|
Gaudi Framework, version v22r0 |
| Home | Generated: 9 Feb 2011 |
|
|
Gaudi Framework, version v22r0 |
| Home | Generated: 9 Feb 2011 |
#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 <list>#include <map>#include <string>#include <vector>#include <sstream>#include <dirent.h>#include <errno.h>
Go to the source code of this file.
| #define CORE_L1_DTLB_MISS_CYCLES 10 |
Definition at line 40 of file pfm_gen_analysis.cpp.
| #define CORE_L2_HIT_CYCLES 14.5 |
Definition at line 39 of file pfm_gen_analysis.cpp.
| #define CORE_L2_MISS_CYCLES 200 |
Definition at line 38 of file pfm_gen_analysis.cpp.
| #define CORE_LCP_STALL_CYCLES 6 |
Definition at line 41 of file pfm_gen_analysis.cpp.
| #define CORE_OVERLAPPING_CYCLES 6 |
Definition at line 43 of file pfm_gen_analysis.cpp.
| #define CORE_SPAN_ACROSS_CACHE_LINE_CYCLES 20 |
Definition at line 44 of file pfm_gen_analysis.cpp.
| #define CORE_UNKNOWN_ADDR_STORE_CYCLES 5 |
Definition at line 42 of file pfm_gen_analysis.cpp.
| #define EXPECTED_CPI 0.25 |
Definition at line 65 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP 35 |
Definition at line 58 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L2_MISS_L3_HIT_SNOOP 60 |
Definition at line 59 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L2_MISS_L3_HITM 75 |
Definition at line 60 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT 200 |
Definition at line 56 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD 180 |
Definition at line 61 of file pfm_gen_analysis.cpp.
| #define I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT 350 |
Definition at line 57 of file pfm_gen_analysis.cpp.
| #define I7_L1_DTLB_WALK_COMPLETED_CYCLES 35 |
Definition at line 47 of file pfm_gen_analysis.cpp.
| #define I7_L1_ITLB_WALK_COMPLETED_CYCLES 35 |
Definition at line 48 of file pfm_gen_analysis.cpp.
| #define I7_L2_HIT_CYCLES 6 |
Definition at line 49 of file pfm_gen_analysis.cpp.
| #define I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES 225 |
Definition at line 53 of file pfm_gen_analysis.cpp.
| #define I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES 180 |
Definition at line 55 of file pfm_gen_analysis.cpp.
| #define I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES 360 |
Definition at line 54 of file pfm_gen_analysis.cpp.
| #define I7_L3_UNSHARED_HIT_CYCLES 35 |
Definition at line 50 of file pfm_gen_analysis.cpp.
| #define I7_OTHER_CORE_L2_HIT_CYCLES 60 |
Definition at line 51 of file pfm_gen_analysis.cpp.
| #define I7_OTHER_CORE_L2_HITM_CYCLES 75 |
Definition at line 52 of file pfm_gen_analysis.cpp.
| #define MAX_ARCH_NAME_LENGTH 20 |
Definition at line 81 of file pfm_gen_analysis.cpp.
| #define MAX_CMASK_STR_LENGTH 5 |
Definition at line 82 of file pfm_gen_analysis.cpp.
| #define MAX_EVENT_NAME_LENGTH 150 |
Definition at line 78 of file pfm_gen_analysis.cpp.
| #define MAX_FILENAME_LENGTH 1024 |
Definition at line 67 of file pfm_gen_analysis.cpp.
| #define MAX_INV_STR_LENGTH 5 |
Definition at line 83 of file pfm_gen_analysis.cpp.
| #define MAX_LIB_LENGTH 5000 |
Definition at line 71 of file pfm_gen_analysis.cpp.
| #define MAX_LIB_MOD_LENGTH 7000 |
Definition at line 72 of file pfm_gen_analysis.cpp.
| #define MAX_LINE_LENGTH 20000 |
Definition at line 77 of file pfm_gen_analysis.cpp.
| #define MAX_MODULE_NAME_LENGTH 250 |
Definition at line 79 of file pfm_gen_analysis.cpp.
| #define MAX_MODULES 1000 |
Definition at line 63 of file pfm_gen_analysis.cpp.
| #define MAX_SAMPLE_INDEX_LENGTH 10000 |
Definition at line 68 of file pfm_gen_analysis.cpp.
| #define MAX_SIMPLE_LIB_LENGTH 300 |
Definition at line 75 of file pfm_gen_analysis.cpp.
| #define MAX_SIMPLE_LIB_MOD_LENGTH 500 |
Definition at line 76 of file pfm_gen_analysis.cpp.
| #define MAX_SIMPLE_SYM_LENGTH 300 |
Definition at line 73 of file pfm_gen_analysis.cpp.
| #define MAX_SIMPLE_SYM_MOD_LENGTH 500 |
Definition at line 74 of file pfm_gen_analysis.cpp.
| #define MAX_SP_STR_LENGTH 50 |
Definition at line 84 of file pfm_gen_analysis.cpp.
| #define MAX_SYM_LENGTH 15000 |
Definition at line 69 of file pfm_gen_analysis.cpp.
| #define MAX_SYM_MOD_LENGTH 20000 |
Definition at line 70 of file pfm_gen_analysis.cpp.
| #define MAX_VALUE_STRING_LENGTH 250 |
Definition at line 80 of file pfm_gen_analysis.cpp.
| #define PIPE_BUFFER_LENGTH 1000 |
Definition at line 86 of file pfm_gen_analysis.cpp.
| void calc_core_deriv_values | ( | double | totalCycles | ) |
Definition at line 469 of file pfm_gen_analysis.cpp.
00470 { 00471 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 00472 { 00473 (it->second)["Total Cycles"] = (it->second)["UNHALTED_CORE_CYCLES"]; 00474 (it->second)["Stalled Cycles"] = (it->second)["RS_UOPS_DISPATCHED CMASK=1 INV=1"]; 00475 (it->second)["L2 Miss Impact"] = (it->second)["MEM_LOAD_RETIRED:L2_LINE_MISS"] * CORE_L2_MISS_CYCLES; 00476 (it->second)["L2 Hit Impact"] = ((it->second)["MEM_LOAD_RETIRED:L1D_LINE_MISS"] - (it->second)["MEM_LOAD_RETIRED:L2_LINE_MISS"]) * CORE_L2_HIT_CYCLES; 00477 (it->second)["L1 DTLB Miss Impact"] = (it->second)["MEM_LOAD_RETIRED:DTLB_MISS"] * CORE_L1_DTLB_MISS_CYCLES; 00478 (it->second)["LCP Stalls Impact"] = (it->second)["ILD_STALL"] * CORE_LCP_STALL_CYCLES; 00479 (it->second)["Loads Blocked by Unknown Address Store Impact"] = (it->second)["LOAD_BLOCK:STA"] * CORE_UNKNOWN_ADDR_STORE_CYCLES; 00480 (it->second)["Loads Overlapped with Stores Impact"] = (it->second)["LOAD_BLOCK:OVERLAP_STORE"] * CORE_OVERLAPPING_CYCLES; 00481 (it->second)["Loads Spanning across Cache Lines Impact"] = (it->second)["LOAD_BLOCK:UNTIL_RETIRE"] * CORE_SPAN_ACROSS_CACHE_LINE_CYCLES; 00482 (it->second)["Store-Fwd Stalls Impact"] = (it->second)["Loads Blocked by Unknown Address Store Impact"] + (it->second)["Loads Overlapped with Stores Impact"] + (it->second)["Loads Spanning across Cache Lines Impact"]; 00483 (it->second)["Counted Stalled Cycles"] = (it->second)["L2 Miss Impact"] + (it->second)["L2 Hit Impact"] + (it->second)["LCP Stalls Impact"] + (it->second)["L1 DTLB Miss Impact"] + (it->second)["Store-Fwd Stalls Impact"]; 00484 (it->second)["Instructions Retired"] = (it->second)["INSTRUCTIONS_RETIRED"]; 00485 (it->second)["ITLB Miss Rate in %"] = ((it->second)["ITLB_MISS_RETIRED"]/(it->second)["INSTRUCTIONS_RETIRED"])*100; 00486 (it->second)["Branch Instructions"] = (it->second)["BRANCH_INSTRUCTIONS_RETIRED"]; 00487 (it->second)["Load Instructions"] = (it->second)["INST_RETIRED:LOADS"]; 00488 (it->second)["Store Instructions"] = (it->second)["INST_RETIRED:STORES"]; 00489 (it->second)["Other Instructions"] = (it->second)["INST_RETIRED:OTHER"] - (it->second)["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"] - (it->second)["BRANCH_INSTRUCTIONS_RETIRED"]; 00490 (it->second)["% of Mispredicted Branches"] = ((it->second)["MISPREDICTED_BRANCH_RETIRED"]/(it->second)["BRANCH_INSTRUCTIONS_RETIRED"])*100; 00491 (it->second)["Packed SIMD Computational Instructions"] = (it->second)["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"]; 00492 (it->second)["Counted Instructions Retired"] = (it->second)["Branch Instructions"] + (it->second)["Load Instructions"] + (it->second)["Store Instructions"] + (it->second)["Other Instructions"] + (it->second)["Packed SIMD Computational Instructions"]; 00493 (it->second)["CPI"] = (it->second)["UNHALTED_CORE_CYCLES"]/(it->second)["INSTRUCTIONS_RETIRED"]; 00494 00495 double localPerformanceImprovement = (it->second)["CPI"]/EXPECTED_CPI; 00496 double cyclesAfterImprovement = (it->second)["UNHALTED_CORE_CYCLES"]/localPerformanceImprovement; 00497 double totalCyclesAfterImprovement = totalCycles-(it->second)["UNHALTED_CORE_CYCLES"]+cyclesAfterImprovement; 00498 (it->second)["iMargin"] = 100-(totalCyclesAfterImprovement/totalCycles)*100; 00499 00500 (it->second)["% of Total Cycles"] = (it->second)["RS_UOPS_DISPATCHED CMASK=1 INV=1"]*100/(it->second)["UNHALTED_CORE_CYCLES"]; 00501 (it->second)["L2 Miss % of counted Stalled Cycles"] =(it->second)["L2 Miss Impact"]*100/(it->second)["Counted Stalled Cycles"]; 00502 (it->second)["L2 Hit % of counted Stalled Cycles"] =(it->second)["L2 Hit Impact"]*100/(it->second)["Counted Stalled Cycles"]; 00503 (it->second)["L1 DTLB Miss % of counted Stalled Cycles"] =(it->second)["L1 DTLB Miss Impact"]*100/(it->second)["Counted Stalled Cycles"]; 00504 (it->second)["LCP Stalls % of counted Stalled Cycles"] =(it->second)["LCP Stalls Impact"]*100/(it->second)["Counted Stalled Cycles"]; 00505 (it->second)["Store-Fwd Stalls % of counted Stalled Cycles"] =(it->second)["Store-Fwd Stalls Impact"]*100/(it->second)["Counted Stalled Cycles"]; 00506 (it->second)["Loads Blocked % of Store-Fwd Stalls Cycles"] =(it->second)["Loads Blocked by Unknown Address Store Impact"]*100/(it->second)["Store-Fwd Stalls Impact"]; 00507 (it->second)["Loads Overlapped % of Store-Fwd Stalls Cycles"] =(it->second)["Loads Overlapped with Stores Impact"]*100/(it->second)["Store-Fwd Stalls Impact"]; 00508 (it->second)["Loads Spanning % of Store-Fwd Stalls Cycles"] =(it->second)["Loads Spanning across Cache Lines Impact"]*100/(it->second)["Store-Fwd Stalls Impact"]; 00509 00510 (it->second)["Load % of all Instructions"] =(it->second)["INST_RETIRED:LOADS"]*100/(it->second)["Counted Instructions Retired"]; 00511 (it->second)["Store % of all Instructions"] =(it->second)["INST_RETIRED:STORES"]*100/(it->second)["Counted Instructions Retired"]; 00512 (it->second)["Branch % of all Instructions"] =(it->second)["BRANCH_INSTRUCTIONS_RETIRED"]*100/(it->second)["Counted Instructions Retired"]; 00513 (it->second)["Packed SIMD % of all Instructions"] =(it->second)["SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"]*100/(it->second)["Counted Instructions Retired"]; 00514 (it->second)["Other % of all Instructions"] =(it->second)["Other Instructions"]*100/(it->second)["Counted Instructions Retired"]; 00515 } 00516 }
| void calc_nhm_deriv_values | ( | double | totalCycles | ) |
Definition at line 624 of file pfm_gen_analysis.cpp.
