-modifying the coefficients parsing code to be more readable, also modifies how it is set in the XML configuration file

This code modification avoids use of the string::assign function that seems to causing memory problems

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 14886]

4 files changed, 103 insertions, 38 deletions

diff --git a/configs/GTX480/gpuwattch_gtx480.xml b/configs/GTX480/gpuwattch_gtx480.xml
index 9a3f8a1..af24630 100755
--- a/configs/GTX480/gpuwattch_gtx480.xml
+++ b/configs/GTX480/gpuwattch_gtx480.xml
@@ -29,7 +29,37 @@
 		<param name="physical_address_width" value="32"/>
 		<param name="virtual_memory_page_size" value="4096"/>
 		<param name="idle_core_power" value="1.59"/><!-- idle core power for GTX479 -->
-		<param name="scaling_coefficients" value="10,0.0884816,10,10,8,10,4.12782,10,2.48832,10,10,10,4.29982,0.387764,0.0714269,0.14302,0.01,0.546811,0.485351,0.806633,0.818073,1.9207,100,100,100,87.9303,100,10,4.3548,10"/>	
+		<!--param name="scaling_coefficients" value="10,0.0884816,10,10,8,10,4.12782,10,2.48832,10,10,10,4.29982,0.387764,0.0714269,0.14302,0.01,0.546811,0.485351,0.806633,0.818073,1.9207,100,100,100,87.9303,100,10,4.3548,10"/-->
+		<param name="TOT_INST" value="10"/>
+		<param name="FP_INT" value="0.0884816"/>
+		<param name="IC_H" value="10"/>
+		<param name="IC_M" value="10"/>
+		<param name="DC_RH" value="8"/>
+		<param name="DC_RM" value="10"/>
+		<param name="DC_WH" value="4.12782"/>
+		<param name="DC_WM" value="10"/>
+		<param name="TC_H" value="2.48832"/>
+		<param name="TC_M" value="10"/>
+		<param name="CC_H" value="10"/>
+		<param name="CC_M" value="10"/>
+		<param name="SHRD_ACC" value="4.29982"/>
+		<param name="REG_RD" value="0.387764"/>
+		<param name="REG_WR" value="0.0714269"/>
+		<param name="NON_REG_OPs" value="0.14302"/>
+		<param name="SP_ACC" value="0.01"/>
+		<param name="SFU_ACC" value="0.546811"/>
+		<param name="FPU_ACC" value="0.485351"/>
+		<param name="MEM_RD" value="0.806633"/>
+		<param name="MEM_WR" value="0.818073"/>
+		<param name="MEM_PRE" value="1.9207"/>
+		<param name="L2_RH" value="100"/>
+		<param name="L2_RM" value="100"/>
+		<param name="L2_WH" value="100"/>
+		<param name="L2_WM" value="87.9303"/>
+		<param name="NOC_A" value="100"/>
+		<param name="PIPE_A" value="10"/>
+		<param name="IDLE_CORE_N" value="4.3548"/>
+		<param name="CONST_DYNAMICN" value="10"/>	
 		<stat name="num_idle_cores" value="0"/><!-- Average Number of idle cores during this period -->
 		<stat name="total_cycles" value="total_cycles_match_mcpat"/>
 		<stat name="idle_cycles" value="idle_cycles_match_mcpat"/>
diff --git a/src/gpuwattch/XML_Parse.cc b/src/gpuwattch/XML_Parse.cc
index 0b82822..ef475d4 100644
--- a/src/gpuwattch/XML_Parse.cc
+++ b/src/gpuwattch/XML_Parse.cc
@@ -103,7 +103,39 @@ void ParseXML::parse(char* filepath)
 		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"virtual_address_width")==0) {sys.virtual_address_width=atoi(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

 		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"physical_address_width")==0) {sys.physical_address_width=atoi(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

 		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"virtual_memory_page_size")==0) {sys.virtual_memory_page_size=atoi(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

