retired"];
721  ( it->second )["Near Calls % of Total Branches Retired"] =
722  ( it->second )["BR_INST_RETIRED:NEAR_CALL"] * 100 / ( it->second )["Total Branch Instructions Retired"];
723 
724  ( it->second )["Instruction Starvation % of Total Cycles"] =
725  ( ( it->second )["UOPS_ISSUED:ANY CMASK=1 INV=1"] - ( it->second )["RESOURCE_STALLS:ANY"] ) * 100 /
726  ( it->second )["CPU_CLK_UNHALTED:THREAD_P"];
727  ( it->second )["% of Total Cycles spent handling FP exceptions"] =
728  ( it->second )["UOPS_DECODED:MS CMASK=1"] * 100 / ( it->second )["CPU_CLK_UNHALTED:THREAD_P"];
729  ( it->second )["# of Instructions per Call"] =
730  ( it->second )["INST_RETIRED:ANY_P"] / ( it->second )["BR_INST_EXEC:NEAR_CALLS"];
731 
732  ( it->second )["Instructions Retired"] = ( it->second )["INST_RETIRED:ANY_P"];
733  ( it->second )["ITLB Miss Rate in %"] =
734  ( ( it->second )["ITLB_MISS_RETIRED"] / ( it->second )["INST_RETIRED:ANY_P"] ) * 100;
735 
736  ( it->second )["Branch Instructions"] = ( it->second )["BR_INST_RETIRED:ALL_BRANCHES"];
737  ( it->second )["Load Instructions"] = ( it->second )["MEM_INST_RETIRED:LOADS"];
738  ( it->second )["Store Instructions"] = ( it->second )["MEM_INST_RETIRED:STORES"];
739  ( it->second )["Other Instructions"] =
740  ( it->second )["Instructions Retired"] - ( it->second )["MEM_INST_RETIRED:LOADS"] -
741  ( it->second )["MEM_INST_RETIRED:STORES"] - ( it->second )["BR_INST_RETIRED:ALL_BRANCHES"];
742  ( it->second )["Packed UOPS Retired"] =
743  ( it->second )["SSEX_UOPS_RETIRED:PACKED_DOUBLE"] + ( it->second )["SSEX_UOPS_RETIRED:PACKED_SINGLE"];
744  ( it->second )["CPI"] = ( it->second )["CPU_CLK_UNHALTED:THREAD_P"] / ( it->second )["INST_RETIRED:ANY_P"];
745 
746  double localPerformanceImprovement = ( it->second )["CPI"] / EXPECTED_CPI;
747  double cyclesAfterImprovement = ( it->second )["CPU_CLK_UNHALTED:THREAD_P"] / localPerformanceImprovement;
748  double totalCyclesAfterImprovement =
749  totalCycles - ( it->second )["CPU_CLK_UNHALTED:THREAD_P"] + cyclesAfterImprovement;
750  ( it->second )["iMargin"] = 100 - ( totalCyclesAfterImprovement / totalCycles ) * 100;
751 
752  ( it->second )["Load % of all Instructions"] =
753  ( it->second )["MEM_INST_RETIRED:LOADS"] * 100 / ( it->second )["INST_RETIRED:ANY_P"];
754  ( it->second )["Store % of all Instructions"] =
755  ( it->second )["MEM_INST_RETIRED:STORES"] * 100 / ( it->second )["INST_RETIRED:ANY_P"];
756  ( it->second )["Branch % of all Instructions"] =
757  ( it->second )["BR_INST_RETIRED:ALL_BRANCHES"] * 100 / ( it->second )["INST_RETIRED:ANY_P"];
758  ( it->second )["Other % of all Instructions"] =
759  ( it->second )["Other Instructions"] * 100 / ( it->second )["INST_RETIRED:ANY_P"];
760 
761  ( it->second )["Packed % of all UOPS Retired"] =
762  ( it->second )["Packed UOPS Retired"] * 100 / ( it->second )["UOPS_RETIRED:ANY"];
763  }
764 }
I7_L3_UNSHARED_HIT_CYCLES
#define I7_L3_UNSHARED_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:51
I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES
#define I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:55
I7_IFETCH_L2_MISS_L3_HITM
#define I7_IFETCH_L2_MISS_L3_HITM
Definition: pfm_gen_analysis.cpp:61
EXPECTED_CPI
#define EXPECTED_CPI
Definition: pfm_gen_analysis.cpp:66
I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD
#define I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD
Definition: pfm_gen_analysis.cpp:62
I7_OTHER_CORE_L2_HITM_CYCLES
#define I7_OTHER_CORE_L2_HITM_CYCLES
Definition: pfm_gen_analysis.cpp:53
std::map
STL class.
std::string
STL class.
I7_L1_ITLB_WALK_COMPLETED_CYCLES
#define I7_L1_ITLB_WALK_COMPLETED_CYCLES
Definition: pfm_gen_analysis.cpp:49
I7_L2_HIT_CYCLES
#define I7_L2_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:50
I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES
#define I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:54
I7_OTHER_CORE_L2_HIT_CYCLES
#define I7_OTHER_CORE_L2_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:52
I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES
#define I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES
Definition: pfm_gen_analysis.cpp:56
I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP
#define I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP
Definition: pfm_gen_analysis.cpp:59
I7_L1_DTLB_WALK_COMPLETED_CYCLES
#define I7_L1_DTLB_WALK_COMPLETED_CYCLES
Definition: pfm_gen_analysis.cpp:48
I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT
#define I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT
Definition: pfm_gen_analysis.cpp:57
I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT
#define I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT
Definition: pfm_gen_analysis.cpp:58
I7_IFETCH_L2_MISS_L3_HIT_SNOOP
#define I7_IFETCH_L2_MISS_L3_HIT_SNOOP
Definition: pfm_gen_analysis.cpp:60
void calc_post_deriv_values ( )

Definition at line 1554 of file pfm_gen_analysis.cpp.

1554  {
1555  if ( nehalem ) {
1556  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1557  ++it ) {
1558  double simdnorm =
1559  1. - normalize( "Packed % of all UOPS Retired", ( it->second )["Packed % of all UOPS Retired"], 1 );
1560  double misspnorm = normalize( "% of Mispredicted Branches", ( it->second )["% of Mispredicted Branches"], 1 );
1561  double stallnorm = normalize( "Stalled Cycles", ( it->second )["Stalled Cycles"], 1 );
1562  ( it->second )["iFactor"] = stallnorm * ( simdnorm + misspnorm + stallnorm );
1563  }
1564  } else {
1565  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1566  ++it ) {
1567  double simdnorm =
1568  1. - normalize( "Packed SIMD % of all Instructions", ( it->second )["Packed SIMD % of all Instructions"], 1 );
1569  double misspnorm = normalize( "% of Mispredicted Branches", ( it->second )["% of Mispredicted Branches"], 1 );
1570  double stallnorm = normalize( "Stalled Cycles", ( it->second )["Stalled Cycles"], 1 );
1571  ( it->second )["iFactor"] = stallnorm * ( simdnorm + misspnorm + stallnorm );
1572  }
1573  }
1574 }
normalize
double normalize(std::string field, double value, double normalizeTo)
Definition: pfm_gen_analysis.cpp:1535
std::map
STL class.
std::string
STL class.
bool check_for_core_caa_events ( )

Definition at line 342 of file pfm_gen_analysis.cpp.

342  {
343  for ( std::vector<std::string>::const_iterator it = core_caa_events.begin(); it != core_caa_events.end(); ++it ) {
344  if ( find( C_events.begin(), C_events.end(), ( *it ) ) == C_events.end() ) {
345  fprintf( stderr, "ERROR: Cannot find event %s!!!\naborting...\n", ( *it ).c_str() );
346  return false;
347  }
348  }
349  return true;
350 }
std::vector::end
T end(T...args)
std::find
T find(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
bool check_for_nhm_caa_events ( )

Definition at line 352 of file pfm_gen_analysis.cpp.

352  {
353  for ( std::vector<std::string>::const_iterator it = nhm_caa_events.begin(); it != nhm_caa_events.end(); ++it ) {
354  if ( find( C_events.begin(), C_events.end(), ( *it ) ) == C_events.end() ) {
355  fprintf( stderr, "ERROR: Cannot find event %s!!!\naborting...\n", ( *it ).c_str() );
356  return false;
357  }
358  }
359  return true;
360 }
std::vector::end
T end(T...args)
std::find
T find(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
int finalize_S_html_pages ( const char *  dir)

Definition at line 1347 of file pfm_gen_analysis.cpp.

1347  {
1348  for ( std::map<std::string, unsigned int>::const_iterator i = modules_tot_samples.begin();
1349  i != modules_tot_samples.end(); i++ ) {
1350  char module_filename[MAX_FILENAME_LENGTH];
1351  strcpy( module_filename, dir );
1352  strcat( module_filename, "/HTML/" );
1353  strcat( module_filename, ( i->first ).c_str() );
1354  strcat( module_filename, ".html" );
1355  FILE* module_file = fopen( module_filename, "a" );
1356  if ( module_file == NULL ) {
1357  fprintf( stderr, "ERROR: Unable to append to file: %s\naborting...\n", module_filename );
1358  exit( 1 );
1359  }
1360  fprintf( module_file, "</body>\n</html>\n" );
1361  if ( fclose( module_file ) ) {
1362  fprintf( stderr, "ERROR: Cannot close file %s!!!\naborting...\n", module_filename );
1363  exit( 1 );
1364  }
1365  }
1366  return 0;
1367 }
std::strcat
T strcat(T...args)
std::map::end
T end(T...args)
std::map
STL class.
std::fclose
T fclose(T...args)
std::exit
T exit(T...args)
std::fopen
T fopen(T...args)
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
std::map::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
const char* func_name ( const char *  demangled_symbol)

Definition at line 859 of file pfm_gen_analysis.cpp.