00625 { 00626 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 00627 { 00628 (it->second)["Total Cycles"] = (it->second)["CPU_CLK_UNHALTED:THREAD_P"]; 00629 00630 (it->second)["L2 Hit Impact"] = (it->second)["MEM_LOAD_RETIRED:L2_HIT"] * I7_L2_HIT_CYCLES; 00631 (it->second)["L3 Unshared Hit Impact"] = (it->second)["MEM_LOAD_RETIRED:L3_UNSHARED_HIT"] * I7_L3_UNSHARED_HIT_CYCLES; 00632 if((it->second)["MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM"]>(it->second)["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"]) 00633 { 00634 (it->second)["L2 Other Core Hit Impact"] = ((it->second)["MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM"] - (it->second)["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"])* I7_OTHER_CORE_L2_HIT_CYCLES; 00635 } 00636 else 00637 { 00638 (it->second)["L2 Other Core Hit Impact"] = 0.0; 00639 } 00640 (it->second)["L2 Other Core Hit Modified Impact"] = (it->second)["MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"] * I7_OTHER_CORE_L2_HITM_CYCLES; 00641 (it->second)["L3 Miss -> Local DRAM Hit Impact"] = (it->second)["MEM_UNCORE_RETIRED:LOCAL_DRAM"] * I7_L3_MISS_LOCAL_DRAM_HIT_CYCLES; 00642 (it->second)["L3 Miss -> Remote DRAM Hit Impact"] = (it->second)["MEM_UNCORE_RETIRED:REMOTE_DRAM"] * I7_L3_MISS_REMOTE_DRAM_HIT_CYCLES; 00643 (it->second)["L3 Miss -> Remote Cache Hit Impact"] = (it->second)["MEM_UNCORE_RETIRED:REMOTE_CACHE_LOCAL_HOME_HIT"] * I7_L3_MISS_REMOTE_CACHE_HIT_CYCLES; 00644 (it->second)["L3 Miss -> Total Impact"] = (it->second)["L3 Miss -> Local DRAM Hit Impact"] + (it->second)["L3 Miss -> Remote DRAM Hit Impact"] + (it->second)["L3 Miss -> Remote Cache Hit Impact"]; 00645 (it->second)["L1 DTLB Miss Impact"] = (it->second)["DTLB_LOAD_MISSES:WALK_COMPLETED"] * I7_L1_DTLB_WALK_COMPLETED_CYCLES; 00646 (it->second)["Counted Stalled Cycles due to Load Ops"] = (it->second)["L3 Miss -> Total Impact"] + (it->second)["L2 Hit Impact"] + (it->second)["L1 DTLB Miss Impact"] + (it->second)["L3 Unshared Hit Impact"] + (it->second)["L2 Other Core Hit Modified Impact"] + (it->second)["L2 Other Core Hit Impact"]; 00647 (it->second)["Cycles spent during DIV & SQRT Ops"] = (it->second)["ARITH:CYCLES_DIV_BUSY"]; 00648 (it->second)["Total Counted Stalled Cycles"] = (it->second)["Counted Stalled Cycles due to Load Ops"] + (it->second)["Cycles spent during DIV & SQRT Ops"]; 00649 (it->second)["Stalled Cycles"] = (it->second)["Total Counted Stalled Cycles"]; //TO BE FIXED when UOPS_EXECUTED:0x3f is fixed!! 00650 (it->second)["% of Total Cycles"] = (it->second)["Stalled Cycles"] * 100 / (it->second)["CPU_CLK_UNHALTED:THREAD_P"]; //TO BE FIXED!! see above 00651 (it->second)["L3 Miss % of Load Stalls"] = (it->second)["L3 Miss -> Total Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00652 (it->second)["L2 Hit % of Load Stalls"] = (it->second)["L2 Hit Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00653 (it->second)["L1 DTLB Miss % of Load Stalls"] = (it->second)["L1 DTLB Miss Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00654 (it->second)["L3 Unshared Hit % of Load Stalls"] = (it->second)["L3 Unshared Hit Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00655 (it->second)["L2 Other Core Hit % of Load Stalls"] = (it->second)["L2 Other Core Hit Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00656 (it->second)["L2 Other Core Hit Modified % of Load Stalls"] = (it->second)["L2 Other Core Hit Modified Impact"] * 100 / (it->second)["Counted Stalled Cycles due to Load Ops"]; 00657 (it->second)["DIV & SQRT Ops % of counted Stalled Cycles"] = (it->second)["Cycles spent during DIV & SQRT Ops"] * 100 / (it->second)["Total Counted Stalled Cycles"]; 00658 00659 (it->second)["Cycles IFETCH served by Local DRAM"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM"] * I7_IFETCH_L3_MISS_LOCAL_DRAM_HIT; 00660 (it->second)["Cycles IFETCH served by L3 (Modified)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM"] * I7_IFETCH_L2_MISS_L3_HITM; 00661 (it->second)["Cycles IFETCH served by L3 (Clean Snoop)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP"] * I7_IFETCH_L2_MISS_L3_HIT_SNOOP; 00662 (it->second)["Cycles IFETCH served by Remote L2"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD"] * I7_IFETCH_L3_MISS_REMOTE_CACHE_FWD; 00663 (it->second)["Cycles IFETCH served by Remote DRAM"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM"] * I7_IFETCH_L3_MISS_REMOTE_DRAM_HIT; 00664 (it->second)["Cycles IFETCH served by L3 (No Snoop)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT"] * I7_IFETCH_L2_MISS_L3_HIT_NO_SNOOP; 00665 (it->second)["Total L2 IFETCH miss Impact"] = (it->second)["Cycles IFETCH served by Local DRAM"] + (it->second)["Cycles IFETCH served by L3 (Modified)"] + (it->second)["Cycles IFETCH served by L3 (Clean Snoop)"] + (it->second)["Cycles IFETCH served by Remote L2"] + (it->second)["Cycles IFETCH served by Remote DRAM"] + (it->second)["Cycles IFETCH served by L3 (No Snoop)"]; 00666 (it->second)["Local DRAM IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by Local DRAM"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00667 (it->second)["L3 (Modified) IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by L3 (Modified)"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00668 (it->second)["L3 (Clean Snoop) IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by L3 (Clean Snoop)"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00669 (it->second)["Remote L2 IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by Remote L2"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00670 (it->second)["Remote DRAM IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by Remote DRAM"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00671 (it->second)["L3 (No Snoop) IFECTHes % Impact"] = (it->second)["Cycles IFETCH served by L3 (No Snoop)"] * 100 / (it->second)["Total L2 IFETCH miss Impact"]; 00672 (it->second)["Total L2 IFETCH misses"] = (it->second)["L2_RQSTS:IFETCH_MISS"]; 00673 (it->second)["% of IFETCHes served by Local DRAM"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00674 (it->second)["% of IFETCHes served by L3 (Modified)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00675 (it->second)["% of IFETCHes served by L3 (Clean Snoop)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00676 (it->second)["% of IFETCHes served by Remote L2"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00677 (it->second)["% of IFETCHes served by Remote DRAM"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00678 (it->second)["% of IFETCHes served by L3 (No Snoop)"] = (it->second)["OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT"] * 100 / (it->second)["L2_RQSTS:IFETCH_MISS"]; 00679 (it->second)["% of L2 IFETCH misses"] = (it->second)["L2_RQSTS:IFETCH_MISS"] * 100 / ((it->second)["L2_RQSTS:IFETCH_MISS"] + (it->second)["L2_RQSTS:IFETCH_HIT"]); 00680 (it->second)["L1 ITLB Miss Impact"] = (it->second)["ITLB_MISSES:WALK_COMPLETED"] * I7_L1_ITLB_WALK_COMPLETED_CYCLES; 00681 00682 (it->second)["Total Branch Instructions Executed"] = (it->second)["BR_INST_EXEC:ANY"]; 00683 (it->second)["% of Mispredicted Branches"] = (it->second)["BR_MISP_EXEC:ANY"] * 100 / (it->second)["BR_INST_EXEC:ANY"]; 00684 (it->second)["Direct Near Calls % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:DIRECT_NEAR_CALL"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00685 (it->second)["Indirect Near Calls % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:INDIRECT_NEAR_CALL"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00686 (it->second)["Indirect Near Non-Calls % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:INDIRECT_NON_CALL"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00687 (it->second)["All Near Calls % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:NEAR_CALLS"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00688 (it->second)["All Non Calls % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:NON_CALLS"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00689 (it->second)["All Returns % of Total Branches Executed"] = (it->second)["BR_INST_EXEC:RETURN_NEAR"] * 100 / (it->second)["Total Branch Instructions Executed"]; 00690 (it->second)["Total Branch Instructions Retired"] = (it->second)["BR_INST_RETIRED:ALL_BRANCHES"]; 00691 (it->second)["Conditionals % of Total Branches Retired"] = (it->second)["BR_INST_RETIRED:CONDITIONAL"] * 100 / (it->second)["Total Branch Instructions Retired"]; 00692 (it->second)["Near Calls % of Total Branches Retired"] = (it->second)["BR_INST_RETIRED:NEAR_CALL"] * 100 / (it->second)["Total Branch Instructions Retired"]; 00693 00694 (it->second)["Instruction Starvation % of Total Cycles"] = ((it->second)["UOPS_ISSUED:ANY CMASK=1 INV=1"] - (it->second)["RESOURCE_STALLS:ANY"])* 100 / (it->second)["CPU_CLK_UNHALTED:THREAD_P"]; 00695 (it->second)["% of Total Cycles spent handling FP exceptions"] = (it->second)["UOPS_DECODED:MS CMASK=1"]* 100 / (it->second)["CPU_CLK_UNHALTED:THREAD_P"]; 00696 (it->second)["# of Instructions per Call"] = (it->second)["INST_RETIRED:ANY_P"] / (it->second)["BR_INST_EXEC:NEAR_CALLS"]; 00697 00698 (it->second)["Instructions Retired"] = (it->second)["INST_RETIRED:ANY_P"]; 00699 (it->second)["ITLB Miss Rate in %"] = ((it->second)["ITLB_MISS_RETIRED"] / (it->second)["INST_RETIRED:ANY_P"]) * 100; 00700 00701 (it->second)["Branch Instructions"] = (it->second)["BR_INST_RETIRED:ALL_BRANCHES"]; 00702 (it->second)["Load Instructions"] = (it->second)["MEM_INST_RETIRED:LOADS"]; 00703 (it->second)["Store Instructions"] = (it->second)["MEM_INST_RETIRED:STORES"]; 00704 (it->second)["Other Instructions"] = (it->second)["Instructions Retired"] - (it->second)["MEM_INST_RETIRED:LOADS"] - (it->second)["MEM_INST_RETIRED:STORES"] - (it->second)["BR_INST_RETIRED:ALL_BRANCHES"]; 00705 (it->second)["Packed UOPS Retired"] = (it->second)["SSEX_UOPS_RETIRED:PACKED_DOUBLE"] + (it->second)["SSEX_UOPS_RETIRED:PACKED_SINGLE"]; 00706 (it->second)["CPI"] = (it->second)["CPU_CLK_UNHALTED:THREAD_P"] / (it->second)["INST_RETIRED:ANY_P"]; 00707 00708 double localPerformanceImprovement = (it->second)["CPI"]/EXPECTED_CPI; 00709 double cyclesAfterImprovement = (it->second)["CPU_CLK_UNHALTED:THREAD_P"]/localPerformanceImprovement; 00710 double totalCyclesAfterImprovement = totalCycles-(it->second)["CPU_CLK_UNHALTED:THREAD_P"]+cyclesAfterImprovement; 00711 (it->second)["iMargin"] = 100-(totalCyclesAfterImprovement/totalCycles)*100; 00712 00713 (it->second)["Load % of all Instructions"] = (it->second)["MEM_INST_RETIRED:LOADS"] * 100 / (it->second)["INST_RETIRED:ANY_P"]; 00714 (it->second)["Store % of all Instructions"] = (it->second)["MEM_INST_RETIRED:STORES"] * 100 / (it->second)["INST_RETIRED:ANY_P"]; 00715 (it->second)["Branch % of all Instructions"] = (it->second)["BR_INST_RETIRED:ALL_BRANCHES"] * 100 / (it->second)["INST_RETIRED:ANY_P"]; 00716 (it->second)["Other % of all Instructions"] = (it->second)["Other Instructions"] * 100 / (it->second)["INST_RETIRED:ANY_P"]; 00717 00718 (it->second)["Packed % of all UOPS Retired"] = (it->second)["Packed UOPS Retired"] * 100 / (it->second)["UOPS_RETIRED:ANY"]; 00719 } 00720 }
| void calc_post_deriv_values | ( | ) |
Definition at line 1687 of file pfm_gen_analysis.cpp.