-		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"idle_core_power")==0) {sys.idle_core_power=(double)atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"idle_core_power")==0) {sys.idle_core_power=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[TOT_INST])==0) {sys.scaling_coefficients[TOT_INST]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[FP_INT])==0) {sys.scaling_coefficients[FP_INT]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[IC_H])==0) {sys.scaling_coefficients[IC_H]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[IC_M])==0) {sys.scaling_coefficients[IC_M]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[DC_RH])==0) {sys.scaling_coefficients[DC_RH]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[DC_RM])==0) {sys.scaling_coefficients[DC_RM]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[DC_WH])==0) {sys.scaling_coefficients[DC_WH]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[DC_WM])==0) {sys.scaling_coefficients[DC_WM]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[TC_H])==0) {sys.scaling_coefficients[TC_H]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[TC_M])==0) {sys.scaling_coefficients[TC_M]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[CC_H])==0) {sys.scaling_coefficients[CC_H]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[CC_M])==0) {sys.scaling_coefficients[CC_M]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[SHRD_ACC])==0) {sys.scaling_coefficients[SHRD_ACC]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[REG_RD])==0) {sys.scaling_coefficients[REG_RD]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[REG_WR])==0) {sys.scaling_coefficients[REG_WR]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[NON_REG_OPs])==0) {sys.scaling_coefficients[NON_REG_OPs]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[SP_ACC])==0) {sys.scaling_coefficients[SP_ACC]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[SFU_ACC])==0) {sys.scaling_coefficients[SFU_ACC]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[FPU_ACC])==0) {sys.scaling_coefficients[FPU_ACC]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[MEM_RD])==0) {sys.scaling_coefficients[MEM_RD]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[MEM_WR])==0) {sys.scaling_coefficients[MEM_WR]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[MEM_PRE])==0) {sys.scaling_coefficients[MEM_PRE]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[L2_RH])==0) {sys.scaling_coefficients[L2_RH]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[L2_RM])==0) {sys.scaling_coefficients[L2_RM]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[L2_WH])==0) {sys.scaling_coefficients[L2_WH]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[L2_WM])==0) {sys.scaling_coefficients[L2_WM]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[NOC_A])==0) {sys.scaling_coefficients[NOC_A]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[PIPE_A])==0) {sys.scaling_coefficients[PIPE_A]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[IDLE_CORE_N])==0) {sys.scaling_coefficients[IDLE_CORE_N]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+        if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),perf_count_label[CONST_DYNAMICN])==0) {sys.scaling_coefficients[CONST_DYNAMICN]=atof(xNode2.getChildNode("param",i).getAttribute("value"));continue;}

+

+/*

 		if (strcmp(xNode2.getChildNode("param",i).getAttribute("name"),"scaling_coefficients")==0)

 		{

 			strtmp.assign(xNode2.getChildNode("param",i).getAttribute("value"));

@@ -126,6 +158,7 @@ void ParseXML::parse(char* filepath)
 			chtmp1[0]='\0';

 			continue;

 		}

+*/

 	}

 

 //	if (sys.Private_L2 && sys.number_of_cores!=sys.number_of_L2s)

@@ -1701,6 +1734,7 @@ void ParseXML::initialize() //Initialize all
 	sys.opt_clockrate=true;

 	sys.opt_area=false;

 	sys.interconnect_projection_type=1;

+	sys.idle_core_power=0;

 	int i,j;

 	for (i=0; i<=63; i++)

 	{

diff --git a/src/gpuwattch/XML_Parse.h b/src/gpuwattch/XML_Parse.h
index 7921719..53e408b 100644
--- a/src/gpuwattch/XML_Parse.h
+++ b/src/gpuwattch/XML_Parse.h
@@ -56,6 +56,42 @@ ToXMLStringTool tx,tx2;
 */

 //all subnodes at the level of system.core(0-n)

 //cache_policy is added into cache property arrays;//0 no write or write-though with non-write allocate;1 write-back with write-allocate

