Merge branch 'FEATURE-0242-HardwareIdle-Mothership' of github.com:POE…

…TSII/Orchestrator into FEATURE-0242-HardwareIdle-Mothership

Source/OrchBase/Composer.cpp

@@ -38,6 +38,9 @@ ComposerGraphI_t::ComposerGraphI_t()
 
     compilationFlags = "";
     provenanceCache = "";
+
+    idleInstructionBinary = "";
+    idleDataBinary = "";
 }
 
 ComposerGraphI_t::ComposerGraphI_t(GraphI_t* graphIIn, std::string& outputPath)
@@ -59,6 +62,9 @@ ComposerGraphI_t::ComposerGraphI_t(GraphI_t* graphIIn, std::string& outputPath)
 
     compilationFlags = "";
     provenanceCache = "";
+
+    idleInstructionBinary = "";
+    idleDataBinary = "";
 }
 
 ComposerGraphI_t::~ComposerGraphI_t()
@@ -126,6 +132,11 @@ void ComposerGraphI_t::Dump(unsigned off,FILE* file)
         default:        fprintf(file, "**INVALID**\n");
     }
 
+
+    fprintf(file, "  Hardware Idle Instructions: %s\n",
+                    idleInstructionBinary.c_str());
+    fprintf(file, "  Hardware Idle Data:         %s\n",idleDataBinary.c_str());
+
     fprintf(file, "\nNitty gritty details:\n");
     fprintf(file, "  Device type strs map size:  %lu \n",
                         static_cast<unsigned long>(devTStrsMap.size()));
@@ -1012,6 +1023,34 @@ bool Composer::isCompiled(GraphI_t* graphI)
     return builderGraphI->compiled;
 }
 
+/******************************************************************************
+ * Public method to get the paths for the HW Idle binaries
+ *****************************************************************************/
+bool Composer::getDummyPaths(GraphI_t* graphI, std::string& instrBin, 
+                                std::string& dataBin)
+{
+    ComposerGraphI_t* builderGraphI;
+
+    ComposerGraphIMap_t::iterator srch = graphIMap.find(graphI);
+    if (srch == graphIMap.end())
+    {   // The Graph Instance has not been seen before, so not compiled.
+        return false;
+
+    } else {
+        builderGraphI = srch->second;
+    }
+
+    if(builderGraphI->idleInstructionBinary == "" || 
+        builderGraphI->idleDataBinary == "")
+    {   // We are missing a binary path, return false.
+        return false;
+    }
+
+    instrBin = builderGraphI->idleInstructionBinary;
+    dataBin = builderGraphI->idleDataBinary;
+    return true;
+}
+
 /******************************************************************************
  * Invoke a clean and then a degenerate
  *****************************************************************************/
@@ -1211,6 +1250,37 @@ int Composer::checkBinaries(ComposerGraphI_t* builderGraphI)
     std::string taskDir(builderGraphI->outputDir);
     std::string elfPath(taskDir + "/bin");
 
+
+    // Check that the "dummy" binaries for HW idle were generated.
+    FILE* dummyBinary;
+
+    // Check Dummy Instruction binary and add to GraphI
+    std::string dummyPath = elfPath + "/dummy_code.v";
+    dummyBinary = fopen(dummyPath.c_str(), "r");
+    if(dummyBinary == PNULL)
+    { // Failed to open binary
+        fprintf(fd,"\tFailed to open dummy instruction binary %s after compilation\n",
+                    dummyPath.c_str());
+        return -1;
+    }
+    fclose(dummyBinary);
+    builderGraphI->idleInstructionBinary = dummyPath;
+
+
+    // Check Dummy Data binary and add to GraphI
+    dummyPath = elfPath + "/threadCtxInit_data.v";
+    dummyBinary = fopen(dummyPath.c_str(), "r");
+    if(dummyBinary == PNULL)
+    { // Failed to open binary
+        fprintf(fd,"\tFailed to open dummy data binary %s after compilation\n",
+                    dummyPath.c_str());
+        return -1;
+    }
+    fclose(dummyBinary);
+    builderGraphI->idleDataBinary = dummyPath;
+
+
+
     // Check that the core binaries were made and link to each core.
     WALKSET(P_core*,(*(builderGraphI->cores)),coreNode)
     {

Source/OrchBase/Composer.h

@@ -98,6 +98,9 @@ typedef struct ComposerGraphI_t
     ssLoopMode_t softswitchLoopMode;
     bool softswitchRequestIdle;
 
+    std::string idleInstructionBinary;   // Dummy instruction binary for HW Idle
+    std::string idleDataBinary;          // Dummy data binary for HW Idle
+
     // Constructors/Destructors
     ComposerGraphI_t();
     ComposerGraphI_t(GraphI_t*, std::string&);
@@ -140,6 +143,8 @@ int         addFlags(GraphI_t*, std::string&);
 bool        isGenerated(GraphI_t*);
 bool        isCompiled(GraphI_t*);
 