859  {
860  char* operator_string_begin = const_cast<char*>( strstr( demangled_symbol, "operator" ) );
861  if ( operator_string_begin != NULL ) {
862  char* operator_string_end = operator_string_begin + 8;
863  while ( *operator_string_end == ' ' ) operator_string_end++;
864  if ( strstr( operator_string_end, "delete[]" ) == operator_string_end ) {
865  operator_string_end += 8;
866  *operator_string_end = '\0';
867  } else if ( strstr( operator_string_end, "delete" ) == operator_string_end ) {
868  operator_string_end += 6;
869  *operator_string_end = '\0';
870  } else if ( strstr( operator_string_end, "new[]" ) == operator_string_end ) {
871  operator_string_end += 5;
872  *operator_string_end = '\0';
873  } else if ( strstr( operator_string_end, "new" ) == operator_string_end ) {
874  operator_string_end += 3;
875  *operator_string_end = '\0';
876  } else if ( strstr( operator_string_end, ">>=" ) == operator_string_end ) {
877  operator_string_end += 3;
878  *operator_string_end = '\0';
879  } else if ( strstr( operator_string_end, "<<=" ) == operator_string_end ) {
880  operator_string_end += 3;
881  *operator_string_end = '\0';
882  } else if ( strstr( operator_string_end, "->*" ) == operator_string_end ) {
883  operator_string_end += 3;
884  *operator_string_end = '\0';
885  } else if ( strstr( operator_string_end, "<<" ) == operator_string_end ) {
886  operator_string_end += 2;
887  *operator_string_end = '\0';
888  } else if ( strstr( operator_string_end, ">>" ) == operator_string_end ) {
889  operator_string_end += 2;
890  *operator_string_end = '\0';
891  } else if ( strstr( operator_string_end, ">=" ) == operator_string_end ) {
892  operator_string_end += 2;
893  *operator_string_end = '\0';
894  } else if ( strstr( operator_string_end, "<=" ) == operator_string_end ) {
895  operator_string_end += 2;
896  *operator_string_end = '\0';
897  } else if ( strstr( operator_string_end, "==" ) == operator_string_end ) {
898  operator_string_end += 2;
899  *operator_string_end = '\0';
900  } else if ( strstr( operator_string_end, "!=" ) == operator_string_end ) {
901  operator_string_end += 2;
902  *operator_string_end = '\0';
903  } else if ( strstr( operator_string_end, "|=" ) == operator_string_end ) {
904  operator_string_end += 2;
905  *operator_string_end = '\0';
906  } else if ( strstr( operator_string_end, "&=" ) == operator_string_end ) {
907  operator_string_end += 2;
908  *operator_string_end = '\0';
909  } else if ( strstr( operator_string_end, "^=" ) == operator_string_end ) {
910  operator_string_end += 2;
911  *operator_string_end = '\0';
912  } else if ( strstr( operator_string_end, "%=" ) == operator_string_end ) {
913  operator_string_end += 2;
914  *operator_string_end = '\0';
915  } else if ( strstr( operator_string_end, "/=" ) == operator_string_end ) {
916  operator_string_end += 2;
917  *operator_string_end = '\0';
918  } else if ( strstr( operator_string_end, "*=" ) == operator_string_end ) {
919  operator_string_end += 2;
920  *operator_string_end = '\0';
921  } else if ( strstr( operator_string_end, "-=" ) == operator_string_end ) {
922  operator_string_end += 2;
923  *operator_string_end = '\0';
924  } else if ( strstr( operator_string_end, "+=" ) == operator_string_end ) {
925  operator_string_end += 2;
926  *operator_string_end = '\0';
927  } else if ( strstr( operator_string_end, "&&" ) == operator_string_end ) {
928  operator_string_end += 2;
929  *operator_string_end = '\0';
930  } else if ( strstr( operator_string_end, "||" ) == operator_string_end ) {
931  operator_string_end += 2;
932  *operator_string_end = '\0';
933  } else if ( strstr( operator_string_end, "[]" ) == operator_string_end ) {
934  operator_string_end += 2;
935  *operator_string_end = '\0';
936  } else if ( strstr( operator_string_end, "()" ) == operator_string_end ) {
937  operator_string_end += 2;
938  *operator_string_end = '\0';
939  } else if ( strstr( operator_string_end, "++" ) == operator_string_end ) {
940  operator_string_end += 2;
941  *operator_string_end = '\0';
942  } else if ( strstr( operator_string_end, "--" ) == operator_string_end ) {
943  operator_string_end += 2;
944  *operator_string_end = '\0';
945  } else if ( strstr( operator_string_end, "->" ) == operator_string_end ) {
946  operator_string_end += 2;
947  *operator_string_end = '\0';
948  } else if ( strstr( operator_string_end, "<" ) == operator_string_end ) {
949  operator_string_end += 1;
950  *operator_string_end = '\0';
951  } else if ( strstr( operator_string_end, ">" ) == operator_string_end ) {
952  operator_string_end += 1;
953  *operator_string_end = '\0';
954  } else if ( strstr( operator_string_end, "~" ) == operator_string_end ) {
955  operator_string_end += 1;
956  *operator_string_end = '\0';
957  } else if ( strstr( operator_string_end, "!" ) == operator_string_end ) {
958  operator_string_end += 1;
959  *operator_string_end = '\0';
960  } else if ( strstr( operator_string_end, "+" ) == operator_string_end ) {
961  operator_string_end += 1;
962  *operator_string_end = '\0';
963  } else if ( strstr( operator_string_end, "-" ) == operator_string_end ) {
964  operator_string_end += 1;
965  *operator_string_end = '\0';
966  } else if ( strstr( operator_string_end, "*" ) == operator_string_end ) {
967  operator_string_end += 1;
968  *operator_string_end = '\0';
969  } else if ( strstr( operator_string_end, "/" ) == operator_string_end ) {
970  operator_string_end += 1;
971  *operator_string_end = '\0';
972  } else if ( strstr( operator_string_end, "%" ) == operator_string_end ) {
973  operator_string_end += 1;
974  *operator_string_end = '\0';
975  } else if ( strstr( operator_string_end, "^" ) == operator_string_end ) {
976  operator_string_end += 1;
977  *operator_string_end = '\0';
978  } else if ( strstr( operator_string_end, "&" ) == operator_string_end ) {
979  operator_string_end += 1;
980  *operator_string_end = '\0';
981  } else if ( strstr( operator_string_end, "|" ) == operator_string_end ) {
982  operator_string_end += 1;
983  *operator_string_end = '\0';
984  } else if ( strstr( operator_string_end, "," ) == operator_string_end ) {
985  operator_string_end += 1;
986  *operator_string_end = '\0';
987  } else if ( strstr( operator_string_end, "=" ) == operator_string_end ) {
988  operator_string_end += 1;
989  *operator_string_end = '\0';
990  }
991  return operator_string_begin;
992  }
993  char* end_of_demangled_name = const_cast<char*>( strrchr( demangled_symbol, ')' ) );
994  if ( end_of_demangled_name != NULL ) {
995  int pars = 1;
996  char c;
997  while ( pars > 0 && end_of_demangled_name != demangled_symbol ) {
998  c = *( --end_of_demangled_name );
999  if ( c == ')' ) {
1000  pars++;
1001  } else if ( c == '(' ) {
1002  pars--;
1003  }
1004  }
1005  } else {
1006  return demangled_symbol;
1007  }
1008  char* end_of_func_name = end_of_demangled_name;
1009  if ( end_of_func_name != NULL ) {
1010  *end_of_func_name = '\0';
1011  char c = *( --end_of_func_name );
1012  if ( c == '>' ) {
1013  int pars = 1;
1014  while ( pars > 0 && end_of_func_name != demangled_symbol ) {
1015  c = *( --end_of_func_name );
1016  if ( c == '>' ) {
1017  pars++;
1018  } else if ( c == '<' ) {
1019  pars--;
1020  }
1021  }
1022  *end_of_func_name = '\0';
1023  }
1024  c = *( --end_of_func_name );
1025  while ( isalnum( c ) || c == '_' || c == '~' ) { c = *( --end_of_func_name ); }
1026  return ++end_of_func_name;
1027  }
1028  return demangled_symbol;
1029 }
std::isalnum
T isalnum(T...args)
std::strrchr
T strrchr(T...args)
std::strstr
T strstr(T...args)
double getTotalCycles ( )

Definition at line 1580 of file pfm_gen_analysis.cpp.

1580  {
1581  double sum = 0;
1582  if ( nehalem ) {
1583  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1584  ++it ) {
1585  sum += ( it->second )["CPU_CLK_UNHALTED:THREAD_P"];
1586  }
1587  } else {
1588  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1589  ++it ) {
1590  sum += ( it->second )["UNHALTED_CORE_CYCLES"];
1591  }
1592  }
1593  return sum;
1594 }
sum
double sum(double x, double y, double z)
Definition: 3DoubleFuncTest.cpp:25
std::map
STL class.
std::string
STL class.
void html_special_chars ( const char *  s,
char *  s_mod 
)

Definition at line 828 of file pfm_gen_analysis.cpp.

828  {
829  int n = strlen( s );
830  *s_mod = '\0';
831  for ( int i = 0; i < n; i++ ) {
832  switch ( s[i] ) {
833  case '<':
834  strcat( s_mod, "&lt;" );
835  break;
836  case '>':
837  strcat( s_mod, "&gt;" );
838  break;
839  case '&':
840  strcat( s_mod, "&amp;" );
841  break;
842  case '"':
843  strcat( s_mod, "&quot;" );
844  break;
845  default:
846  char to_app[2];
847  to_app[0] = s[i];
848  to_app[1] = '\0';
849  strcat( s_mod, to_app );
850  break;
851  }
852  }
853  return;
854 }
std::strcat
T strcat(T...args)
std::strlen
T strlen(T...args)
gaudirun.s
string s
Definition: gaudirun.py:312
void init_core_caa_events ( )

Definition at line 274 of file pfm_gen_analysis.cpp.