01688 { 01689 if(nehalem) 01690 { 01691 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01692 { 01693 double simdnorm = 1. - normalize("Packed % of all UOPS Retired", (it->second)["Packed % of all UOPS Retired"], 1); 01694 double misspnorm = normalize("% of Mispredicted Branches", (it->second)["% of Mispredicted Branches"], 1); 01695 double stallnorm = normalize("Stalled Cycles", (it->second)["Stalled Cycles"], 1); 01696 (it->second)["iFactor"] = stallnorm*(simdnorm + misspnorm + stallnorm); 01697 } 01698 } 01699 else 01700 { 01701 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01702 { 01703 double simdnorm = 1. - normalize("Packed SIMD % of all Instructions", (it->second)["Packed SIMD % of all Instructions"], 1); 01704 double misspnorm = normalize("% of Mispredicted Branches", (it->second)["% of Mispredicted Branches"], 1); 01705 double stallnorm = normalize("Stalled Cycles", (it->second)["Stalled Cycles"], 1); 01706 (it->second)["iFactor"] = stallnorm*(simdnorm + misspnorm + stallnorm); 01707 } 01708 } 01709 }
| bool check_for_core_caa_events | ( | ) |
Definition at line 393 of file pfm_gen_analysis.cpp.
00394 { 00395 for(std::vector<std::string>::const_iterator it=core_caa_events.begin(); it!=core_caa_events.end(); ++it) 00396 { 00397 if(find(C_events.begin(), C_events.end(), (*it))==C_events.end()) 00398 { 00399 fprintf(stderr, "ERROR: Cannot find event %s!!!\naborting...\n", (*it).c_str()); 00400 return false; 00401 } 00402 } 00403 return true; 00404 }
| bool check_for_nhm_caa_events | ( | ) |
Definition at line 406 of file pfm_gen_analysis.cpp.
00407 { 00408 for(std::vector<std::string>::const_iterator it=nhm_caa_events.begin(); it!=nhm_caa_events.end(); ++it) 00409 { 00410 if(find(C_events.begin(), C_events.end(), (*it))==C_events.end()) 00411 { 00412 fprintf(stderr, "ERROR: Cannot find event %s!!!\naborting...\n", (*it).c_str()); 00413 return false; 00414 } 00415 } 00416 return true; 00417 }
| int finalize_S_html_pages | ( | const char * | dir | ) |
Definition at line 1472 of file pfm_gen_analysis.cpp.
01473 { 01474 for(std::map<std::string, unsigned int>::const_iterator i = modules_tot_samples.begin(); i != modules_tot_samples.end(); i++) 01475 { 01476 char module_filename[MAX_FILENAME_LENGTH]; 01477 strcpy(module_filename, dir); 01478 strcat(module_filename, "/HTML/"); 01479 strcat(module_filename, (i->first).c_str()); 01480 strcat(module_filename, ".html"); 01481 FILE *module_file = fopen(module_filename, "a"); 01482 if(module_file == NULL) 01483 { 01484 fprintf(stderr, "ERROR: Unable to append to file: %s\naborting...\n", module_filename); 01485 exit(1); 01486 } 01487 fprintf(module_file, "</body>\n</html>\n"); 01488 if(fclose(module_file)) 01489 { 01490 fprintf(stderr, "ERROR: Cannot close file %s!!!\naborting...\n", module_filename); 01491 exit(1); 01492 } 01493 } 01494 return 0; 01495 }
| const char* func_name | ( | const char * | demangled_symbol | ) |
Definition at line 855 of file pfm_gen_analysis.cpp.
00856 { 00857 char *operator_string_begin = strstr(demangled_symbol, "operator"); 00858 if(operator_string_begin != NULL) 00859 { 00860 char *operator_string_end = operator_string_begin+8; 00861 while(*operator_string_end == ' ') operator_string_end++; 00862 if(strstr(operator_string_end, "delete[]")==operator_string_end) 00863 { 00864 operator_string_end+=8; 00865 *operator_string_end='\0'; 00866 } 00867 else if(strstr(operator_string_end, "delete")==operator_string_end) 00868 { 00869 operator_string_end+=6; 00870 *operator_string_end='\0'; 00871 } 00872 else if(strstr(operator_string_end, "new[]")==operator_string_end) 00873 { 00874 operator_string_end+=5; 00875 *operator_string_end='\0'; 00876 } 00877 else if(strstr(operator_string_end, "new")==operator_string_end) 00878 { 00879 operator_string_end+=3; 00880 *operator_string_end='\0'; 00881 } 00882 else if(strstr(operator_string_end, ">>=")==operator_string_end) 00883 { 00884 operator_string_end+=3; 00885 *operator_string_end='\0'; 00886 } 00887 else if(strstr(operator_string_end, "<<=")==operator_string_end) 00888 { 00889 operator_string_end+=3; 00890 *operator_string_end='\0'; 00891 } 00892 else if(strstr(operator_string_end, "->*")==operator_string_end) 00893 { 00894 operator_string_end+=3; 00895 *operator_string_end='\0'; 00896 } 00897 else if(strstr(operator_string_end, "<<")==operator_string_end) 00898 { 00899 operator_string_end+=2; 00900 *operator_string_end='\0'; 00901 } 00902 else if(strstr(operator_string_end, ">>")==operator_string_end) 00903 { 00904 operator_string_end+=2; 00905 *operator_string_end='\0'; 00906 } 00907 else if(strstr(operator_string_end, ">=")==operator_string_end) 00908 { 00909 operator_string_end+=2; 00910 *operator_string_end='\0'; 00911 } 00912 else if(strstr(operator_string_end, "<=")==operator_string_end) 00913 { 00914 operator_string_end+=2; 00915 *operator_string_end='\0'; 00916 } 00917 else if(strstr(operator_string_end, "==")==operator_string_end) 00918 { 00919 operator_string_end+=2; 00920 *operator_string_end='\0'; 00921 } 00922 else if(strstr(operator_string_end, "!=")==operator_string_end) 00923 { 00924 operator_string_end+=2; 00925 *operator_string_end='\0'; 00926 } 00927 else if(strstr(operator_string_end, "|=")==operator_string_end) 00928 { 00929 operator_string_end+=2; 00930 *operator_string_end='\0'; 00931 } 00932 else if(strstr(operator_string_end, "&=")==operator_string_end) 00933 { 00934 operator_string_end+=2; 00935 *operator_string_end='\0'; 00936 } 00937 else if(strstr(operator_string_end, "^=")==operator_string_end) 00938 { 00939 operator_string_end+=2; 00940 *operator_string_end='\0'; 00941 } 00942 else if(strstr(operator_string_end, "%=")==operator_string_end) 00943 { 00944 operator_string_end+=2; 00945 *operator_string_end='\0'; 00946 } 00947 else if(strstr(operator_string_end, "/=")==operator_string_end) 00948 { 00949 operator_string_end+=2; 00950 *operator_string_end='\0'; 00951 } 00952 else if(strstr(operator_string_end, "*=")==operator_string_end) 00953 { 00954 operator_string_end+=2; 00955 *operator_string_end='\0'; 00956 } 00957 else if(strstr(operator_string_end, "-=")==operator_string_end) 00958 { 00959 operator_string_end+=2; 00960 *operator_string_end='\0'; 00961 } 00962 else if(strstr(operator_string_end, "+=")==operator_string_end) 00963 { 00964 operator_string_end+=2; 00965 *operator_string_end='\0'; 00966 } 00967 else if(strstr(operator_string_end, "&&")==operator_string_end) 00968 { 00969 operator_string_end+=2; 00970 *operator_string_end='\0'; 00971 } 00972 else if(strstr(operator_string_end, "||")==operator_string_end) 00973 { 00974 operator_string_end+=2; 00975 *operator_string_end='\0'; 00976 } 00977 else if(strstr(operator_string_end, "[]")==operator_string_end) 00978 { 00979 operator_string_end+=2; 00980 *operator_string_end='\0'; 00981 } 00982 else if(strstr(operator_string_end, "()")==operator_string_end) 00983 { 00984 operator_string_end+=2; 00985 *operator_string_end='\0'; 00986 } 00987 else if(strstr(operator_string_end, "++")==operator_string_end) 00988 { 00989 operator_string_end+=2; 00990 *operator_string_end='\0'; 00991 } 00992 else if(strstr(operator_string_end, "--")==operator_string_end) 00993 { 00994 operator_string_end+=2; 00995 *operator_string_end='\0'; 00996 } 00997 else if(strstr(operator_string_end, "->")==operator_string_end) 00998 { 00999 operator_string_end+=2; 01000 *operator_string_end='\0'; 01001 } 01002 else if(strstr(operator_string_end, "<")==operator_string_end) 01003 { 01004 operator_string_end+=1; 01005 *operator_string_end='\0'; 01006 } 01007 else if(strstr(operator_string_end, ">")==operator_string_end) 01008 { 01009 operator_string_end+=1; 01010 *operator_string_end='\0'; 01011 } 01012 else if(strstr(operator_string_end, "~")==operator_string_end) 01013 { 01014 operator_string_end+=1; 01015 *operator_string_end='\0'; 01016 } 01017 else if(strstr(operator_string_end, "!")