+bool        getDummyPaths(GraphI_t*, std::string&, std::string&);   // Get the hardware idle binary names 
+
 void        Show(FILE * = stdout);
 void        Dump(unsigned = 0,FILE * = stdout);
 

Source/OrchBase/Handlers/CmDepl.cpp

@@ -376,11 +376,11 @@ int CmDepl::DeployGraph(GraphI_t* gi)
         bool soloApp = dynamic_cast<Root *>(par)->pOC->SingleApp();
         if (soloApp)
         {
-            /* TODO: GMB to plonk binary paths here (or at least some way to
-             * grab them). <!> */
-            #warning GMB to define these paths somewhere.
             std::string codePath;
             std::string dataPath;
+
+            // Get the dummy binary paths
+            par->pComposer->getDummyPaths(gi, codePath, dataPath);
 
             emptMessage.Put(0, &codePath);
             emptMessage.Put(1, &dataPath);

Source/Softswitch/Makefile

@@ -142,8 +142,8 @@ VDTARGETS := $(patsubst $(GENINC)/vars_%.h,$(BINDIR)/softswitch_data_%.v,$(wildc
 # targets for thread-local variables
 THREADVARS := $(patsubst $(GENSRC)/%.cpp,%.o,$(wildcard $(GENSRC)/vars_*_*.cpp))
 
-.PHONY: all clean printvars supervisor
-all : $(TARGETS) $(VCTARGETS) $(VDTARGETS) supervisor
+.PHONY: all clean printvars supervisor dummy
+all : $(TARGETS) $(VCTARGETS) $(VDTARGETS) supervisor dummy
 
 # This line can be given a target/prerequisite (tgt : $(TARGETS)) if needed
 #	$(foreach TARGET,$<,cp $(TARGET) $(BINDIR)/$(TARGET))
@@ -155,9 +155,11 @@ SupervisorData.o : supervisor.bin
 	ld -r -b binary -o $@ supervisor.bin
 	rm supervisor.bin
 
+# Supervisor shared object.
 $(BINDIR)/libSupervisor.so: Supervisor.cpp Supervisor.h supervisor_generated.cpp supervisor_generated.h SupervisorApi.h SupervisorApiEntrypoints.h SupervisorData.o
-	$(MPICXX) $(MPICFLAGS) $(MPILDFLAGS) -O3 -Wl,-soname,libSupervisor.so -o $@ $< SupervisorData.o
+	$(MPICXX) $(MPICFLAGS) -pipe $(MPILDFLAGS) -O3 -Wl,-soname,libSupervisor.so -o $@ $< SupervisorData.o
 
+# Split out the code and data from the elf.
 $(BINDIR)/softswitch_code_%.v: $(BINDIR)/softswitch_%.elf
 	$(TINSELBIN)/checkelf.sh $<
 	$(RV_OBJCOPY) -O verilog --only-section=.text $< $@
@@ -176,7 +178,7 @@ io.o : io.c io.h tinsel.h $(TINSELINC)/config.h
 
 ## the pattern substitution should find all vars_x_y.o created from vars_x_y.cpp
 .SECONDEXPANSION:
-$(BINDIR)/softswitch_%.elf : softswitch_main.o softswitch_common.o softswitch.o vars_%.o $$(patsubst $$(SECEXPCPPS),$$(SECEXPOS),$$(wildcard $(GENSRC)/vars_$$*_*.cpp)) handlers_%.o io.o link_%.ld entry.o libc.a libgcc.a
+$(BINDIR)/softswitch_%.elf : softswitch_main.o softswitch_common.o softswitch.o softswitch_vars.o vars_%.o $$(patsubst $$(SECEXPCPPS),$$(SECEXPOS),$$(wildcard $(GENSRC)/vars_$$*_*.cpp)) handlers_%.o io.o link_%.ld entry.o libc.a libgcc.a
 	$(RV_LD) $(LDFLAGS) -T link_$(notdir $*).ld -o $@ $(filter-out link_$(notdir $*).ld, $^)
 
 # this convoluted rule (we hope) will establish as a prerequisite for any file
@@ -205,5 +207,28 @@ link_%.ld : genld.sh
 $(HL)/%.o :
 	make -C $(TINSELHL)
 