274  {
275  core_caa_events.push_back( "BRANCH_INSTRUCTIONS_RETIRED" );
276  core_caa_events.push_back( "ILD_STALL" );
277  core_caa_events.push_back( "INST_RETIRED:LOADS" );
278  core_caa_events.push_back( "INST_RETIRED:OTHER" );
279  core_caa_events.push_back( "INST_RETIRED:STORES" );
280  core_caa_events.push_back( "INSTRUCTIONS_RETIRED" );
281  core_caa_events.push_back( "LOAD_BLOCK:OVERLAP_STORE" );
282  core_caa_events.push_back( "LOAD_BLOCK:STA" );
283  core_caa_events.push_back( "LOAD_BLOCK:UNTIL_RETIRE" );
284  core_caa_events.push_back( "MEM_LOAD_RETIRED:DTLB_MISS" );
285  core_caa_events.push_back( "MEM_LOAD_RETIRED:L1D_LINE_MISS" );
286  core_caa_events.push_back( "MEM_LOAD_RETIRED:L2_LINE_MISS" );
287  core_caa_events.push_back( "MISPREDICTED_BRANCH_RETIRED" );
288  // core_caa_events.push_back("RS_UOPS_DISPATCHED");
289  // core_caa_events.push_back("RS_UOPS_DISPATCHED CMASK=1");
290  core_caa_events.push_back( "RS_UOPS_DISPATCHED CMASK=1 INV=1" );
291  core_caa_events.push_back( "SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE" );
292  core_caa_events.push_back( "UNHALTED_CORE_CYCLES" );
293  // core_caa_events.push_back("UOPS_RETIRED:ANY");
294  // core_caa_events.push_back("UOPS_RETIRED:FUSED");
295  // core_caa_events.push_back("IDLE_DURING_DIV");
296 }
std::vector::push_back
T push_back(T...args)
void init_core_caa_events_displ ( )

Definition at line 362 of file pfm_gen_analysis.cpp.

362  {
363  core_caa_events_displ.push_back( "Total Cycles" );
364  core_caa_events_displ.push_back( "Stalled Cycles" );
365  core_caa_events_displ.push_back( "% of Total Cycles" );
366  core_caa_events_displ.push_back( "Instructions Retired" );
367  core_caa_events_displ.push_back( "CPI" );
368  core_caa_events_displ.push_back( "" );
369  core_caa_events_displ.push_back( "iMargin" );
370  core_caa_events_displ.push_back( "iFactor" );
371  core_caa_events_displ.push_back( "" );
372  core_caa_events_displ.push_back( "Counted Stalled Cycles" );
373  core_caa_events_displ.push_back( "" );
374  core_caa_events_displ.push_back( "L2 Miss Impact" );
375  core_caa_events_displ.push_back( "L2 Miss % of counted Stalled Cycles" );
376  core_caa_events_displ.push_back( "" );
377  core_caa_events_displ.push_back( "L2 Hit Impact" );
378  core_caa_events_displ.push_back( "L2 Hit % of counted Stalled Cycles" );
379  core_caa_events_displ.push_back( "" );
380  core_caa_events_displ.push_back( "L1 DTLB Miss Impact" );
381  core_caa_events_displ.push_back( "L1 DTLB Miss % of counted Stalled Cycles" );
382  core_caa_events_displ.push_back( "" );
383  core_caa_events_displ.push_back( "LCP Stalls Impact" );
384  core_caa_events_displ.push_back( "LCP Stalls % of counted Stalled Cycles" );
385  core_caa_events_displ.push_back( "" );
386  core_caa_events_displ.push_back( "Store-Fwd Stalls Impact" );
387  core_caa_events_displ.push_back( "Store-Fwd Stalls % of counted Stalled Cycles" );
388  core_caa_events_displ.push_back( "" );
389  core_caa_events_displ.push_back( "Loads Blocked by Unknown Address Store Impact" );
390  core_caa_events_displ.push_back( "Loads Blocked % of Store-Fwd Stalls Cycles" );
391  core_caa_events_displ.push_back( "Loads Overlapped with Stores Impact" );
392  core_caa_events_displ.push_back( "Loads Overlapped % of Store-Fwd Stalls Cycles" );
393  core_caa_events_displ.push_back( "Loads Spanning across Cache Lines Impact" );
394  core_caa_events_displ.push_back( "Loads Spanning % of Store-Fwd Stalls Cycles" );
395  core_caa_events_displ.push_back( "" );
396  core_caa_events_displ.push_back( "Load Instructions" );
397  core_caa_events_displ.push_back( "Load % of all Instructions" );
398  core_caa_events_displ.push_back( "Store Instructions" );
399  core_caa_events_displ.push_back( "Store % of all Instructions" );
400  core_caa_events_displ.push_back( "Branch Instructions" );
401  core_caa_events_displ.push_back( "Branch % of all Instructions" );
402  core_caa_events_displ.push_back( "Packed SIMD Computational Instructions" );
403  core_caa_events_displ.push_back( "Packed SIMD % of all Instructions" );
404  core_caa_events_displ.push_back( "Other Instructions" );
405  core_caa_events_displ.push_back( "Other % of all Instructions" );
406  core_caa_events_displ.push_back( "" );
407  core_caa_events_displ.push_back( "ITLB Miss Rate in %" );
408  core_caa_events_displ.push_back( "% of Mispredicted Branches" );
409 }
std::vector::push_back
T push_back(T...args)
void init_nhm_caa_events ( )

Definition at line 298 of file pfm_gen_analysis.cpp.

298  {
299  nhm_caa_events.push_back( "ARITH:CYCLES_DIV_BUSY" );
300  nhm_caa_events.push_back( "BR_INST_EXEC:ANY" );
301  nhm_caa_events.push_back( "BR_INST_EXEC:DIRECT_NEAR_CALL" );
302  nhm_caa_events.push_back( "BR_INST_EXEC:INDIRECT_NEAR_CALL" );
303  nhm_caa_events.push_back( "BR_INST_EXEC:INDIRECT_NON_CALL" );
304  nhm_caa_events.push_back( "BR_INST_EXEC:NEAR_CALLS" );
305  nhm_caa_events.push_back( "BR_INST_EXEC:NON_CALLS" );
306  nhm_caa_events.push_back( "BR_INST_EXEC:RETURN_NEAR" );
307  nhm_caa_events.push_back( "BR_INST_RETIRED:ALL_BRANCHES" );
308  nhm_caa_events.push_back( "BR_INST_RETIRED:CONDITIONAL" );
309  nhm_caa_events.push_back( "BR_INST_RETIRED:NEAR_CALL" );
310  nhm_caa_events.push_back( "BR_MISP_EXEC:ANY" );
311  nhm_caa_events.push_back( "CPU_CLK_UNHALTED:THREAD_P" );
312  nhm_caa_events.push_back( "DTLB_LOAD_MISSES:WALK_COMPLETED" );
313  nhm_caa_events.push_back( "INST_RETIRED:ANY_P" );
314  nhm_caa_events.push_back( "ITLB_MISSES:WALK_COMPLETED" );
315  nhm_caa_events.push_back( "L2_RQSTS:IFETCH_HIT" );
316  nhm_caa_events.push_back( "L2_RQSTS:IFETCH_MISS" );
317  nhm_caa_events.push_back( "MEM_INST_RETIRED:LOADS" );
318  nhm_caa_events.push_back( "MEM_INST_RETIRED:STORES" );
319  nhm_caa_events.push_back( "MEM_LOAD_RETIRED:L2_HIT" );
320  nhm_caa_events.push_back( "MEM_LOAD_RETIRED:L3_MISS" );
321  nhm_caa_events.push_back( "MEM_LOAD_RETIRED:L3_UNSHARED_HIT" );
322  nhm_caa_events.push_back( "MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM" );
323  nhm_caa_events.push_back( "MEM_UNCORE_RETIRED:LOCAL_DRAM" );
324  nhm_caa_events.push_back( "MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM" );
325  nhm_caa_events.push_back( "MEM_UNCORE_RETIRED:REMOTE_CACHE_LOCAL_HOME_HIT" );
326  nhm_caa_events.push_back( "MEM_UNCORE_RETIRED:REMOTE_DRAM" );
327  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM" );
328  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM" );
329  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP" );
330  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD" );
331  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM" );
332  nhm_caa_events.push_back( "OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT" );
333  nhm_caa_events.push_back( "RESOURCE_STALLS:ANY" );
334  nhm_caa_events.push_back( "SSEX_UOPS_RETIRED:PACKED_DOUBLE" );
335  nhm_caa_events.push_back( "SSEX_UOPS_RETIRED:PACKED_SINGLE" );
336  nhm_caa_events.push_back( "UOPS_DECODED:MS CMASK=1" );
337  nhm_caa_events.push_back( "UOPS_ISSUED:ANY CMASK=1 INV=1" );
338  nhm_caa_events.push_back( "ITLB_MISS_RETIRED" );
339  nhm_caa_events.push_back( "UOPS_RETIRED:ANY" );
340 }
std::vector::push_back
T push_back(T...args)
void init_nhm_caa_events_displ ( )

Definition at line 492 of file pfm_gen_analysis.cpp.