==operator_string_end) 01018 { 01019 operator_string_end+=1; 01020 *operator_string_end='\0'; 01021 } 01022 else if(strstr(operator_string_end, "+")==operator_string_end) 01023 { 01024 operator_string_end+=1; 01025 *operator_string_end='\0'; 01026 } 01027 else if(strstr(operator_string_end, "-")==operator_string_end) 01028 { 01029 operator_string_end+=1; 01030 *operator_string_end='\0'; 01031 } 01032 else if(strstr(operator_string_end, "*")==operator_string_end) 01033 { 01034 operator_string_end+=1; 01035 *operator_string_end='\0'; 01036 } 01037 else if(strstr(operator_string_end, "/")==operator_string_end) 01038 { 01039 operator_string_end+=1; 01040 *operator_string_end='\0'; 01041 } 01042 else if(strstr(operator_string_end, "%")==operator_string_end) 01043 { 01044 operator_string_end+=1; 01045 *operator_string_end='\0'; 01046 } 01047 else if(strstr(operator_string_end, "^")==operator_string_end) 01048 { 01049 operator_string_end+=1; 01050 *operator_string_end='\0'; 01051 } 01052 else if(strstr(operator_string_end, "&")==operator_string_end) 01053 { 01054 operator_string_end+=1; 01055 *operator_string_end='\0'; 01056 } 01057 else if(strstr(operator_string_end, "|")==operator_string_end) 01058 { 01059 operator_string_end+=1; 01060 *operator_string_end='\0'; 01061 } 01062 else if(strstr(operator_string_end, ",")==operator_string_end) 01063 { 01064 operator_string_end+=1; 01065 *operator_string_end='\0'; 01066 } 01067 else if(strstr(operator_string_end, "=")==operator_string_end) 01068 { 01069 operator_string_end+=1; 01070 *operator_string_end='\0'; 01071 } 01072 return operator_string_begin; 01073 } 01074 char *end_of_demangled_name = strrchr(demangled_symbol, ')'); 01075 if(end_of_demangled_name != NULL) 01076 { 01077 int pars = 1; 01078 char c; 01079 while(pars>0 && end_of_demangled_name!=demangled_symbol) 01080 { 01081 c = *(--end_of_demangled_name); 01082 if(c==')') 01083 { 01084 pars++; 01085 } 01086 else if(c=='(') 01087 { 01088 pars--; 01089 } 01090 } 01091 } 01092 else 01093 { 01094 return demangled_symbol; 01095 } 01096 char *end_of_func_name = end_of_demangled_name; 01097 if(end_of_func_name != NULL) 01098 { 01099 *end_of_func_name = '\0'; 01100 char c = *(--end_of_func_name); 01101 if(c=='>') 01102 { 01103 int pars = 1; 01104 while(pars>0 && end_of_func_name!=demangled_symbol) 01105 { 01106 c = *(--end_of_func_name); 01107 if(c=='>') 01108 { 01109 pars++; 01110 } 01111 else if(c=='<') 01112 { 01113 pars--; 01114 } 01115 } 01116 *end_of_func_name = '\0'; 01117 } 01118 c = *(--end_of_func_name); 01119 while(isalnum(c) || c=='_' || c=='~') 01120 { 01121 c = *(--end_of_func_name); 01122 } 01123 return ++end_of_func_name; 01124 } 01125 return demangled_symbol; 01126 }
| double getTotalCycles | ( | ) |
Definition at line 1715 of file pfm_gen_analysis.cpp.
01716 { 01717 double sum=0; 01718 if(nehalem) 01719 { 01720 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01721 { 01722 sum += (it->second)["CPU_CLK_UNHALTED:THREAD_P"]; 01723 } 01724 } 01725 else 01726 { 01727 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01728 { 01729 sum += (it->second)["UNHALTED_CORE_CYCLES"]; 01730 } 01731 } 01732 return sum; 01733 }
| void html_special_chars | ( | const char * | s, | |
| char * | s_mod | |||
| ) |
Definition at line 821 of file pfm_gen_analysis.cpp.
00822 { 00823 int n = strlen(s); 00824 *s_mod = '\0'; 00825 for (int i=0; i < n; i++) 00826 { 00827 switch(s[i]) 00828 { 00829 case '<': 00830 strcat(s_mod, "<"); 00831 break; 00832 case '>': 00833 strcat(s_mod, ">"); 00834 break; 00835 case '&': 00836 strcat(s_mod, "&"); 00837 break; 00838 case '"': 00839 strcat(s_mod, """); 00840 break; 00841 default: 00842 char to_app[2]; 00843 to_app[0]=s[i]; 00844 to_app[1]='\0'; 00845 strcat(s_mod, to_app); 00846 break; 00847 } 00848 } 00849 return; 00850 }
| void init_core_caa_events | ( | ) |
Definition at line 323 of file pfm_gen_analysis.cpp.
00324 { 00325 core_caa_events.push_back("BRANCH_INSTRUCTIONS_RETIRED"); 00326 core_caa_events.push_back("ILD_STALL"); 00327 core_caa_events.push_back("INST_RETIRED:LOADS"); 00328 core_caa_events.push_back("INST_RETIRED:OTHER"); 00329 core_caa_events.push_back("INST_RETIRED:STORES"); 00330 core_caa_events.push_back("INSTRUCTIONS_RETIRED"); 00331 core_caa_events.push_back("LOAD_BLOCK:OVERLAP_STORE"); 00332 core_caa_events.push_back("LOAD_BLOCK:STA"); 00333 core_caa_events.push_back("LOAD_BLOCK:UNTIL_RETIRE"); 00334 core_caa_events.push_back("MEM_LOAD_RETIRED:DTLB_MISS"); 00335 core_caa_events.push_back("MEM_LOAD_RETIRED:L1D_LINE_MISS"); 00336 core_caa_events.push_back("MEM_LOAD_RETIRED:L2_LINE_MISS"); 00337 core_caa_events.push_back("MISPREDICTED_BRANCH_RETIRED"); 00338 //core_caa_events.push_back("RS_UOPS_DISPATCHED"); 00339 //core_caa_events.push_back("RS_UOPS_DISPATCHED CMASK=1"); 00340 core_caa_events.push_back("RS_UOPS_DISPATCHED CMASK=1 INV=1"); 00341 core_caa_events.push_back("SIMD_COMP_INST_RETIRED:PACKED_SINGLE:PACKED_DOUBLE"); 00342 core_caa_events.push_back("UNHALTED_CORE_CYCLES"); 00343 //core_caa_events.push_back("UOPS_RETIRED:ANY"); 00344 //core_caa_events.push_back("UOPS_RETIRED:FUSED"); 00345 //core_caa_events.push_back("IDLE_DURING_DIV"); 00346 }
| void init_core_caa_events_displ | ( | ) |
Definition at line 419 of file pfm_gen_analysis.cpp.
00420 { 00421 core_caa_events_displ.push_back("Total Cycles"); 00422 core_caa_events_displ.push_back("Stalled Cycles"); 00423 core_caa_events_displ.push_back("% of Total Cycles"); 00424 core_caa_events_displ.push_back("Instructions Retired"); 00425 core_caa_events_displ.push_back("CPI"); 00426 core_caa_events_displ.push_back(""); 00427 core_caa_events_displ.push_back("iMargin"); 00428 core_caa_events_displ.push_back("iFactor"); 00429 core_caa_events_displ.push_back(""); 00430 core_caa_events_displ.push_back("Counted Stalled Cycles"); 00431 core_caa_events_displ.push_back(""); 00432 core_caa_events_displ.push_back("L2 Miss Impact"); 00433 core_caa_events_displ.push_back("L2 Miss % of counted Stalled Cycles"); 00434 core_caa_events_displ.push_back(""); 00435 core_caa_events_displ.push_back("L2 Hit Impact"); 00436 core_caa_events_displ.push_back("L2 Hit % of counted Stalled Cycles"); 00437 core_caa_events_displ.push_back(""); 00438 core_caa_events_displ.push_back("L1 DTLB Miss Impact"); 00439 core_caa_events_displ.push_back("L1 DTLB Miss % of counted Stalled Cycles"); 00440 core_caa_events_displ.push_back(""); 00441 core_caa_events_displ.push_back("LCP Stalls Impact"); 00442 core_caa_events_displ.push_back("LCP Stalls % of counted Stalled Cycles"); 00443 core_caa_events_displ.push_back(""); 00444 core_caa_events_displ.push_back("Store-Fwd Stalls Impact"); 00445 core_caa_events_displ.push_back("Store-Fwd Stalls % of counted Stalled Cycles"); 00446 core_caa_events_displ.push_back(""); 00447 core_caa_events_displ.push_back("Loads Blocked by Unknown Address Store Impact"); 00448 core_caa_events_displ.push_back("Loads Blocked % of Store-Fwd Stalls Cycles"); 00449 core_caa_events_displ.push_back("Loads Overlapped with Stores Impact"); 00450 core_caa_events_displ.push_back("Loads Overlapped % of Store-Fwd Stalls Cycles"); 00451 core_caa_events_displ.push_back("Loads Spanning across Cache Lines Impact"); 00452 core_caa_events_displ.push_back("Loads Spanning % of Store-Fwd Stalls Cycles"); 00453 core_caa_events_displ.push_back(""); 00454 core_caa_events_displ.push_back("Load Instructions"); 00455 core_caa_events_displ.push_back("Load % of all Instructions"); 00456 core_caa_events_displ.push_back("Store Instructions"); 00457 core_caa_events_displ.push_back("Store % of all Instructions"); 00458 core_caa_events_displ.push_back("Branch Instructions"); 00459 core_caa_events_displ.push_back("Branch % of all Instructions"); 00460 core_caa_events_displ.push_back("Packed SIMD Computational Instructions"); 00461 core_caa_events_displ.push_back("Packed SIMD % of all Instructions"); 00462 core_caa_events_displ.push_back("Other Instructions"); 00463 core_caa_events_displ.push_back("Other % of all Instructions"); 00464 core_caa_events_displ.push_back(""); 00465 core_caa_events_displ.push_back("ITLB Miss Rate in %"); 00466 core_caa_events_displ.push_back("% of Mispredicted Branches"); 00467 }
| void init_nhm_caa_events | ( | ) |
Definition at line 348 of file pfm_gen_analysis.cpp.