+static const char * perf_count_label[] = {"TOT_INST,", "FP_INT,", "IC_H,", "IC_M,", "DC_RH,", "DC_RM,", "DC_WH,", "DC_WM,",

+                                "TC_H,", "TC_M,", "CC_H,", "CC_M,", "SHRD_ACC,", "REG_RD,", "REG_WR,", "NON_REG_OPs,",

+                                "SP_ACC,", "SFU_ACC,", "FPU_ACC,", "MEM_RD,","MEM_WR,", "MEM_PRE,", "L2_RH,", "L2_RM,", "L2_WH,",

+                                "L2_WM,", "NOC_A,", "PIPE_A,", "IDLE_CORE_N,", "CONST_DYNAMICN"};

+enum perf_count_t {

+   TOT_INST=0,

+   FP_INT,

+   IC_H,

+   IC_M,

+   DC_RH,

+   DC_RM,

+   DC_WH,

+   DC_WM,

+   TC_H,

+   TC_M,

+   CC_H,

+   CC_M,

+   SHRD_ACC,

+   REG_RD,

+   REG_WR,

+   NON_REG_OPs,

+   SP_ACC,

+   SFU_ACC,

+   FPU_ACC,

+   MEM_RD,

+   MEM_WR,

+   MEM_PRE,

+   L2_RH,

+   L2_RM,

+   L2_WH,

+   L2_WM,

+   NOC_A,

+   PIPE_A,

+   IDLE_CORE_N,

+   CONST_DYNAMICN

+};

 

 typedef struct{

 	int prediction_width;

diff --git a/src/gpuwattch/gpgpu_sim_wrapper.cc b/src/gpuwattch/gpgpu_sim_wrapper.cc
index 2792da0..ffd4650 100644
--- a/src/gpuwattch/gpgpu_sim_wrapper.cc
+++ b/src/gpuwattch/gpgpu_sim_wrapper.cc
@@ -31,42 +31,7 @@
 #define NUM_COMPONENTS_MODELLED 18
 #define NUM_PERFORMANCE_COUNTERS 30
 
-static const char * perf_count_label[] = {"TOT_INST,", "FP_INT,", "IC_H,", "IC_M,", "DC_RH,", "DC_RM,", "DC_WH,", "DC_WM,",
-								"TC_H,", "TC_M,", "CC_H,", "CC_M,", "SHRD_ACC,", "REG_RD,", "REG_WR,", "NON_REG_OPs,",
-								"SP_ACC,", "SFU_ACC,", "FPU_ACC,", "MEM_RD,","MEM_WR,", "MEM_PRE,", "L2_RH,", "L2_RM,", "L2_WH,",
-								"L2_WM,", "NOC_A,", "PIPE_A,", "IDLE_CORE_N,", "CONST_DYNAMICN"};
-enum perf_count_t {
-   TOT_INST=0,
-   FP_INT,
-   IC_H,
-   IC_M,
-   DC_RH,
-   DC_RM,
-   DC_WH,
-   DC_WM,
-   TC_H,
-   TC_M,
-   CC_H,
-   CC_M,
-   SHRD_ACC,
-   REG_RD,
-   REG_WR,
-   NON_REG_OPs,
-   SP_ACC,
-   SFU_ACC,
-   FPU_ACC,
-   MEM_RD,
-   MEM_WR,
-   MEM_PRE,
-   L2_RH,
-   L2_RM,
-   L2_WH,
-   L2_WM,
-   NOC_A,
-   PIPE_A,
-   IDLE_CORE_N,
-   CONST_DYNAMICN
-};
+
 
 static const char * pwr_cmp_label[] = {"IBP,", "ICP,", "DCP,", "TCP,", "CCP,", "SHRDP,", "RFP,", "SPP,",
 								"SFUP,", "FPUP,", "SCHEDP,", "L2CP,", "MCP,", "NOCP,", "DRAMP,",