492  {
493  nhm_caa_events_displ.push_back( "Total Cycles" );
494  nhm_caa_events_displ.push_back( "Instructions Retired" );
495  nhm_caa_events_displ.push_back( "CPI" );
496  nhm_caa_events_displ.push_back( "" );
497  nhm_caa_events_displ.push_back( "iMargin" );
498  nhm_caa_events_displ.push_back( "iFactor" );
499  nhm_caa_events_displ.push_back( "" );
500  nhm_caa_events_displ.push_back( "Stalled Cycles" );
501  nhm_caa_events_displ.push_back( "% of Total Cycles" );
502  nhm_caa_events_displ.push_back( "Total Counted Stalled Cycles" );
503  nhm_caa_events_displ.push_back( "" );
504  nhm_caa_events_displ.push_back( "Instruction Starvation % of Total Cycles" );
505  nhm_caa_events_displ.push_back( "# of Instructions per Call" );
506  nhm_caa_events_displ.push_back( "% of Total Cycles spent handling FP exceptions" );
507  nhm_caa_events_displ.push_back( "" );
508  nhm_caa_events_displ.push_back( "Counted Stalled Cycles due to Load Ops" );
509  nhm_caa_events_displ.push_back( "" );
510  nhm_caa_events_displ.push_back( "L2 Hit Impact" );
511  nhm_caa_events_displ.push_back( "L2 Hit % of Load Stalls" );
512  nhm_caa_events_displ.push_back( "" );
513  nhm_caa_events_displ.push_back( "L3 Unshared Hit Impact" );
514  nhm_caa_events_displ.push_back( "L3 Unshared Hit % of Load Stalls" );
515  nhm_caa_events_displ.push_back( "" );
516  nhm_caa_events_displ.push_back( "L2 Other Core Hit Impact" );
517  nhm_caa_events_displ.push_back( "L2 Other Core Hit % of Load Stalls" );
518  nhm_caa_events_displ.push_back( "" );
519  nhm_caa_events_displ.push_back( "L2 Other Core Hit Modified Impact" );
520  nhm_caa_events_displ.push_back( "L2 Other Core Hit Modified % of Load Stalls" );
521  nhm_caa_events_displ.push_back( "" );
522  nhm_caa_events_displ.push_back( "L3 Miss -> Local DRAM Hit Impact" );
523  nhm_caa_events_displ.push_back( "L3 Miss -> Remote DRAM Hit Impact" );
524  nhm_caa_events_displ.push_back( "L3 Miss -> Remote Cache Hit Impact" );
525  nhm_caa_events_displ.push_back( "L3 Miss -> Total Impact" );
526  nhm_caa_events_displ.push_back( "L3 Miss % of Load Stalls" );
527  nhm_caa_events_displ.push_back( "" );
528  nhm_caa_events_displ.push_back( "L1 DTLB Miss Impact" );
529  nhm_caa_events_displ.push_back( "L1 DTLB Miss % of Load Stalls" );
530  nhm_caa_events_displ.push_back( "" );
531  nhm_caa_events_displ.push_back( "Cycles spent during DIV & SQRT Ops" );
532  nhm_caa_events_displ.push_back( "DIV & SQRT Ops % of counted Stalled Cycles" );
533  nhm_caa_events_displ.push_back( "" );
534  nhm_caa_events_displ.push_back( "Total L2 IFETCH misses" );
535  nhm_caa_events_displ.push_back( "% of L2 IFETCH misses" );
536  nhm_caa_events_displ.push_back( "" );
537  nhm_caa_events_displ.push_back( "% of IFETCHes served by Local DRAM" );
538  nhm_caa_events_displ.push_back( "% of IFETCHes served by L3 (Modified)" );
539  nhm_caa_events_displ.push_back( "% of IFETCHes served by L3 (Clean Snoop)" );
540  nhm_caa_events_displ.push_back( "% of IFETCHes served by Remote L2" );
541  nhm_caa_events_displ.push_back( "% of IFETCHes served by Remote DRAM" );
542  nhm_caa_events_displ.push_back( "% of IFETCHes served by L3 (No Snoop)" );
543  nhm_caa_events_displ.push_back( "" );
544  nhm_caa_events_displ.push_back( "Total L2 IFETCH miss Impact" );
545  nhm_caa_events_displ.push_back( "" );
546  nhm_caa_events_displ.push_back( "Cycles IFETCH served by Local DRAM" );
547  nhm_caa_events_displ.push_back( "Local DRAM IFECTHes % Impact" );
548  nhm_caa_events_displ.push_back( "" );
549  nhm_caa_events_displ.push_back( "Cycles IFETCH served by L3 (Modified)" );
550  nhm_caa_events_displ.push_back( "L3 (Modified) IFECTHes % Impact" );
551  nhm_caa_events_displ.push_back( "" );
552  nhm_caa_events_displ.push_back( "Cycles IFETCH served by L3 (Clean Snoop)" );
553  nhm_caa_events_displ.push_back( "L3 (Clean Snoop) IFECTHes % Impact" );
554  nhm_caa_events_displ.push_back( "" );
555  nhm_caa_events_displ.push_back( "Cycles IFETCH served by Remote L2" );
556  nhm_caa_events_displ.push_back( "Remote L2 IFECTHes % Impact" );
557  nhm_caa_events_displ.push_back( "" );
558  nhm_caa_events_displ.push_back( "Cycles IFETCH served by Remote DRAM" );
559  nhm_caa_events_displ.push_back( "Remote DRAM IFECTHes % Impact" );
560  nhm_caa_events_displ.push_back( "" );
561  nhm_caa_events_displ.push_back( "Cycles IFETCH served by L3 (No Snoop)" );
562  nhm_caa_events_displ.push_back( "L3 (No Snoop) IFECTHes % Impact" );
563  nhm_caa_events_displ.push_back( "" );
564  nhm_caa_events_displ.push_back( "Total Branch Instructions Executed" );
565  nhm_caa_events_displ.push_back( "% of Mispredicted Branches" );
566  nhm_caa_events_displ.push_back( "" );
567  nhm_caa_events_displ.push_back( "Direct Near Calls % of Total Branches Executed" );
568  nhm_caa_events_displ.push_back( "Indirect Near Calls % of Total Branches Executed" );
569  nhm_caa_events_displ.push_back( "Indirect Near Non-Calls % of Total Branches Executed" );
570  nhm_caa_events_displ.push_back( "All Near Calls % of Total Branches Executed" );
571  nhm_caa_events_displ.push_back( "All Non Calls % of Total Branches Executed" );
572  nhm_caa_events_displ.push_back( "All Returns % of Total Branches Executed" );
573  nhm_caa_events_displ.push_back( "" );
574  nhm_caa_events_displ.push_back( "Total Branch Instructions Retired" );
575  nhm_caa_events_displ.push_back( "Conditionals % of Total Branches Retired" );
576  nhm_caa_events_displ.push_back( "Near Calls % of Total Branches Retired" );
577  nhm_caa_events_displ.push_back( "" );
578  nhm_caa_events_displ.push_back( "L1 ITLB Miss Impact" );
579  nhm_caa_events_displ.push_back( "ITLB Miss Rate in %" );
580  nhm_caa_events_displ.push_back( "" );
581  nhm_caa_events_displ.push_back( "Branch Instructions" );
582  nhm_caa_events_displ.push_back( "Branch % of all Instructions" );
583  nhm_caa_events_displ.push_back( "" );
584  nhm_caa_events_displ.push_back( "Load Instructions" );
585  nhm_caa_events_displ.push_back( "Load % of all Instructions" );
586  nhm_caa_events_displ.push_back( "" );
587  nhm_caa_events_displ.push_back( "Store Instructions" );
588  nhm_caa_events_displ.push_back( "Store % of all Instructions" );
589  nhm_caa_events_displ.push_back( "" );
590  nhm_caa_events_displ.push_back( "Other Instructions" );
591  nhm_caa_events_displ.push_back( "Other % of all Instructions" );
592  nhm_caa_events_displ.push_back( "" );
593  nhm_caa_events_displ.push_back( "Packed UOPS Retired" );
594  nhm_caa_events_displ.push_back( "Packed % of all UOPS Retired" );
595 }
std::vector::push_back
T push_back(T...args)
int main ( int  argc,
char *  argv[] 
)

Definition at line 1599 of file pfm_gen_analysis.cpp.