00349 { 00350 nhm_caa_events.push_back("ARITH:CYCLES_DIV_BUSY"); 00351 nhm_caa_events.push_back("BR_INST_EXEC:ANY"); 00352 nhm_caa_events.push_back("BR_INST_EXEC:DIRECT_NEAR_CALL"); 00353 nhm_caa_events.push_back("BR_INST_EXEC:INDIRECT_NEAR_CALL"); 00354 nhm_caa_events.push_back("BR_INST_EXEC:INDIRECT_NON_CALL"); 00355 nhm_caa_events.push_back("BR_INST_EXEC:NEAR_CALLS"); 00356 nhm_caa_events.push_back("BR_INST_EXEC:NON_CALLS"); 00357 nhm_caa_events.push_back("BR_INST_EXEC:RETURN_NEAR"); 00358 nhm_caa_events.push_back("BR_INST_RETIRED:ALL_BRANCHES"); 00359 nhm_caa_events.push_back("BR_INST_RETIRED:CONDITIONAL"); 00360 nhm_caa_events.push_back("BR_INST_RETIRED:NEAR_CALL"); 00361 nhm_caa_events.push_back("BR_MISP_EXEC:ANY"); 00362 nhm_caa_events.push_back("CPU_CLK_UNHALTED:THREAD_P"); 00363 nhm_caa_events.push_back("DTLB_LOAD_MISSES:WALK_COMPLETED"); 00364 nhm_caa_events.push_back("INST_RETIRED:ANY_P"); 00365 nhm_caa_events.push_back("ITLB_MISSES:WALK_COMPLETED"); 00366 nhm_caa_events.push_back("L2_RQSTS:IFETCH_HIT"); 00367 nhm_caa_events.push_back("L2_RQSTS:IFETCH_MISS"); 00368 nhm_caa_events.push_back("MEM_INST_RETIRED:LOADS"); 00369 nhm_caa_events.push_back("MEM_INST_RETIRED:STORES"); 00370 nhm_caa_events.push_back("MEM_LOAD_RETIRED:L2_HIT"); 00371 nhm_caa_events.push_back("MEM_LOAD_RETIRED:L3_MISS"); 00372 nhm_caa_events.push_back("MEM_LOAD_RETIRED:L3_UNSHARED_HIT"); 00373 nhm_caa_events.push_back("MEM_LOAD_RETIRED:OTHER_CORE_L2_HIT_HITM"); 00374 nhm_caa_events.push_back("MEM_UNCORE_RETIRED:LOCAL_DRAM"); 00375 nhm_caa_events.push_back("MEM_UNCORE_RETIRED:OTHER_CORE_L2_HITM"); 00376 nhm_caa_events.push_back("MEM_UNCORE_RETIRED:REMOTE_CACHE_LOCAL_HOME_HIT"); 00377 nhm_caa_events.push_back("MEM_UNCORE_RETIRED:REMOTE_DRAM"); 00378 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:LOCAL_DRAM"); 00379 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HITM"); 00380 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:OTHER_CORE_HIT_SNP"); 00381 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_CACHE_FWD"); 00382 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:REMOTE_DRAM"); 00383 nhm_caa_events.push_back("OFFCORE_RESPONSE_0:DMND_IFETCH:UNCORE_HIT"); 00384 nhm_caa_events.push_back("RESOURCE_STALLS:ANY"); 00385 nhm_caa_events.push_back("SSEX_UOPS_RETIRED:PACKED_DOUBLE"); 00386 nhm_caa_events.push_back("SSEX_UOPS_RETIRED:PACKED_SINGLE"); 00387 nhm_caa_events.push_back("UOPS_DECODED:MS CMASK=1"); 00388 nhm_caa_events.push_back("UOPS_ISSUED:ANY CMASK=1 INV=1"); 00389 nhm_caa_events.push_back("ITLB_MISS_RETIRED"); 00390 nhm_caa_events.push_back("UOPS_RETIRED:ANY"); 00391 }
| void init_nhm_caa_events_displ | ( | ) |
Definition at line 518 of file pfm_gen_analysis.cpp.
00519 { 00520 nhm_caa_events_displ.push_back("Total Cycles"); 00521 nhm_caa_events_displ.push_back("Instructions Retired"); 00522 nhm_caa_events_displ.push_back("CPI"); 00523 nhm_caa_events_displ.push_back(""); 00524 nhm_caa_events_displ.push_back("iMargin"); 00525 nhm_caa_events_displ.push_back("iFactor"); 00526 nhm_caa_events_displ.push_back(""); 00527 nhm_caa_events_displ.push_back("Stalled Cycles"); 00528 nhm_caa_events_displ.push_back("% of Total Cycles"); 00529 nhm_caa_events_displ.push_back("Total Counted Stalled Cycles"); 00530 nhm_caa_events_displ.push_back(""); 00531 nhm_caa_events_displ.push_back("Instruction Starvation % of Total Cycles"); 00532 nhm_caa_events_displ.push_back("# of Instructions per Call"); 00533 nhm_caa_events_displ.push_back("% of Total Cycles spent handling FP exceptions"); 00534 nhm_caa_events_displ.push_back(""); 00535 nhm_caa_events_displ.push_back("Counted Stalled Cycles due to Load Ops"); 00536 nhm_caa_events_displ.push_back(""); 00537 nhm_caa_events_displ.push_back("L2 Hit Impact"); 00538 nhm_caa_events_displ.push_back("L2 Hit % of Load Stalls"); 00539 nhm_caa_events_displ.push_back(""); 00540 nhm_caa_events_displ.push_back("L3 Unshared Hit Impact"); 00541 nhm_caa_events_displ.push_back("L3 Unshared Hit % of Load Stalls"); 00542 nhm_caa_events_displ.push_back(""); 00543 nhm_caa_events_displ.push_back("L2 Other Core Hit Impact"); 00544 nhm_caa_events_displ.push_back("L2 Other Core Hit % of Load Stalls"); 00545 nhm_caa_events_displ.push_back(""); 00546 nhm_caa_events_displ.push_back("L2 Other Core Hit Modified Impact"); 00547 nhm_caa_events_displ.push_back("L2 Other Core Hit Modified % of Load Stalls"); 00548 nhm_caa_events_displ.push_back(""); 00549 nhm_caa_events_displ.push_back("L3 Miss -> Local DRAM Hit Impact"); 00550 nhm_caa_events_displ.push_back("L3 Miss -> Remote DRAM Hit Impact"); 00551 nhm_caa_events_displ.push_back("L3 Miss -> Remote Cache Hit Impact"); 00552 nhm_caa_events_displ.push_back("L3 Miss -> Total Impact"); 00553 nhm_caa_events_displ.push_back("L3 Miss % of Load Stalls"); 00554 nhm_caa_events_displ.push_back(""); 00555 nhm_caa_events_displ.push_back("L1 DTLB Miss Impact"); 00556 nhm_caa_events_displ.push_back("L1 DTLB Miss % of Load Stalls"); 00557 nhm_caa_events_displ.push_back(""); 00558 nhm_caa_events_displ.push_back("Cycles spent during DIV & SQRT Ops"); 00559 nhm_caa_events_displ.push_back("DIV & SQRT Ops % of counted Stalled Cycles"); 00560 nhm_caa_events_displ.push_back(""); 00561 nhm_caa_events_displ.push_back("Total L2 IFETCH misses"); 00562 nhm_caa_events_displ.push_back("% of L2 IFETCH misses"); 00563 nhm_caa_events_displ.push_back(""); 00564 nhm_caa_events_displ.push_back("% of IFETCHes served by Local DRAM"); 00565 nhm_caa_events_displ.push_back("% of IFETCHes served by L3 (Modified)"); 00566 nhm_caa_events_displ.push_back("% of IFETCHes served by L3 (Clean Snoop)"); 00567 nhm_caa_events_displ.push_back("% of IFETCHes served by Remote L2"); 00568 nhm_caa_events_displ.push_back("% of IFETCHes served by Remote DRAM"); 00569 nhm_caa_events_displ.push_back("% of IFETCHes served by L3 (No Snoop)"); 00570 nhm_caa_events_displ.push_back(""); 00571 nhm_caa_events_displ.push_back("Total L2 IFETCH miss Impact"); 00572 nhm_caa_events_displ.push_back(""); 00573 nhm_caa_events_displ.push_back("Cycles IFETCH served by Local DRAM"); 00574 nhm_caa_events_displ.push_back("Local DRAM IFECTHes % Impact"); 00575 nhm_caa_events_displ.push_back(""); 00576 nhm_caa_events_displ.push_back("Cycles IFETCH served by L3 (Modified)"); 00577 nhm_caa_events_displ.push_back("L3 (Modified) IFECTHes % Impact"); 00578 nhm_caa_events_displ.push_back(""); 00579 nhm_caa_events_displ.push_back("Cycles IFETCH served by L3 (Clean Snoop)"); 00580 nhm_caa_events_displ.push_back("L3 (Clean Snoop) IFECTHes % Impact"); 00581 nhm_caa_events_displ.push_back(""); 00582 nhm_caa_events_displ.push_back("Cycles IFETCH served by Remote L2"); 00583 nhm_caa_events_displ.push_back("Remote L2 IFECTHes % Impact"); 00584 nhm_caa_events_displ.push_back(""); 00585 nhm_caa_events_displ.push_back("Cycles IFETCH served by Remote DRAM"); 00586 nhm_caa_events_displ.push_back("Remote DRAM IFECTHes % Impact"); 00587 nhm_caa_events_displ.push_back(""); 00588 nhm_caa_events_displ.push_back("Cycles IFETCH served by L3 (No Snoop)"); 00589 nhm_caa_events_displ.push_back("L3 (No Snoop) IFECTHes % Impact"); 00590 nhm_caa_events_displ.push_back(""); 00591 nhm_caa_events_displ.push_back("Total Branch Instructions Executed"); 00592 nhm_caa_events_displ.push_back("% of Mispredicted Branches"); 00593 nhm_caa_events_displ.push_back(""); 00594 nhm_caa_events_displ.push_back("Direct Near Calls % of Total Branches Executed"); 00595 nhm_caa_events_displ.push_back("Indirect Near Calls % of Total Branches Executed"); 00596 nhm_caa_events_displ.push_back("Indirect Near Non-Calls % of Total Branches Executed"); 00597 nhm_caa_events_displ.push_back("All Near Calls % of Total Branches Executed"); 00598 nhm_caa_events_displ.push_back("All Non Calls % of Total Branches Executed"); 00599 nhm_caa_events_displ.push_back("All Returns % of Total Branches Executed"); 00600 nhm_caa_events_displ.push_back(""); 00601 nhm_caa_events_displ.push_back("Total Branch Instructions Retired"); 00602 nhm_caa_events_displ.push_back("Conditionals % of Total Branches Retired"); 00603 nhm_caa_events_displ.push_back("Near Calls % of Total Branches Retired"); 00604 nhm_caa_events_displ.push_back(""); 00605 nhm_caa_events_displ.push_back("L1 ITLB Miss Impact"); 00606 nhm_caa_events_displ.push_back("ITLB Miss Rate in %"); 00607 nhm_caa_events_displ.push_back(""); 00608 nhm_caa_events_displ.push_back("Branch Instructions"); 00609 nhm_caa_events_displ.push_back("Branch % of all Instructions"); 00610 nhm_caa_events_displ.push_back(""); 00611 nhm_caa_events_displ.push_back("Load Instructions"); 00612 nhm_caa_events_displ.push_back("Load % of all Instructions"); 00613 nhm_caa_events_displ.push_back(""); 00614 nhm_caa_events_displ.push_back("Store Instructions"); 00615 nhm_caa_events_displ.push_back("Store % of all Instructions"); 00616 nhm_caa_events_displ.push_back(""); 00617 nhm_caa_events_displ.push_back("Other Instructions"); 00618 nhm_caa_events_displ.push_back("Other % of all Instructions"); 00619 nhm_caa_events_displ.push_back(""); 00620 nhm_caa_events_displ.push_back("Packed UOPS Retired"); 00621 nhm_caa_events_displ.push_back("Packed % of all UOPS Retired"); 00622 }
| int main | ( | int | argc, | |
| char * | argv[] | |||
| ) |
Definition at line 1738 of file pfm_gen_analysis.cpp.