1599  {
1600  if ( argc < 2 || argc > 4 ) {
1601  printf( "\n\nUsage: %s DIRECTORY [--caa] [--csv]\n\n", argv[0] );
1602  exit( 1 );
1603  }
1604 
1605  bool caa = false;
1606  bool csv = false;
1607  for ( int i = 2; i < argc; i++ ) {
1608  if ( !strcmp( argv[i], "--caa" ) ) caa = true;
1609  if ( !strcmp( argv[i], "--csv" ) ) csv = true;
1610  }
1611 
1612  char dir[MAX_FILENAME_LENGTH];
1613  strcpy( dir, argv[1] );
1614  if ( !csv ) {
1615  strcat( dir, "/HTML" );
1616  int res = mkdir( dir, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH );
1617  if ( res != 0 ) {
1618  fprintf( stderr, "ERROR: Cannot create directory %s\naborting...\n", dir );
1619  exit( 1 );
1620  }
1621  }
1622 
1623  DIR* dp;
1624  struct dirent* dirp;
1625  int num_of_modules = 0;
1626  if ( ( dp = opendir( argv[1] ) ) == NULL ) {
1627  printf( "Error(%d) opening %s\n", errno, argv[1] );
1628  return errno;
1629  }
1630  while ( ( dirp = readdir( dp ) ) != NULL ) {
1631  if ( strstr( dirp->d_name, "_S_" ) != NULL && strstr( dirp->d_name, ".txt.gz" ) != NULL && !csv ) {
1632  if ( read_S_events( argv[1], dirp->d_name ) ) {
1633  fprintf( stderr, "ERROR: Cannot read file %s\naborting...\n", dirp->d_name );
1634  exit( 1 );
1635  }
1636  }
1637  }
1638  closedir( dp );
1639  sort( S_events.begin(), S_events.end() );
1640  if ( ( dp = opendir( argv[1] ) ) == NULL ) {
1641  printf( "Error(%d) opening %s\n", errno, argv[1] );
1642  return errno;
1643  }
1644  while ( ( dirp = readdir( dp ) ) != NULL ) {
1645  if ( strstr( dirp->d_name, "_S_" ) != NULL && strstr( dirp->d_name, ".txt.gz" ) != NULL && !csv ) {
1646  if ( read_S_file( argv[1], dirp->d_name ) ) {
1647  fprintf( stderr, "ERROR: Cannot read file %s\naborting...\n", dirp->d_name );
1648  exit( 1 );
1649  }
1650  } else if ( strstr( dirp->d_name, "_C_" ) != NULL && strstr( dirp->d_name, ".txt" ) != NULL ) {
1651  int res = read_C_file( argv[1], dirp->d_name );
1652  if ( res > num_of_modules ) { num_of_modules = res; }
1653  }
1654  }
1655  closedir( dp );
1656 
1657  if ( !csv ) {
1658  if ( finalize_S_html_pages( argv[1] ) ) {
1659  fprintf( stderr, "ERROR: Cannot finalize HTML pages!!!\naborting...\n" );
1660  exit( 1 );
1661  }
1662  }
1663 
1664  char filepath[MAX_FILENAME_LENGTH];
1665  bzero( filepath, MAX_FILENAME_LENGTH );
1666  if ( !csv )
1667  sprintf( filepath, "%s/HTML/index.html", argv[1] );
1668  else
1669  sprintf( filepath, "%s/results.csv", argv[1] );
1670  FILE* fp = fopen( filepath, "w" );
1671  if ( fp == NULL ) {
1672  fprintf( stderr, "ERROR: Cannot create file index.html!!!\naborting...\n" );
1673  exit( 1 );
1674  }
1675 
1676  if ( caa ) {
1677  double totalCycles;
1678  if ( !nehalem ) {
1679  init_core_caa_events();
1680  if ( !check_for_core_caa_events() ) {
1681  fprintf( stderr, "(core) ERROR: One or more events for CAA missing!\naborting...\n" );
1682  exit( 1 );
1683  }
1684  init_core_caa_events_displ();
1685  totalCycles = getTotalCycles();
1686  calc_core_deriv_values( totalCycles );
1687  calc_post_deriv_values();
1688  if ( !csv ) {
1689  put_C_header( fp, core_caa_events_displ );
1690  put_C_modules( fp, core_caa_events_displ );
1691  } else {
1692  put_C_header_csv( fp, core_caa_events_displ );
1693  put_C_modules_csv( fp, core_caa_events_displ );
1694  }
1695  } else {
1696  init_nhm_caa_events();
1697  if ( !check_for_nhm_caa_events() ) {
1698  fprintf( stderr, "(nehalem) ERROR: One or more events for CAA missing!\naborting...\n" );
1699  exit( 1 );
1700  }
1701  init_nhm_caa_events_displ();
1702  totalCycles = getTotalCycles();
1703  calc_nhm_deriv_values( totalCycles );
1704  calc_post_deriv_values();
1705  if ( !csv ) {
1706  put_C_header( fp, nhm_caa_events_displ );
1707  put_C_modules( fp, nhm_caa_events_displ );
1708  } else {
1709  put_C_header_csv( fp, nhm_caa_events_displ );
1710  put_C_modules_csv( fp, nhm_caa_events_displ );
1711  }
1712  }
1713  if ( !csv ) put_C_footer( fp );
1714  fclose( fp );
1715  } else {
1716  if ( !csv ) {
1717  put_C_header( fp, C_events );
1718  put_C_modules( fp, C_events );
1719  put_C_footer( fp );
1720  } else {
1721  put_C_header_csv( fp, C_events );
1722  put_C_modules_csv( fp, C_events );
1723  }
1724  fclose( fp );
1725  }
1726  if ( !csv ) {
1727  char src[MAX_FILENAME_LENGTH];
1728  char dst[MAX_FILENAME_LENGTH];
1729  sprintf( src, "sorttable.js" );
1730  sprintf( dst, "%s/HTML/sorttable.js", argv[1] );
1731  int fd_src = open( src, O_RDONLY );
1732  if ( fd_src == -1 ) {
1733  fprintf( stderr, "ERROR: Cannot open file \"%s\"!\naborting...\n", src );
1734  exit( 1 );
1735  }
1736  int fd_dst = open( dst, O_WRONLY | O_CREAT | O_TRUNC, 0644 );
1737  if ( fd_dst == -1 ) {
1738  fprintf( stderr, "ERROR: Cannot open file \"%s\" (%s)!\naborting...\n", dst, strerror( errno ) );
1739  exit( 1 );
1740  }
1741  char c;
1742  while ( read( fd_src, &c, 1 ) ) {
1743  if ( write( fd_dst, &c, 1 ) == -1 ) {
1744  std::cerr << "ERROR: failed to write to " << dst << std::endl;
1745  exit( 1 );
1746  }
1747  }
1748  close( fd_dst );
1749  close( fd_src );
1750  }
1751  return 0;
1752 }
std::strcat
T strcat(T...args)
put_C_footer
void put_C_footer(FILE *fp)
Definition: pfm_gen_analysis.cpp:1494
std::strerror
T strerror(T...args)
init_core_caa_events_displ
void init_core_caa_events_displ()
Definition: pfm_gen_analysis.cpp:362
put_C_modules
void put_C_modules(FILE *fp, std::vector< std::string > &columns)
Definition: pfm_gen_analysis.cpp:1467
gaudirun.argv
list argv
Definition: gaudirun.py:294
std::endl
T endl(T...args)
init_core_caa_events
void init_core_caa_events()
Definition: pfm_gen_analysis.cpp:274
hivetimeline.read
def read(f, regex='.*', skipevents=0)
Definition: hivetimeline.py:22
read_S_file
int read_S_file(const char *dir, const char *filename)
Definition: pfm_gen_analysis.cpp:1188
check_for_nhm_caa_events
bool check_for_nhm_caa_events()
Definition: pfm_gen_analysis.cpp:352
std::vector::end
T end(T...args)
calc_post_deriv_values
void calc_post_deriv_values()
Definition: pfm_gen_analysis.cpp:1554
calc_nhm_deriv_values
void calc_nhm_deriv_values(double totalCycles)
Definition: pfm_gen_analysis.cpp:597
init_nhm_caa_events_displ
void init_nhm_caa_events_displ()
Definition: pfm_gen_analysis.cpp:492
finalize_S_html_pages
int finalize_S_html_pages(const char *dir)
Definition: pfm_gen_analysis.cpp:1347
std::fclose
T fclose(T...args)
put_C_header
void put_C_header(FILE *fp, std::vector< std::string > &columns)
Definition: pfm_gen_analysis.cpp:1438
init_nhm_caa_events
void init_nhm_caa_events()
Definition: pfm_gen_analysis.cpp:298
calc_core_deriv_values
void calc_core_deriv_values(double totalCycles)
Definition: pfm_gen_analysis.cpp:411
std::exit
T exit(T...args)
std::strcmp
T strcmp(T...args)
std::fopen
T fopen(T...args)
read_S_events
int read_S_events(const char *dir, const char *filename)
Definition: pfm_gen_analysis.cpp:1301
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
put_C_header_csv
void put_C_header_csv(FILE *fp, std::vector< std::string > &columns)
Definition: pfm_gen_analysis.cpp:1499
std::vector::begin
T begin(T...args)
check_for_core_caa_events
bool check_for_core_caa_events()
Definition: pfm_gen_analysis.cpp:342
std::strstr
T strstr(T...args)
getTotalCycles
double getTotalCycles()
Definition: pfm_gen_analysis.cpp:1580
std::sort
T sort(T...args)
put_C_modules_csv
void put_C_modules_csv(FILE *fp, std::vector< std::string > &columns)
Definition: pfm_gen_analysis.cpp:1510
std::printf
T printf(T...args)
read_C_file
int read_C_file(const char *dir, const char *filename)
Definition: pfm_gen_analysis.cpp:1373
double normalize ( std::string  field,
double  value,
double  normalizeTo 
)

Definition at line 1535 of file pfm_gen_analysis.cpp.

1535  {
1536  double max = 0;
1537  double counter_value;
1538  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1539  ++it ) {
1540  counter_value = ( it->second )[field];
1541  if ( max < counter_value ) max = counter_value;
1542  }
1543  if ( value > 0 && max > 0 && normalizeTo > 0 ) {
1544  return 1. * value / max * normalizeTo;
1545  } else
1546  return 0;
1547 }
max
EventIDBase max(const EventIDBase &lhs, const EventIDBase &rhs)
Definition: EventIDBase.h:215
std::map
STL class.
std::string
STL class.
void put_C_footer ( FILE *  fp)

Definition at line 1494 of file pfm_gen_analysis.cpp.

1494  {
1495  fprintf( fp, "</table>\n</body>\n</html>\n" );
1496  return;
1497 }
std::fprintf
T fprintf(T...args)
void put_C_header ( FILE *  fp,
std::vector< std::string > &  columns 
)

Definition at line 1438 of file pfm_gen_analysis.cpp.

1438  {
1439  fprintf(
1440  fp,
1441  "<!DOCTYPE html PUBLIC \"-//W3C//DTD HTML 4.01 Transitional//EN\" \"http://www.w3.org/TR/html4/loose.dtd\">\n" );
1442  fprintf( fp, "<html>\n" );
1443  fprintf( fp, "<head>\n" );
1444  fprintf( fp, "<title>\n" );
1445  fprintf( fp, "Analysis Result\n" );
1446  fprintf( fp, "</title>\n" );
1447  fprintf( fp, "<script src=\"sorttable.js\"></script>\n" );
1448  fprintf( fp, "<style>\ntable.sortable thead "
1449  "{\nbackground-color:#eee;\ncolor:#666666;\nfont-weight:bold;\ncursor:default;\nfont-family:courier;\n}"
1450  "\n</style>\n" );
1451  fprintf( fp, "</head>\n" );
1452  fprintf( fp, "<body link=\"black\">\n" );
1453  fprintf( fp, "<h1>RESULTS:</h1><br/>Click for detailed symbol view...<p/>\n" );
1454  fprintf( fp, "<table class=\"sortable\" cellpadding=\"5\">\n" );
1455  fprintf( fp, "<tr>\n" );
1456  fprintf( fp, "<th>MODULE NAME</th>\n" );
1457  for ( std::vector<std::string>::const_iterator it = columns.begin(); it != columns.end(); ++it ) {
1458  if ( strlen( it->c_str() ) == 0 )
1459  fprintf( fp, "<th bgcolor=\"#FFFFFF\">&nbsp;</th>\n" );
1460  else
1461  fprintf( fp, "<th>%s</th>\n", ( *it ).c_str() );
1462  }
1463  fprintf( fp, "</tr>\n" );
1464  return;
1465 }
std::vector::end
T end(T...args)
std::strlen
T strlen(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
void put_C_header_csv ( FILE *  fp,
std::vector< std::string > &  columns 
)

Definition at line 1499 of file pfm_gen_analysis.cpp.

1499  {
1500  fprintf( fp, "MODULE NAME" );
1501  for ( std::vector<std::string>::const_iterator it = columns.begin(); it != columns.end(); ++it ) {
1502  if ( strlen( it->c_str() ) == 0 ) {
1503  } else
1504  fprintf( fp, ",%s", ( *it ).c_str() );
1505  }
1506  fprintf( fp, "\n" );
1507  return;
1508 }
std::vector::end
T end(T...args)
std::strlen
T strlen(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
void put_C_modules ( FILE *  fp,
std::vector< std::string > &  columns 
)

Definition at line 1467 of file pfm_gen_analysis.cpp.

1467  {
1468  int index = 0;
1469  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1470  ++it ) {
1471  if ( index % 2 )
1472  fprintf( fp, "<tr bgcolor=\"#FFFFCC\">\n" );
1473  else
1474  fprintf( fp, "<tr bgcolor=\"#CCFFCC\">\n" );
1475  fprintf( fp, "<td style=\"font-family:monospace;font-size:large;color:Black\"><a href=\"%s.html\">%s</a></td>\n",
1476  ( it->first ).c_str(), ( it->first ).c_str() );
1477  for ( std::vector<std::string>::const_iterator jt = columns.begin(); jt != columns.end(); ++jt ) {
1478  if ( strlen( jt->c_str() ) == 0 ) {
1479  fprintf( fp, "<td bgcolor=\"#FFFFFF\">&nbsp;</td>" );
1480  } else {
1481  if ( ( it->second ).find( *jt ) == ( it->second ).end() ) {
1482  fprintf( stderr, "ERROR: Cannot find derivate value \"%s\"!!!\naborting...\n", ( *jt ).c_str() );
1483  exit( 1 );
1484  }
1485  fprintf( fp, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\" align=\"right\">%.2f</td>\n",
1486  ( it->second )[*jt] );
1487  }
1488  }
1489  fprintf( fp, "</tr>\n" );
1490  index++;
1491  }
1492 }
std::vector::end
T end(T...args)
std::map
STL class.
std::string
STL class.
std::exit
T exit(T...args)
std::strlen
T strlen(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
void put_C_modules_csv ( FILE *  fp,
std::vector< std::string > &  columns 
)

Definition at line 1510 of file pfm_gen_analysis.cpp.

1510  {
1511  for ( std::map<std::string, std::map<std::string, double>>::iterator it = C_modules.begin(); it != C_modules.end();
1512  ++it ) {
1513  fprintf( fp, "%s", ( it->first ).c_str() );
1514  for ( std::vector<std::string>::const_iterator jt = columns.begin(); jt != columns.end(); ++jt ) {
1515  if ( strlen( jt->c_str() ) == 0 ) {
1516  } else {
1517  if ( ( it->second ).find( *jt ) == ( it->second ).end() ) {
1518  fprintf( stderr, "ERROR: Cannot find derivate value \"%s\"!!!\naborting...\n", ( *jt ).c_str() );
1519  exit( 1 );
1520  }
1521  fprintf( fp, ",%.2f", ( it->second )[*jt] );
1522  }
1523  }
1524  fprintf( fp, "\n" );
1525  }
1526 }
std::vector::end
T end(T...args)
std::map
STL class.
std::string
STL class.
std::exit
T exit(T...args)
std::strlen
T strlen(T...args)
std::vector
STL class.
std::vector::begin
T begin(T...args)
std::fprintf
T fprintf(T...args)
void put_S_module ( S_module *  cur_module,
const char *  dir 
)

Definition at line 1036 of file pfm_gen_analysis.cpp.

1036  {
1037  char module_name[MAX_MODULE_NAME_LENGTH];
1038  bzero( module_name, MAX_MODULE_NAME_LENGTH );
1039  strcpy( module_name, ( cur_module->get_module_name() ).c_str() );
1040  char module_filename[MAX_FILENAME_LENGTH];
1041  bzero( module_filename, MAX_FILENAME_LENGTH );
1042  strcpy( module_filename, dir );
1043  strcat( module_filename, "/HTML/" );
1044  strcat( module_filename, module_name );
1045  strcat( module_filename, ".html" );
1046  char event[MAX_EVENT_NAME_LENGTH];
1047  bzero( event, MAX_EVENT_NAME_LENGTH );
1048  strcpy( event, ( cur_module->get_event() ).c_str() );
1049  std::map<std::string, unsigned int>::iterator result = modules_tot_samples.find( cur_module->get_module_name() );
1050  FILE* module_file;
1051  if ( result == modules_tot_samples.end() ) // not found
1052  {
1053  if ( ( !strcmp( event, "UNHALTED_CORE_CYCLES" ) && !nehalem ) ||
1054  ( !strcmp( event, "CPU_CLK_UNHALTED:THREAD_P" ) && nehalem ) ) {
1055  modules_tot_samples.insert(
1056  std::pair<std::string, unsigned int>( cur_module->get_module_name(), cur_module->get_total_num_samples() ) );
1057  } else {
1058  modules_tot_samples.insert( std::pair<std::string, unsigned int>( cur_module->get_module_name(), 0 ) );
1059  }
1060  module_file = fopen( module_filename, "w" );
1061  if ( module_file == NULL ) {
1062  fprintf( stderr, "ERROR: Cannot create file %s!!!\naborting...\n", module_filename );
1063  exit( 1 );
1064  }
1065  fprintf( module_file, "<!DOCTYPE html PUBLIC \"-//W3C//DTD HTML 4.01 Transitional//EN\" "
1066  "\"http://www.w3.org/TR/html4/loose.dtd\">\n" );
1067  fprintf( module_file, "<html>\n" );
1068  fprintf( module_file, "<head>\n" );
1069  fprintf( module_file, "<title>\n" );
1070  fprintf( module_file, "%s\n", module_name );
1071  fprintf( module_file, "</title>\n" );
1072  fprintf( module_file, "</head>\n" );
1073  fprintf( module_file, "<body>\n" );
1074  fprintf( module_file, "<h2>%s</h2><br/>Events Sampled:<br/>\n", module_name );
1075  fprintf( module_file, "<ul>\n" );
1076  for ( std::vector<std::string>::const_iterator it = S_events.begin(); it != S_events.end(); ++it ) {
1077  fprintf( module_file, "<li><a href=\"#%s\">%s</a></li>\n", it->c_str(), it->c_str() );
1078  }
1079  fprintf( module_file, "</ul>\n" );
1080  } // if(result == modules_tot_samples.end()) //not found
1081  else {
1082  if ( ( !strcmp( event, "UNHALTED_CORE_CYCLES" ) && !nehalem ) ||
1083  ( !strcmp( event, "CPU_CLK_UNHALTED:THREAD_P" ) && nehalem ) ) {
1084  modules_tot_samples[cur_module->get_module_name()] = cur_module->get_total_num_samples();
1085  }
1086  module_file = fopen( module_filename, "a" );
1087  } // else:: if(result != modules_tot_samples.end()) //found!!
1088  char event_str[MAX_EVENT_NAME_LENGTH];
1089  bzero( event_str, MAX_EVENT_NAME_LENGTH );
1090  strcpy( event_str, event );
1091  if ( cur_module->get_c_mask() > 0 ) {
1092  sprintf( event_str + strlen( event_str ), " CMASK=%d", cur_module->get_c_mask() );
1093  }
1094  if ( cur_module->get_inv_mask() > 0 ) {
1095  sprintf( event_str + strlen( event_str ), " INV=%d", cur_module->get_inv_mask() );
1096  }
1097  fprintf( module_file, "<a name=\"%s\"><a>\n", event_str );
1098  fprintf( module_file, "<table cellpadding=\"5\">\n" );
1099  fprintf( module_file, "<tr bgcolor=\"#EEEEEE\">\n" );
1100  fprintf( module_file,
1101  "<th colspan=\"6\" align=\"left\">%s -- cmask: %u -- invmask: %u -- Total Samples: %u -- "
1102  "Sampling Period: %d</th>\n",
1103  event, cur_module->get_c_mask(), cur_module->get_inv_mask(), cur_module->get_total_num_samples(),
1104  cur_module->get_smpl_period() );
1105  fprintf( module_file, "</tr>\n" );
1106  fprintf( module_file, "<tr bgcolor=\"#EEEEEE\">\n" );
1107  fprintf( module_file, "<th align=\"left\">Samples</th>\n" );
1108  fprintf( module_file, "<th align=\"left\">Percentage</th>\n" );
1109  fprintf( module_file, "<th align=\"left\">Symbol Name</th>\n" );
1110  fprintf( module_file, "<th align=\"left\">Library Name</th>\n" );
1111  fprintf( module_file, "<th align=\"left\">Complete Signature</th>\n" );
1112  fprintf( module_file, "<th align=\"left\">Library Pathname</th>\n" );
1113  fprintf( module_file, "</tr>\n" );
1114  for ( int j = 0; j < 20; j++ ) {
1115  char sym[MAX_SYM_LENGTH];
1116  char sym_mod[MAX_SYM_MOD_LENGTH];
1117  char lib[MAX_LIB_LENGTH];
1118  char lib_mod[MAX_LIB_MOD_LENGTH];
1119  char simple_sym[MAX_SIMPLE_SYM_LENGTH];
1120  char simple_sym_mod[MAX_SIMPLE_SYM_MOD_LENGTH];
1121  char simple_lib[MAX_SIMPLE_LIB_LENGTH];
1122  char simple_lib_mod[MAX_SIMPLE_LIB_MOD_LENGTH];
1123 
1124  bzero( sym, MAX_SYM_LENGTH );
1125  bzero( sym_mod, MAX_SYM_MOD_LENGTH );
1126  bzero( lib, MAX_LIB_LENGTH );
1127  bzero( lib_mod, MAX_LIB_MOD_LENGTH );
1128  bzero( simple_sym, MAX_SIMPLE_SYM_LENGTH );
1129  bzero( simple_sym_mod, MAX_SIMPLE_SYM_MOD_LENGTH );
1130  bzero( simple_lib, MAX_SIMPLE_LIB_LENGTH );
1131  bzero( simple_lib_mod, MAX_SIMPLE_LIB_MOD_LENGTH );
1132 
1133  char index[MAX_SAMPLE_INDEX_LENGTH];
1134  bzero( index, MAX_SAMPLE_INDEX_LENGTH );
1135  unsigned int value;
1136  bool res = cur_module->get_max( index, value );
1137  if ( !res ) break;
1138  char* sym_end = strchr( index, '%' );
1139  if ( sym_end == NULL ) // error
1140  {
1141  fprintf( stderr, "ERROR: Invalid sym and lib name! : %s\naborting...\n", index );
1142  exit( 1 );
1143  }
1144  memcpy( sym, index, strlen( index ) - strlen( sym_end ) );
1145  strcpy( lib, sym_end + 1 );
1146  char temp[MAX_SYM_LENGTH];
1147  bzero( temp, MAX_SYM_LENGTH );
1148  strcpy( temp, sym );
1149  strcpy( simple_sym, ( func_name( temp ) ) );
1150  if ( strrchr( lib, '/' ) != NULL && *( strrchr( lib, '/' ) + 1 ) != '\0' ) {
1151  strcpy( simple_lib, strrchr( lib, '/' ) + 1 );
1152  } else {
1153  strcpy( simple_lib, lib );
1154  }
1155  if ( j % 2 != 0 ) {
1156  fprintf( module_file, "<tr bgcolor=\"#FFFFCC\">\n" );
1157  } else {
1158  fprintf( module_file, "<tr bgcolor=\"#CCFFCC\">\n" );
1159  }
1160  fprintf( module_file, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\">%u</td>\n", value );
1161  fprintf( module_file, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\">%f%%</td>\n",
1162  ( ( (double)( value ) ) / ( (double)( cur_module->get_total_num_samples() ) ) ) * 100 );
1163  html_special_chars( simple_sym, simple_sym_mod );
1164  html_special_chars( simple_lib, simple_lib_mod );
1165  html_special_chars( sym, sym_mod );
1166  html_special_chars( lib, lib_mod );
1167  fprintf( module_file, "<td style=\"font-family:courier;\">%s</td>\n", simple_sym_mod );
1168  fprintf( module_file, "<td style=\"font-family:courier;\">%s</td>\n", simple_lib_mod );
1169  fprintf( module_file, "<td style=\"font-family:courier;\">%s</td>\n", sym_mod );
1170  fprintf( module_file, "<td style=\"font-family:courier;\">%s</td>\n</tr>\n", lib_mod );
1171  }
1172  fprintf( module_file, "</table><br/><br/>\n" );
1173  int res = fclose( module_file );
1174  if ( res ) {
1175  fprintf( stderr, "ERROR: Cannot close file %s!!!\naborting...\n", module_filename );
1176  exit( 1 );
1177  }
1178  return;
1179 }
MAX_EVENT_NAME_LENGTH
#define MAX_EVENT_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:79
S_module::get_inv_mask
unsigned int get_inv_mask()
Definition: pfm_gen_analysis.cpp:803
func_name
const char * func_name(const char *demangled_symbol)
Definition: pfm_gen_analysis.cpp:859
MAX_SAMPLE_INDEX_LENGTH
#define MAX_SAMPLE_INDEX_LENGTH
Definition: pfm_gen_analysis.cpp:69
std::strcat
T strcat(T...args)
MAX_LIB_LENGTH
#define MAX_LIB_LENGTH
Definition: pfm_gen_analysis.cpp:72
MAX_SYM_LENGTH
#define MAX_SYM_LENGTH
Definition: pfm_gen_analysis.cpp:70
std::map::end
T end(T...args)
S_module::get_smpl_period
unsigned int get_smpl_period()
Definition: pfm_gen_analysis.cpp:802
std::map
STL class.
MAX_SIMPLE_SYM_LENGTH
#define MAX_SIMPLE_SYM_LENGTH
Definition: pfm_gen_analysis.cpp:74
MAX_LIB_MOD_LENGTH
#define MAX_LIB_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:73
std::fclose
T fclose(T...args)
MAX_SIMPLE_LIB_MOD_LENGTH
#define MAX_SIMPLE_LIB_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:77
std::exit
T exit(T...args)
std::memcpy
T memcpy(T...args)
std::strlen
T strlen(T...args)
html_special_chars
void html_special_chars(const char *s, char *s_mod)
Definition: pfm_gen_analysis.cpp:828
std::strcmp
T strcmp(T...args)
std::fopen
T fopen(T...args)
std::strrchr
T strrchr(T...args)
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
std::map::insert
T insert(T...args)
std::map::find
T find(T...args)
std::vector
STL class.
MAX_SIMPLE_SYM_MOD_LENGTH
#define MAX_SIMPLE_SYM_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:75
std::vector::begin
T begin(T...args)
S_module::get_event
std::string get_event()
Definition: pfm_gen_analysis.cpp:806
S_module::get_total_num_samples
unsigned int get_total_num_samples()
Definition: pfm_gen_analysis.cpp:821
MAX_MODULE_NAME_LENGTH
#define MAX_MODULE_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:80
MAX_SIMPLE_LIB_LENGTH
#define MAX_SIMPLE_LIB_LENGTH
Definition: pfm_gen_analysis.cpp:76
std::fprintf
T fprintf(T...args)
S_module::get_max
bool get_max(char *index, unsigned int &value)
Definition: pfm_gen_analysis.cpp:811
S_module::get_module_name
std::string get_module_name()
Definition: pfm_gen_analysis.cpp:820
MAX_SYM_MOD_LENGTH
#define MAX_SYM_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:71
std::strchr
T strchr(T...args)
S_module::get_c_mask
unsigned int get_c_mask()
Definition: pfm_gen_analysis.cpp:804
int read_C_file ( const char *  dir,
const char *  filename 
)