01739 { 01740 if(argc<2 || argc>4) 01741 { 01742 printf("\n\nUsage: %s DIRECTORY [--caa] [--csv]\n\n", argv[0]); 01743 exit(1); 01744 } 01745 01746 bool caa = false; 01747 bool csv = false; 01748 for(int i=2; i<argc; i++) 01749 { 01750 if(!strcmp(argv[i], "--caa")) caa = true; 01751 if(!strcmp(argv[i], "--csv")) csv = true; 01752 } 01753 01754 char dir[MAX_FILENAME_LENGTH]; 01755 strcpy(dir, argv[1]); 01756 if(!csv) 01757 { 01758 strcat(dir, "/HTML"); 01759 int res = mkdir(dir, S_IRWXU | S_IRWXG | S_IROTH | S_IXOTH); 01760 if(res!=0) 01761 { 01762 fprintf(stderr, "ERROR: Cannot create directory %s\naborting...\n", dir); 01763 exit(1); 01764 } 01765 } 01766 01767 DIR *dp; 01768 struct dirent *dirp; 01769 int num_of_modules = 0; 01770 if((dp = opendir(argv[1]))==NULL) 01771 { 01772 printf("Error(%d) opening %s\n", errno, argv[1]); 01773 return errno; 01774 } 01775 while((dirp = readdir(dp))!=NULL) 01776 { 01777 if(strstr(dirp->d_name, "_S_")!=NULL && strstr(dirp->d_name, ".txt.gz")!=NULL && !csv) 01778 { 01779 if(read_S_events(argv[1], dirp->d_name)) 01780 { 01781 fprintf(stderr, "ERROR: Cannot read file %s\naborting...\n", dirp->d_name); 01782 exit(1); 01783 } 01784 } 01785 } 01786 closedir(dp); 01787 sort(S_events.begin(), S_events.end()); 01788 if((dp = opendir(argv[1]))==NULL) 01789 { 01790 printf("Error(%d) opening %s\n", errno, argv[1]); 01791 return errno; 01792 } 01793 while((dirp = readdir(dp))!=NULL) 01794 { 01795 if(strstr(dirp->d_name, "_S_")!=NULL && strstr(dirp->d_name, ".txt.gz")!=NULL && !csv) 01796 { 01797 if(read_S_file(argv[1], dirp->d_name)) 01798 { 01799 fprintf(stderr, "ERROR: Cannot read file %s\naborting...\n", dirp->d_name); 01800 exit(1); 01801 } 01802 } 01803 else if(strstr(dirp->d_name, "_C_")!=NULL && strstr(dirp->d_name, ".txt")!=NULL) 01804 { 01805 int res = read_C_file(argv[1], dirp->d_name); 01806 if(res>num_of_modules) 01807 { 01808 num_of_modules = res; 01809 } 01810 } 01811 } 01812 closedir(dp); 01813 01814 if(!csv) 01815 { 01816 if(finalize_S_html_pages(argv[1])) 01817 { 01818 fprintf(stderr, "ERROR: Cannot finalize HTML pages!!!\naborting...\n"); 01819 exit(1); 01820 } 01821 } 01822 01823 char filepath[MAX_FILENAME_LENGTH]; 01824 bzero(filepath, MAX_FILENAME_LENGTH); 01825 if(!csv) sprintf(filepath, "%s/HTML/index.html", argv[1]); 01826 else sprintf(filepath, "%s/results.csv", argv[1]); 01827 FILE *fp = fopen(filepath, "w"); 01828 if(fp == NULL) 01829 { 01830 fprintf(stderr, "ERROR: Cannot create file index.html!!!\naborting...\n"); 01831 exit(1); 01832 } 01833 01834 if(caa) 01835 { 01836 double totalCycles; 01837 if(!nehalem) 01838 { 01839 init_core_caa_events(); 01840 if(!check_for_core_caa_events()) 01841 { 01842 fprintf(stderr, "(core) ERROR: One or more events for CAA missing!\naborting...\n"); 01843 exit(1); 01844 } 01845 init_core_caa_events_displ(); 01846 totalCycles = getTotalCycles(); 01847 calc_core_deriv_values(totalCycles); 01848 calc_post_deriv_values(); 01849 if(!csv) 01850 { 01851 put_C_header(fp, core_caa_events_displ); 01852 put_C_modules(fp, core_caa_events_displ); 01853 } 01854 else 01855 { 01856 put_C_header_csv(fp, core_caa_events_displ); 01857 put_C_modules_csv(fp, core_caa_events_displ); 01858 } 01859 } 01860 else 01861 { 01862 init_nhm_caa_events(); 01863 if(!check_for_nhm_caa_events()) 01864 { 01865 fprintf(stderr, "(nehalem) ERROR: One or more events for CAA missing!\naborting...\n"); 01866 exit(1); 01867 } 01868 init_nhm_caa_events_displ(); 01869 totalCycles = getTotalCycles(); 01870 calc_nhm_deriv_values(totalCycles); 01871 calc_post_deriv_values(); 01872 if(!csv) 01873 { 01874 put_C_header(fp, nhm_caa_events_displ); 01875 put_C_modules(fp, nhm_caa_events_displ); 01876 } 01877 else 01878 { 01879 put_C_header_csv(fp, nhm_caa_events_displ); 01880 put_C_modules_csv(fp, nhm_caa_events_displ); 01881 } 01882 } 01883 if(!csv) put_C_footer(fp); 01884 fclose(fp); 01885 } 01886 else 01887 { 01888 if(!csv) 01889 { 01890 put_C_header(fp, C_events); 01891 put_C_modules(fp, C_events); 01892 put_C_footer(fp); 01893 } 01894 else 01895 { 01896 put_C_header_csv(fp, C_events); 01897 put_C_modules_csv(fp, C_events); 01898 } 01899 fclose(fp); 01900 } 01901 if(!csv) 01902 { 01903 char src[MAX_FILENAME_LENGTH]; 01904 char dst[MAX_FILENAME_LENGTH]; 01905 sprintf(src, "sorttable.js"); 01906 sprintf(dst, "%s/HTML/sorttable.js", argv[1]); 01907 int fd_src = open(src, O_RDONLY); 01908 if(fd_src == -1) 01909 { 01910 fprintf(stderr, "ERROR: Cannot open file \"%s\"!\naborting...\n", src); 01911 exit(1); 01912 } 01913 int fd_dst = open(dst, O_WRONLY|O_CREAT|O_TRUNC, 0644); 01914 if(fd_dst == -1) 01915 { 01916 fprintf(stderr, "ERROR: Cannot open file \"%s\" (%s)!\naborting...\n", dst, strerror(errno)); 01917 exit(1); 01918 } 01919 char c; 01920 while(read(fd_src, &c, 1)) 01921 { 01922 write(fd_dst, &c, 1); 01923 } 01924 close(fd_dst); 01925 close(fd_src); 01926 } 01927 return 0; 01928 }
| double normalize | ( | std::string | field, | |
| double | value, | |||
| double | normalizeTo | |||
| ) |
Definition at line 1666 of file pfm_gen_analysis.cpp.
01667 { 01668 double max = 0; 01669 double counter_value; 01670 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01671 { 01672 counter_value = (it->second)[field]; 01673 if(max < counter_value) max = counter_value; 01674 } 01675 if(value>0 && max>0 && normalizeTo>0) 01676 { 01677 return 1.*value/max*normalizeTo; 01678 } 01679 else return 0; 01680 }
| void put_C_footer | ( | FILE * | fp | ) |
Definition at line 1619 of file pfm_gen_analysis.cpp.
| void put_C_header | ( | FILE * | fp, | |
| std::vector< std::string > & | columns | |||
| ) |
Definition at line 1565 of file pfm_gen_analysis.cpp.
01566 { 01567 fprintf(fp, "<!DOCTYPE html PUBLIC \"-//W3C//DTD HTML 4.01 Transitional//EN\" \"http://www.w3.org/TR/html4/loose.dtd\">\n"); 01568 fprintf(fp, "<html>\n"); 01569 fprintf(fp, "<head>\n"); 01570 fprintf(fp, "<title>\n"); 01571 fprintf(fp, "Analysis Result\n"); 01572 fprintf(fp, "</title>\n"); 01573 fprintf(fp, "<script src=\"sorttable.js\"></script>\n"); 01574 fprintf(fp, "<style>\ntable.sortable thead {\nbackground-color:#eee;\ncolor:#666666;\nfont-weight:bold;\ncursor:default;\nfont-family:courier;\n}\n</style>\n"); 01575 fprintf(fp, "</head>\n"); 01576 fprintf(fp, "<body link=\"black\">\n"); 01577 fprintf(fp, "<h1>RESULTS:</h1><br/>Click for detailed symbol view...<p/>\n"); 01578 fprintf(fp, "<table class=\"sortable\" cellpadding=\"5\">\n"); 01579 fprintf(fp, "<tr>\n"); 01580 fprintf(fp, "<th>MODULE NAME</th>\n"); 01581 for(std::vector<std::string>::const_iterator it = columns.begin(); it != columns.end(); ++it) 01582 { 01583 if(strlen(it->c_str())==0) fprintf(fp, "<th bgcolor=\"#FFFFFF\"> </th>\n"); 01584 else fprintf(fp, "<th>%s</th>\n", (*it).c_str()); 01585 } 01586 fprintf(fp, "</tr>\n"); 01587 return; 01588 }
| void put_C_header_csv | ( | FILE * | fp, | |
| std::vector< std::string > & | columns | |||
| ) |
Definition at line 1625 of file pfm_gen_analysis.cpp.
01626 { 01627 fprintf(fp, "MODULE NAME"); 01628 for(std::vector<std::string>::const_iterator it = columns.begin(); it != columns.end(); ++it) 01629 { 01630 if(strlen(it->c_str())==0); 01631 else fprintf(fp, ",%s", (*it).c_str()); 01632 } 01633 fprintf(fp, "\n"); 01634 return; 01635 }
| void put_C_modules | ( | FILE * | fp, | |
| std::vector< std::string > & | columns | |||
| ) |
Definition at line 1590 of file pfm_gen_analysis.cpp.
01591 { 01592 int index = 0; 01593 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01594 { 01595 if(index%2) fprintf(fp, "<tr bgcolor=\"#FFFFCC\">\n"); 01596 else fprintf(fp, "<tr bgcolor=\"#CCFFCC\">\n"); 01597 fprintf(fp, "<td style=\"font-family:monospace;font-size:large;color:Black\"><a href=\"%s.html\">%s</a></td>\n", (it->first).c_str(), (it->first).c_str()); 01598 for(std::vector<std::string>::const_iterator jt = columns.begin(); jt != columns.end(); ++jt) 01599 { 01600 if(strlen(jt->c_str())==0) 01601 { 01602 fprintf(fp, "<td bgcolor=\"#FFFFFF\"> </td>"); 01603 } 01604 else 01605 { 01606 if((it->second).find(*jt) == (it->second).end()) 01607 { 01608 fprintf(stderr, "ERROR: Cannot find derivate value \"%s\"!!!\naborting...\n", (*jt).c_str()); 01609 exit(1); 01610 } 01611 fprintf(fp, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\" align=\"right\">%.2f</td>\n", (it->second)[*jt]); 01612 } 01613 } 01614 fprintf(fp, "</tr>\n"); 01615 index++; 01616 } 01617 }
| void put_C_modules_csv | ( | FILE * | fp, | |
| std::vector< std::string > & | columns | |||
| ) |
Definition at line 1637 of file pfm_gen_analysis.cpp.
01638 { 01639 for(std::map<std::string, std::map<std::string, double> >::iterator it = C_modules.begin(); it != C_modules.end(); ++it) 01640 { 01641 fprintf(fp, "%s", (it->first).c_str()); 01642 for(std::vector<std::string>::const_iterator jt = columns.begin(); jt != columns.end(); ++jt) 01643 { 01644 if(strlen(jt->c_str())==0); 01645 else 01646 { 01647 if((it->second).find(*jt) == (it->second).end()) 01648 { 01649 fprintf(stderr, "ERROR: Cannot find derivate value \"%s\"!!!\naborting...\n", (*jt).c_str()); 01650 exit(1); 01651 } 01652 fprintf(fp, ",%.2f", (it->second)[*jt]); 01653 } 01654 } 01655 fprintf(fp, "\n"); 01656 } 01657 }
| void put_S_module | ( | S_module * | cur_module, | |
| const char * | dir | |||
| ) |
Definition at line 1133 of file pfm_gen_analysis.cpp.
01134 { 01135 char module_name[MAX_MODULE_NAME_LENGTH]; 01136 bzero(module_name, MAX_MODULE_NAME_LENGTH); 01137 strcpy(module_name, (cur_module->get_module_name()).c_str()); 01138 char module_filename[MAX_FILENAME_LENGTH]; 01139 bzero(module_filename, MAX_FILENAME_LENGTH); 01140 strcpy(module_filename, dir); 01141 strcat(module_filename, "/HTML/"); 01142 strcat(module_filename, module_name); 01143 strcat(module_filename, ".html"); 01144 char event[MAX_EVENT_NAME_LENGTH]; 01145 bzero(event, MAX_EVENT_NAME_LENGTH); 01146 strcpy(event, (cur_module->get_event()).c_str()); 01147 std::map<std::string, unsigned int>::iterator result = modules_tot_samples.find(cur_module->get_module_name()); 01148 FILE *module_file; 01149 if(result == modules_tot_samples.end()) //not found 01150 { 01151 if((!strcmp(event, "UNHALTED_CORE_CYCLES") && !nehalem) || (!strcmp(event, "CPU_CLK_UNHALTED:THREAD_P") && nehalem)) 01152 { 01153 modules_tot_samples.insert(std::pair<std::string, unsigned int>(cur_module->get_module_name(), cur_module->get_total_num_samples())); 01154 } 01155 else 01156 { 01157 modules_tot_samples.insert(std::pair<std::string, unsigned int>(cur_module->get_module_name(), 0)); 01158 } 01159 module_file = fopen(module_filename, "w"); 01160 if(module_file == NULL) 01161 { 01162 fprintf(stderr, "ERROR: Cannot create file %s!!!\naborting...\n", module_filename); 01163 exit(1); 01164 } 01165 fprintf(module_file, "<!DOCTYPE html PUBLIC \"-//W3C//DTD HTML 4.01 Transitional//EN\" \"http://www.w3.org/TR/html4/loose.dtd\">\n"); 01166 fprintf(module_file, "<html>\n"); 01167 fprintf(module_file, "<head>\n"); 01168 fprintf(module_file, "<title>\n"); 01169 fprintf(module_file, "%s\n", module_name); 01170 fprintf(module_file, "</title>\n"); 01171 fprintf(module_file, "</head>\n"); 01172 fprintf(module_file, "<body>\n"); 01173 fprintf(module_file, "<h2>%s</h2><br/>Events Sampled:<br/>\n", module_name); 01174 fprintf(module_file, "<ul>\n"); 01175 for(std::vector<std::string>::const_iterator it = S_events.begin(); it != S_events.end(); ++it) 01176 { 01177 fprintf(module_file, "<li><a href=\"#%s\">%s</a></li>\n", it->c_str(), it->c_str()); 01178 } 01179 fprintf(module_file, "</ul>\n"); 01180 }// if(result == modules_tot_samples.end()) //not found 01181 else 01182 { 01183 if((!strcmp(event, "UNHALTED_CORE_CYCLES") && !nehalem) || (!strcmp(event, "CPU_CLK_UNHALTED:THREAD_P") && nehalem)) 01184 { 01185 modules_tot_samples[cur_module->get_module_name()] = cur_module->get_total_num_samples(); 01186 } 01187 module_file = fopen(module_filename, "a"); 01188 }//else:: if(result != modules_tot_samples.end()) //found!! 01189 char event_str[MAX_EVENT_NAME_LENGTH]; 01190 bzero(event_str, MAX_EVENT_NAME_LENGTH); 01191 strcpy(event_str, event); 01192 if(cur_module->get_c_mask()>0) 01193 { 01194 sprintf(event_str, "%s CMASK=%d", event_str, cur_module->get_c_mask()); 01195 } 01196 if(cur_module->get_inv_mask()>0) 01197 { 01198 sprintf(event_str, "%s INV=%d", event_str, cur_module->get_inv_mask()); 01199 } 01200 fprintf(module_file, "<a name=\"%s\"><a>\n", event_str); 01201 fprintf(module_file, "<table cellpadding=\"5\">\n"); 01202 fprintf(module_file, "<tr bgcolor=\"#EEEEEE\">\n"); 01203 fprintf(module_file, "<th colspan=\"6\" align=\"left\">%s -- cmask: %u -- invmask: %u -- Total Samples: %u -- Sampling Period: %d</th>\n", event, cur_module->get_c_mask(), cur_module->get_inv_mask(), cur_module->get_total_num_samples(), cur_module->get_smpl_period()); 01204 fprintf(module_file, "</tr>\n"); 01205 fprintf(module_file, "<tr bgcolor=\"#EEEEEE\">\n"); 01206 fprintf(module_file, "<th align=\"left\">Samples</th>\n"); 01207 fprintf(module_file, "<th align=\"left\">Percentage</th>\n"); 01208 fprintf(module_file, "<th align=\"left\">Symbol Name</th>\n"); 01209 fprintf(module_file, "<th align=\"left\">Library Name</th>\n"); 01210 fprintf(module_file, "<th align=\"left\">Complete Signature</th>\n"); 01211 fprintf(module_file, "<th align=\"left\">Library Pathname</th>\n"); 01212 fprintf(module_file, "</tr>\n"); 01213 for(int j=0; j<20; j++) 01214 { 01215 char sym[MAX_SYM_LENGTH]; 01216 char sym_mod[MAX_SYM_MOD_LENGTH]; 01217 char lib[MAX_LIB_LENGTH]; 01218 char lib_mod[MAX_LIB_MOD_LENGTH]; 01219 char simple_sym[MAX_SIMPLE_SYM_LENGTH]; 01220 char simple_sym_mod[MAX_SIMPLE_SYM_MOD_LENGTH]; 01221 char simple_lib[MAX_SIMPLE_LIB_LENGTH]; 01222 char simple_lib_mod[MAX_SIMPLE_LIB_MOD_LENGTH]; 01223 01224 bzero(sym, MAX_SYM_LENGTH); 01225 bzero(sym_mod, MAX_SYM_MOD_LENGTH); 01226 bzero(lib, MAX_LIB_LENGTH); 01227 bzero(lib_mod, MAX_LIB_MOD_LENGTH); 01228 bzero(simple_sym, MAX_SIMPLE_SYM_LENGTH); 01229 bzero(simple_sym_mod, MAX_SIMPLE_SYM_MOD_LENGTH); 01230 bzero(simple_lib, MAX_SIMPLE_LIB_LENGTH); 01231 bzero(simple_lib_mod, MAX_SIMPLE_LIB_MOD_LENGTH); 01232 01233 char index[MAX_SAMPLE_INDEX_LENGTH]; 01234 bzero(index, MAX_SAMPLE_INDEX_LENGTH); 01235 unsigned int value; 01236 bool res = cur_module->get_max(index, &value); 01237 if(!res) break; 01238 char *sym_end = strchr(index, '%'); 01239 if(sym_end==NULL) //error 01240 { 01241 fprintf(stderr, "ERROR: Invalid sym and lib name! : %s\naborting...\n", index); 01242 exit(1); 01243 } 01244 strncpy(sym, index, strlen(index)-strlen(sym_end)); 01245 strcpy(lib, sym_end+1); 01246 char temp[MAX_SYM_LENGTH]; 01247 bzero(temp, MAX_SYM_LENGTH); 01248 strcpy(temp, sym); 01249 strcpy(simple_sym, (func_name(temp))); 01250 if(strrchr(lib, '/')!=NULL && *(strrchr(lib, '/')+1)!='\0') 01251 { 01252 strcpy(simple_lib, strrchr(lib, '/')+1); 01253 } 01254 else 01255 { 01256 strcpy(simple_lib, lib); 01257 } 01258 if(j%2!=0) 01259 { 01260 fprintf(module_file, "<tr bgcolor=\"#FFFFCC\">\n"); 01261 } 01262 else 01263 { 01264 fprintf(module_file, "<tr bgcolor=\"#CCFFCC\">\n"); 01265 } 01266 fprintf(module_file, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\">%u</td>\n", value); 01267 fprintf(module_file, "<td style=\"font-family:monospace;font-size:large;color:DarkBlue\">%f%%</td>\n", (((double)(value))/((double)(cur_module->get_total_num_samples())))*100); 01268 html_special_chars(simple_sym, simple_sym_mod); 01269 html_special_chars(simple_lib, simple_lib_mod); 01270 html_special_chars(sym, sym_mod); 01271 html_special_chars(lib, lib_mod); 01272 fprintf(module_file, "<td style=\"font-family:courier;\">%s</td>\n", simple_sym_mod); 01273 fprintf(module_file, "<td style=\"font-family:courier;\">%s</td>\n", simple_lib_mod); 01274 fprintf(module_file, "<td style=\"font-family:courier;\">%s</td>\n", sym_mod); 01275 fprintf(module_file, "<td style=\"font-family:courier;\">%s</td>\n</tr>\n", lib_mod); 01276 } 01277 fprintf(module_file, "</table><br/><br/>\n"); 01278 int res = fclose(module_file); 01279 if(res) 01280 { 01281 fprintf(stderr, "ERROR: Cannot close file %s!!!\naborting...\n", module_filename); 01282 exit(1); 01283 } 01284 return; 01285 }
| int read_C_file | ( | const char * | dir, | |
| const char * | filename | |||
| ) |
Definition at line 1501 of file pfm_gen_analysis.cpp.