Definition at line 1373 of file pfm_gen_analysis.cpp.

1373  {
1374  char event[MAX_EVENT_NAME_LENGTH];
1375  char arch[MAX_ARCH_NAME_LENGTH];
1376  char line[MAX_LINE_LENGTH];
1377  char cmask_str[MAX_CMASK_STR_LENGTH];
1378  char inv_str[MAX_INV_STR_LENGTH];
1379  char sp_str[MAX_SP_STR_LENGTH];
1380  char cur_module_name[MAX_MODULE_NAME_LENGTH];
1381  bzero( line, MAX_LINE_LENGTH );
1382  bzero( event, MAX_EVENT_NAME_LENGTH );
1383  bzero( cur_module_name, MAX_MODULE_NAME_LENGTH );
1384  bzero( arch, MAX_ARCH_NAME_LENGTH );
1385  bzero( line, MAX_LINE_LENGTH );
1386  bzero( cmask_str, MAX_CMASK_STR_LENGTH );
1387  bzero( inv_str, MAX_INV_STR_LENGTH );
1388  bzero( sp_str, MAX_SP_STR_LENGTH );
1389  int number_of_modules = 0;
1390  long cur_sum = 0;
1391  int no_of_values = 0;
1392  char path_name[MAX_FILENAME_LENGTH];
1393  bzero( path_name, MAX_FILENAME_LENGTH );
1394  strcpy( path_name, dir );
1395  strcat( path_name, "/" );
1396  strcat( path_name, filename );
1397  FILE* fp = fopen( path_name, "r" );
1398  int stat = fscanf( fp, "%s %s %s %s %s\n", arch, event, cmask_str, inv_str, sp_str );
1399  if ( stat != 5 ) {
1400  std::cerr << "ERROR: failed to parse " << path_name << std::endl;
1401  exit( 1 );
1402  }
1403  if ( !strcmp( arch, "NHM" ) )
1404  nehalem = true;
1405  else
1406  nehalem = false;
1407  std::string event_str( event );
1408  if ( atoi( cmask_str ) > 0 ) {
1409  event_str += " CMASK=";
1410  event_str += cmask_str;
1411  }
1412  if ( atoi( inv_str ) > 0 ) {
1413  event_str += " INV=";
1414  event_str += inv_str;
1415  }
1416  C_events.push_back( event_str );
1417  while ( fscanf( fp, "%s\n", line ) != EOF ) {
1418  if ( isalpha( line[0] ) ) // module
1419  {
1420  if ( number_of_modules > 0 ) {
1421  C_modules[cur_module_name][event_str] = (double)cur_sum / no_of_values;
1422  cur_sum = 0;
1423  no_of_values = 0;
1424  }
1425  strcpy( cur_module_name, line );
1426  number_of_modules++;
1427  } else if ( isdigit( line[0] ) ) // value
1428  {
1429  cur_sum += strtol( line, NULL, 10 );
1430  no_of_values++;
1431  }
1432  }
1433  C_modules[cur_module_name][event_str] = (double)cur_sum / no_of_values; // last module
1434  fclose( fp );
1435  return number_of_modules;
1436 }
MAX_EVENT_NAME_LENGTH
#define MAX_EVENT_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:79
std::atoi
T atoi(T...args)
std::strcat
T strcat(T...args)
std::isalpha
T isalpha(T...args)
std::fscanf
T fscanf(T...args)
std::endl
T endl(T...args)
MAX_ARCH_NAME_LENGTH
#define MAX_ARCH_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:82
MAX_LINE_LENGTH
#define MAX_LINE_LENGTH
Definition: pfm_gen_analysis.cpp:78
std::fclose
T fclose(T...args)
std::string
STL class.
std::vector::push_back
T push_back(T...args)
MAX_SP_STR_LENGTH
#define MAX_SP_STR_LENGTH
Definition: pfm_gen_analysis.cpp:85
std::exit
T exit(T...args)
std::isdigit
T isdigit(T...args)
std::strcmp
T strcmp(T...args)
std::fopen
T fopen(T...args)
MAX_CMASK_STR_LENGTH
#define MAX_CMASK_STR_LENGTH
Definition: pfm_gen_analysis.cpp:83
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
MAX_INV_STR_LENGTH
#define MAX_INV_STR_LENGTH
Definition: pfm_gen_analysis.cpp:84
std::strtol
T strtol(T...args)
MAX_MODULE_NAME_LENGTH
#define MAX_MODULE_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:80
int read_S_events ( const char *  dir,
const char *  filename 
)

Definition at line 1301 of file pfm_gen_analysis.cpp.

1301  {
1302  char event[MAX_EVENT_NAME_LENGTH];
1303  char arch[MAX_ARCH_NAME_LENGTH];
1304  char line[MAX_LINE_LENGTH];
1305  char cmask_str[MAX_CMASK_STR_LENGTH];
1306  char inv_str[MAX_INV_STR_LENGTH];
1307  char sp_str[MAX_SP_STR_LENGTH];
1308  bzero( line, MAX_LINE_LENGTH );
1309  bzero( event, MAX_EVENT_NAME_LENGTH );
1310  bzero( arch, MAX_ARCH_NAME_LENGTH );
1311  bzero( cmask_str, MAX_CMASK_STR_LENGTH );
1312  bzero( inv_str, MAX_INV_STR_LENGTH );
1313  bzero( sp_str, MAX_SP_STR_LENGTH );
1314  char path_name[MAX_FILENAME_LENGTH];
1315  bzero( path_name, MAX_FILENAME_LENGTH );
1316  strcpy( path_name, dir );
1317  strcat( path_name, "/" );
1318  strcat( path_name, filename );
1319  gzFile res_file = gzopen( path_name, "rb" );
1320  if ( res_file != NULL ) {
1321  bzero( line, MAX_LINE_LENGTH );
1322  gzgets( res_file, line, MAX_LINE_LENGTH );
1323  if ( line[strlen( line ) - 1] == '\n' ) line[strlen( line ) - 1] = '\0';
1324  bzero( event, MAX_EVENT_NAME_LENGTH );
1325  sscanf( line, "%s %s %s %s %s\n", arch, event, cmask_str, inv_str, sp_str );
1326  std::string event_str( event );
1327  if ( atoi( cmask_str ) > 0 ) {
1328  event_str += " CMASK=";
1329  event_str += cmask_str;
1330  }
1331  if ( atoi( inv_str ) > 0 ) {
1332  event_str += " INV=";
1333  event_str += inv_str;
1334  }
1335  S_events.push_back( event_str );
1336  } // if(res_file != NULL)
1337  else {
1338  fprintf( stderr, "ERROR: Unable to open input file: %s\naborting...\n", filename );
1339  exit( 1 );
1340  }
1341  return 0;
1342 }
MAX_EVENT_NAME_LENGTH
#define MAX_EVENT_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:79
std::atoi
T atoi(T...args)
std::strcat
T strcat(T...args)
std::sscanf
T sscanf(T...args)
MAX_ARCH_NAME_LENGTH
#define MAX_ARCH_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:82
MAX_LINE_LENGTH
#define MAX_LINE_LENGTH
Definition: pfm_gen_analysis.cpp:78
std::string
STL class.
std::vector::push_back
T push_back(T...args)
MAX_SP_STR_LENGTH
#define MAX_SP_STR_LENGTH
Definition: pfm_gen_analysis.cpp:85
std::exit
T exit(T...args)
std::strlen
T strlen(T...args)
MAX_CMASK_STR_LENGTH
#define MAX_CMASK_STR_LENGTH
Definition: pfm_gen_analysis.cpp:83
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
MAX_INV_STR_LENGTH
#define MAX_INV_STR_LENGTH
Definition: pfm_gen_analysis.cpp:84
std::fprintf
T fprintf(T...args)
int read_S_file ( const char *  dir,
const char *  filename 
)