01502 { 01503 char event[MAX_EVENT_NAME_LENGTH]; 01504 char arch[MAX_ARCH_NAME_LENGTH]; 01505 char line[MAX_LINE_LENGTH]; 01506 char cmask_str[MAX_CMASK_STR_LENGTH]; 01507 char inv_str[MAX_INV_STR_LENGTH]; 01508 char sp_str[MAX_SP_STR_LENGTH]; 01509 char cur_module_name[MAX_MODULE_NAME_LENGTH]; 01510 bzero(line, MAX_LINE_LENGTH); 01511 bzero(event, MAX_EVENT_NAME_LENGTH); 01512 bzero(cur_module_name, MAX_MODULE_NAME_LENGTH); 01513 bzero(arch, MAX_ARCH_NAME_LENGTH); 01514 bzero(line, MAX_LINE_LENGTH); 01515 bzero(cmask_str, MAX_CMASK_STR_LENGTH); 01516 bzero(inv_str, MAX_INV_STR_LENGTH); 01517 bzero(sp_str, MAX_SP_STR_LENGTH); 01518 int number_of_modules = 0; 01519 long cur_sum = 0; 01520 int no_of_values = 0; 01521 char path_name[MAX_FILENAME_LENGTH]; 01522 bzero(path_name, MAX_FILENAME_LENGTH); 01523 strcpy(path_name, dir); 01524 strcat(path_name, "/"); 01525 strcat(path_name, filename); 01526 FILE *fp = fopen(path_name, "r"); 01527 fscanf(fp, "%s %s %s %s %s\n", arch, event, cmask_str, inv_str, sp_str); 01528 if(!strcmp(arch, "NHM")) nehalem = true; else nehalem = false; 01529 std::string event_str(event); 01530 if(atoi(cmask_str)>0) 01531 { 01532 event_str += " CMASK="; 01533 event_str += cmask_str; 01534 } 01535 if(atoi(inv_str)>0) 01536 { 01537 event_str += " INV="; 01538 event_str += inv_str; 01539 } 01540 C_events.push_back(event_str); 01541 while(fscanf(fp, "%s\n", line)!=EOF) 01542 { 01543 if(isalpha(line[0])) //module 01544 { 01545 if(number_of_modules>0) 01546 { 01547 C_modules[cur_module_name][event_str]=(double)cur_sum/no_of_values; 01548 cur_sum = 0; 01549 no_of_values = 0; 01550 } 01551 strcpy(cur_module_name, line); 01552 number_of_modules++; 01553 } 01554 else if(isdigit(line[0])) //value 01555 { 01556 cur_sum += strtol(line, NULL, 10); 01557 no_of_values++; 01558 } 01559 } 01560 C_modules[cur_module_name][event_str]=(double)cur_sum/no_of_values; //last module 01561 fclose(fp); 01562 return number_of_modules; 01563 }
| int read_S_events | ( | const char * | dir, | |
| const char * | filename | |||
| ) |
Definition at line 1421 of file pfm_gen_analysis.cpp.
01422 { 01423 char event[MAX_EVENT_NAME_LENGTH]; 01424 char arch[MAX_ARCH_NAME_LENGTH]; 01425 char line[MAX_LINE_LENGTH]; 01426 char cmask_str[MAX_CMASK_STR_LENGTH]; 01427 char inv_str[MAX_INV_STR_LENGTH]; 01428 char sp_str[MAX_SP_STR_LENGTH]; 01429 bzero(line, MAX_LINE_LENGTH); 01430 bzero(event, MAX_EVENT_NAME_LENGTH); 01431 bzero(arch, MAX_ARCH_NAME_LENGTH); 01432 bzero(cmask_str, MAX_CMASK_STR_LENGTH); 01433 bzero(inv_str, MAX_INV_STR_LENGTH); 01434 bzero(sp_str, MAX_SP_STR_LENGTH); 01435 char path_name[MAX_FILENAME_LENGTH]; 01436 bzero(path_name, MAX_FILENAME_LENGTH); 01437 strcpy(path_name, dir); 01438 strcat(path_name, "/"); 01439 strcat(path_name, filename); 01440 gzFile res_file = gzopen(path_name, "rb"); 01441 if(res_file != NULL) 01442 { 01443 bzero(line, MAX_LINE_LENGTH); 01444 gzgets(res_file, line, MAX_LINE_LENGTH); 01445 if(line[strlen(line)-1]=='\n') line[strlen(line)-1]='\0'; 01446 bzero(event, MAX_EVENT_NAME_LENGTH); 01447 sscanf(line, "%s %s %s %s %s\n", arch, event, cmask_str, inv_str, sp_str); 01448 std::string event_str(event); 01449 if(atoi(cmask_str)>0) 01450 { 01451 event_str += " CMASK="; 01452 event_str += cmask_str; 01453 } 01454 if(atoi(inv_str)>0) 01455 { 01456 event_str += " INV="; 01457 event_str += inv_str; 01458 } 01459 S_events.push_back(event_str); 01460 }// if(res_file != NULL) 01461 else 01462 { 01463 fprintf(stderr, "ERROR: Unable to open input file: %s\naborting...\n", filename); 01464 exit(1); 01465 } 01466 return 0; 01467 }
| int read_S_file | ( | const char * | dir, | |
| const char * | filename | |||
| ) |
Definition at line 1294 of file pfm_gen_analysis.cpp.
01295 { 01296 char line[MAX_LINE_LENGTH]; 01297 char event[MAX_EVENT_NAME_LENGTH]; 01298 char arch[MAX_ARCH_NAME_LENGTH]; 01299 unsigned int cmask; 01300 unsigned int inv; 01301 unsigned int sp; 01302 char cur_module_name[MAX_MODULE_NAME_LENGTH]; 01303 bzero(line, MAX_LINE_LENGTH); 01304 bzero(event, MAX_EVENT_NAME_LENGTH); 01305 bzero(cur_module_name, MAX_MODULE_NAME_LENGTH); 01306 bzero(arch, MAX_ARCH_NAME_LENGTH); 01307 01308 S_module *cur_module = new S_module(); 01309 unsigned int module_num = 0; 01310 01311 char path_name[MAX_FILENAME_LENGTH]; 01312 bzero(path_name, MAX_FILENAME_LENGTH); 01313 strcpy(path_name, dir); 01314 strcat(path_name, "/"); 01315 strcat(path_name, filename); 01316 gzFile res_file = gzopen(path_name, "rb"); 01317 01318 if(res_file != NULL) 01319 { 01320 bzero(line, MAX_LINE_LENGTH); 01321 gzgets(res_file, line, MAX_LINE_LENGTH); 01322 if(line[strlen(line)-1]=='\n') line[strlen(line)-1]='\0'; 01323 bzero(event, MAX_EVENT_NAME_LENGTH); 01324 sscanf(line, "%s %s %u %u %u", arch, event, &cmask, &inv, &sp); 01325 if(!strcmp(arch, "NHM")) nehalem = true; else nehalem = false; 01326 bzero(line, MAX_LINE_LENGTH); 01327 while(gzgets(res_file, line, MAX_LINE_LENGTH)!=Z_NULL) 01328 { 01329 if(line[strlen(line)-1]=='\n') line[strlen(line)-1]='\0'; 01330 if(strchr(line, ' ')==NULL) //module 01331 { 01332 if(module_num>0) 01333 { 01334 put_S_module(cur_module, dir); 01335 cur_module->clear(); 01336 } 01337 module_num++; 01338 char *end_sym = strchr(line, '%'); 01339 if(end_sym == NULL) //error 01340 { 01341 fprintf(stderr, "ERROR: Invalid module name. \nLINE: %s\naborting...\n", line); 01342 exit(1); 01343 } 01344 bzero(cur_module_name, MAX_MODULE_NAME_LENGTH); 01345 strncpy(cur_module_name, line, strlen(line)-strlen(end_sym)); 01346 cur_module->init(cur_module_name, arch, event, cmask, inv, sp); 01347 cur_module->set_total(atoi(end_sym+1)); 01348 } //module 01349 else //symbol, libName, libOffset, value 01350 { 01351 unsigned int value=0, libOffset=0; 01352 char symbol[MAX_SYM_LENGTH]; 01353 char libName[MAX_LIB_LENGTH]; 01354 char final_sym[MAX_SYM_MOD_LENGTH]; 01355 char final_lib[MAX_LIB_MOD_LENGTH]; 01356 bzero(symbol, MAX_SYM_LENGTH); 01357 bzero(libName, MAX_LIB_LENGTH); 01358 bzero(final_sym, MAX_SYM_MOD_LENGTH); 01359 bzero(final_lib, MAX_LIB_MOD_LENGTH); 01360 01361 sscanf(line, "%s %s %u %u", symbol, libName, &libOffset, &value); 01362 char realPathName_s[FILENAME_MAX]; 01363 bzero(realPathName_s, FILENAME_MAX); 01364 char *realPathName = realpath(libName, realPathName_s); 01365 if(realPathName!=NULL && strlen(realPathName)>0) 01366 { 01367 std::map<std::string, FileInfo>::iterator result; 01368 result = libsInfo.find(realPathName); 01369 if(result == libsInfo.end()) 01370 { 01371 libsInfo[realPathName] = FileInfo(realPathName, true); 01372 } 01373 const char *temp_sym = libsInfo[realPathName].symbolByOffset(libOffset); 01374 if(temp_sym!=NULL && strlen(temp_sym)>0) 01375 { 01376 int status; 01377 char *demangled_symbol = abi::__cxa_demangle(temp_sym, NULL, NULL, &status); 01378 if(status == 0) 01379 { 01380 strcpy(final_sym, demangled_symbol); 01381 free(demangled_symbol); 01382 } 01383 else 01384 { 01385 strcpy(final_sym, temp_sym); 01386 } 01387 } 01388 else 01389 { 01390 strcpy(final_sym, "???"); 01391 } 01392 strcpy(final_lib, realPathName); 01393 } 01394 else 01395 { 01396 strcpy(final_sym, symbol); 01397 strcpy(final_lib, libName); 01398 } 01399 char index[MAX_LINE_LENGTH]; 01400 bzero(index, MAX_LINE_LENGTH); 01401 strcpy(index, final_sym); 01402 strcat(index, "%"); 01403 strcat(index, final_lib); 01404 cur_module->add_sample(index, value); 01405 }// symbol, libName, libOffset, value 01406 bzero(line, MAX_LINE_LENGTH); 01407 }// while(gzgets(res_file, line, MAX_LINE_LENGTH)!=Z_NULL) 01408 put_S_module(cur_module, dir);//last module! 01409 cur_module->clear(); 01410 gzclose(res_file); 01411 }// if(res_file != NULL) 01412 else 01413 { 01414 fprintf(stderr, "ERROR: Unable to open input file: %s\naborting...\n", filename); 01415 exit(1); 01416 } 01417 delete cur_module; //delete it! 01418 return 0; 01419 }
| bool skipString | ( | const char * | strptr, | |
| const char * | srcbuffer, | |||
| const char ** | dstbuffer | |||
| ) |
Definition at line 153 of file pfm_gen_analysis.cpp.
| bool skipWhitespaces | ( | const char * | srcbuffer, | |
| const char ** | destbuffer | |||
| ) |
Definition at line 132 of file pfm_gen_analysis.cpp.
std::vector<std::string> C_events [static] |
Definition at line 315 of file pfm_gen_analysis.cpp.
std::map<std::string, std::map<std::string, double> > C_modules [static] |
Definition at line 314 of file pfm_gen_analysis.cpp.
std::vector<std::string> core_caa_events [static] |
Definition at line 318 of file pfm_gen_analysis.cpp.
std::vector<std::string> core_caa_events_displ [static] |
Definition at line 320 of file pfm_gen_analysis.cpp.
std::map<std::string, FileInfo> libsInfo [static] |
Definition at line 311 of file pfm_gen_analysis.cpp.
std::map<std::string, unsigned int> modules_tot_samples [static] |
Definition at line 310 of file pfm_gen_analysis.cpp.
int nehalem [static] |
Definition at line 312 of file pfm_gen_analysis.cpp.
std::vector<std::string> nhm_caa_events [static] |
Definition at line 319 of file pfm_gen_analysis.cpp.
std::vector<std::string> nhm_caa_events_displ [static] |
Definition at line 321 of file pfm_gen_analysis.cpp.
std::vector<std::string> S_events [static] |
Definition at line 316 of file pfm_gen_analysis.cpp.