Definition at line 1188 of file pfm_gen_analysis.cpp.

1188  {
1189  char line[MAX_LINE_LENGTH];
1190  char event[MAX_EVENT_NAME_LENGTH];
1191  char arch[MAX_ARCH_NAME_LENGTH];
1192  unsigned int cmask;
1193  unsigned int inv;
1194  unsigned int sp;
1195  char cur_module_name[MAX_MODULE_NAME_LENGTH];
1196  bzero( line, MAX_LINE_LENGTH );
1197  bzero( event, MAX_EVENT_NAME_LENGTH );
1198  bzero( cur_module_name, MAX_MODULE_NAME_LENGTH );
1199  bzero( arch, MAX_ARCH_NAME_LENGTH );
1200 
1201  S_module* cur_module = new S_module();
1202  unsigned int module_num = 0;
1203 
1204  char path_name[MAX_FILENAME_LENGTH];
1205  bzero( path_name, MAX_FILENAME_LENGTH );
1206  strcpy( path_name, dir );
1207  strcat( path_name, "/" );
1208  strcat( path_name, filename );
1209  gzFile res_file = gzopen( path_name, "rb" );
1210 
1211  if ( res_file != NULL ) {
1212  bzero( line, MAX_LINE_LENGTH );
1213  gzgets( res_file, line, MAX_LINE_LENGTH );
1214  if ( line[strlen( line ) - 1] == '\n' ) line[strlen( line ) - 1] = '\0';
1215  bzero( event, MAX_EVENT_NAME_LENGTH );
1216  sscanf( line, "%s %s %u %u %u", arch, event, &cmask, &inv, &sp );
1217  if ( !strcmp( arch, "NHM" ) )
1218  nehalem = true;
1219  else
1220  nehalem = false;
1221  bzero( line, MAX_LINE_LENGTH );
1222  while ( gzgets( res_file, line, MAX_LINE_LENGTH ) != Z_NULL ) {
1223  if ( line[strlen( line ) - 1] == '\n' ) line[strlen( line ) - 1] = '\0';
1224  if ( strchr( line, ' ' ) == NULL ) // module
1225  {
1226  if ( module_num > 0 ) {
1227  put_S_module( cur_module, dir );
1228  cur_module->clear();
1229  }
1230  module_num++;
1231  char* end_sym = strchr( line, '%' );
1232  if ( end_sym == NULL ) // error
1233  {
1234  fprintf( stderr, "ERROR: Invalid module name. \nLINE: %s\naborting...\n", line );
1235  exit( 1 );
1236  }
1237  bzero( cur_module_name, MAX_MODULE_NAME_LENGTH );
1238  memcpy( cur_module_name, line, strlen( line ) - strlen( end_sym ) );
1239  cur_module->init( cur_module_name, arch, event, cmask, inv, sp );
1240  cur_module->set_total( atoi( end_sym + 1 ) );
1241  } // module
1242  else // symbol, libName, libOffset, value
1243  {
1244  unsigned int value = 0, libOffset = 0;
1245  char symbol[MAX_SYM_LENGTH];
1246  char libName[MAX_LIB_LENGTH];
1247  char final_sym[MAX_SYM_MOD_LENGTH];
1248  char final_lib[MAX_LIB_MOD_LENGTH];
1249  bzero( symbol, MAX_SYM_LENGTH );
1250  bzero( libName, MAX_LIB_LENGTH );
1251  bzero( final_sym, MAX_SYM_MOD_LENGTH );
1252  bzero( final_lib, MAX_LIB_MOD_LENGTH );
1253 
1254  sscanf( line, "%s %s %u %u", symbol, libName, &libOffset, &value );
1255  char realPathName_s[FILENAME_MAX];
1256  bzero( realPathName_s, FILENAME_MAX );
1257  char* realPathName = realpath( libName, realPathName_s );
1258  if ( realPathName != NULL && strlen( realPathName ) > 0 ) {
1259  std::map<std::string, FileInfo>::iterator result;
1260  result = libsInfo.find( realPathName );
1261  if ( result == libsInfo.end() ) { libsInfo[realPathName] = FileInfo( realPathName, true ); }
1262  const char* temp_sym = libsInfo[realPathName].symbolByOffset( libOffset );
1263  if ( temp_sym != NULL && strlen( temp_sym ) > 0 ) {
1264  int status;
1265  char* demangled_symbol = abi::__cxa_demangle( temp_sym, NULL, NULL, &status );
1266  if ( status == 0 ) {
1267  strcpy( final_sym, demangled_symbol );
1268  free( demangled_symbol );
1269  } else {
1270  strcpy( final_sym, temp_sym );
1271  }
1272  } else {
1273  strcpy( final_sym, "???" );
1274  }
1275  strcpy( final_lib, realPathName );
1276  } else {
1277  strcpy( final_sym, symbol );
1278  strcpy( final_lib, libName );
1279  }
1280  char index[MAX_LINE_LENGTH];
1281  bzero( index, MAX_LINE_LENGTH );
1282  strcpy( index, final_sym );
1283  strcat( index, "%" );
1284  strcat( index, final_lib );
1285  cur_module->add_sample( index, value );
1286  } // symbol, libName, libOffset, value
1287  bzero( line, MAX_LINE_LENGTH );
1288  } // while(gzgets(res_file, line, MAX_LINE_LENGTH)!=Z_NULL)
1289  put_S_module( cur_module, dir ); // last module!
1290  cur_module->clear();
1291  gzclose( res_file );
1292  } // if(res_file != NULL)
1293  else {
1294  fprintf( stderr, "ERROR: Unable to open input file: %s\naborting...\n", filename );
1295  exit( 1 );
1296  }
1297  delete cur_module; // delete it!
1298  return 0;
1299 }
MAX_EVENT_NAME_LENGTH
#define MAX_EVENT_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:79
std::atoi
T atoi(T...args)
std::strcat
T strcat(T...args)
MAX_LIB_LENGTH
#define MAX_LIB_LENGTH
Definition: pfm_gen_analysis.cpp:72
std::sscanf
T sscanf(T...args)
put_S_module
void put_S_module(S_module *cur_module, const char *dir)
Definition: pfm_gen_analysis.cpp:1036
S_module
Definition: pfm_gen_analysis.cpp:767
MAX_SYM_LENGTH
#define MAX_SYM_LENGTH
Definition: pfm_gen_analysis.cpp:70
MAX_ARCH_NAME_LENGTH
#define MAX_ARCH_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:82
S_module::add_sample
void add_sample(const char *index, unsigned int value)
Definition: pfm_gen_analysis.cpp:807
std::free
T free(T...args)
MAX_LINE_LENGTH
#define MAX_LINE_LENGTH
Definition: pfm_gen_analysis.cpp:78
std::map
STL class.
MAX_LIB_MOD_LENGTH
#define MAX_LIB_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:73
std::exit
T exit(T...args)
std::memcpy
T memcpy(T...args)
std::strlen
T strlen(T...args)
S_module::init
void init(const char *name, const char *architecture, const char *event_name, unsigned int c_mask, unsigned int inv_mask, unsigned int smpl_period)
Definition: pfm_gen_analysis.cpp:789
S_module::set_total
void set_total(unsigned int total)
Definition: pfm_gen_analysis.cpp:798
std::strcmp
T strcmp(T...args)
MAX_FILENAME_LENGTH
#define MAX_FILENAME_LENGTH
Definition: pfm_gen_analysis.cpp:68
std::strcpy
T strcpy(T...args)
S_module::clear
void clear()
Definition: pfm_gen_analysis.cpp:780
MAX_MODULE_NAME_LENGTH
#define MAX_MODULE_NAME_LENGTH
Definition: pfm_gen_analysis.cpp:80
std::fprintf
T fprintf(T...args)
MAX_SYM_MOD_LENGTH
#define MAX_SYM_MOD_LENGTH
Definition: pfm_gen_analysis.cpp:71
std::strchr
T strchr(T...args)
bool skipString ( const char *  strptr,
const char *  srcbuffer,
const char **  dstbuffer 
)

Definition at line 137 of file pfm_gen_analysis.cpp.

137  {
138  if ( strncmp( srcbuffer, strptr, strlen( strptr ) ) ) { return false; }
139  *dstbuffer = srcbuffer + strlen( strptr );
140  return true;
141 }
std::strlen
T strlen(T...args)
std::strncmp
T strncmp(T...args)
bool skipWhitespaces ( const char *  srcbuffer,
const char **  destbuffer 
)

Definition at line 123 of file pfm_gen_analysis.cpp.

123  {
124  if ( !isspace( *srcbuffer++ ) ) { return false; }
125  while ( isspace( *srcbuffer ) ) { srcbuffer++; }
126  *destbuffer = srcbuffer;
127  return true;
128 }
std::isspace
T isspace(T...args)