+
+# Rules to build the dummy binaries for Hardware Idle detection.
+dummy: $(BINDIR)/threadCtxInit_data.v $(BINDIR)/dummy_code.v
+
+threadCtxInitGenerator : threadCtxInitGenerator.cpp
+	$(MPICXX) -I$(TINSELINC) -std=c++14 -o $@ $<
+
+dummy.o : dummy.cpp
+	$(RV_CPPC) $(CFLAGS) -Wall -c -o $@ $<
+
+$(BINDIR)/dummy.elf : dummy.o link_0.ld entry.o libc.a libgcc.a
+	$(RV_LD) $(LDFLAGS) -T link_0.ld -o $@ $(filter-out link_0.ld, $^)
+
+$(BINDIR)/dummy_code.v : $(BINDIR)/dummy.elf
+	$(TINSELBIN)/checkelf.sh $<
+	$(RV_OBJCOPY) -O verilog --only-section=.text $< $@
+
+$(BINDIR)/threadCtxInit_data.v : threadCtxInitGenerator
+	./$<
+	mv threadCtxInit_data.v $@
+
+
+# Cleanup on aisle 3
 clean :
-	rm -f *.o *.elf link.ld $(BINDIR)/*.elf $(BINDIR)/*.v $(BINDIR)/libSupervisor.so
+	rm -f *.o *.elf link.ld link_*.ld $(BINDIR)/*.elf $(BINDIR)/*.v $(BINDIR)/libSupervisor.so threadCtxInitGenerator

Source/Softswitch/inc/softswitch_vars.h

@@ -0,0 +1,23 @@
+#ifndef _VARS_H_
+#define _VARS_H_
+#include <cstdint>
+#include "softswitch_common.h"
+
+extern ThreadCtxt_t Thread_0_Context;
+extern ThreadCtxt_t Thread_1_Context;
+extern ThreadCtxt_t Thread_2_Context;
+extern ThreadCtxt_t Thread_3_Context;
+extern ThreadCtxt_t Thread_4_Context;
+extern ThreadCtxt_t Thread_5_Context;
+extern ThreadCtxt_t Thread_6_Context;
+extern ThreadCtxt_t Thread_7_Context;
+extern ThreadCtxt_t Thread_8_Context;
+extern ThreadCtxt_t Thread_9_Context;
+extern ThreadCtxt_t Thread_10_Context;
+extern ThreadCtxt_t Thread_11_Context;
+extern ThreadCtxt_t Thread_12_Context;
+extern ThreadCtxt_t Thread_13_Context;
+extern ThreadCtxt_t Thread_14_Context;
+extern ThreadCtxt_t Thread_15_Context;
+
+#endif /*_VARS_H_*/

Source/Softswitch/src/dummy.cpp

@@ -0,0 +1,9 @@
+#include "tinsel.h"
+
+int main(int argc, char** argv)
+{
+    // Block on tinselIdle to facilitate hardware idle.
+    while(true) tinselIdle(1);
+
+    return 0;
+}

Source/Softswitch/src/softswitch_main.cpp

@@ -14,6 +14,11 @@ void softswitch_main()
 {
     // The thread context will reside at the base of the DRAM memory area.
     PThreadContext* ThreadContext = static_cast<PThreadContext*>(tinselHeapBase());
+
+    if(ThreadContext->numDevInsts == 0)
+    {   // We have no devices loaded, so should do nothing.
+        while(true) tinselIdle(1);
+    }
 
     // Configure rtsBuf
     outPin_t* rtsBuf[ThreadContext->rtsBuffSize+1];

Source/Softswitch/src/softswitch_vars.cpp

@@ -0,0 +1,20 @@
+
+#include "softswitch_common.h"
+#include "softswitch_vars.h"
+
+ThreadCtxt_t Thread_0_Context __attribute__((weak, section (".thr0_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_1_Context __attribute__((weak, section (".thr1_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_2_Context __attribute__((weak, section (".thr2_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_3_Context __attribute__((weak, section (".thr3_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_4_Context __attribute__((weak, section (".thr4_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_5_Context __attribute__((weak, section (".thr5_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_6_Context __attribute__((weak, section (".thr6_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_7_Context __attribute__((weak, section (".thr7_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_8_Context __attribute__((weak, section (".thr8_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_9_Context __attribute__((weak, section (".thr9_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_10_Context __attribute__((weak, section (".thr10_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_11_Context __attribute__((weak, section (".thr11_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_12_Context __attribute__((weak, section (".thr12_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_13_Context __attribute__((weak, section (".thr13_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_14_Context __attribute__((weak, section (".thr14_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
+ThreadCtxt_t Thread_15_Context __attribute__((weak, section (".thr15_base"))) = {0,PNULL,0,PNULL,PNULL,0,PNULL,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};

Source/Softswitch/src/threadCtxInitGenerator.cpp

@@ -0,0 +1,70 @@
+#include "config.h"
+
+#include <string>
+#include <iostream>
+#include <iomanip>
+#include <fstream>
+#include <stdio.h>
+#include <math.h>
+
+
+// Number of bytes per thread partition                         2^21 = 2,097,152
+uint32_t BytesPerDRAMPartition=1<<TinselLogBytesPerDRAMPartition;
+
+// The number of cores that share a DRAM                            2^(2+3) = 32
+uint32_t CoresPerDRAM=1<<(TinselLogCoresPerDCache+TinselLogDCachesPerDRAM);
+
+// The number of threads per DRAM                                    32*16 = 512
+uint32_t ThreadsPerDRAM=CoresPerDRAM*TinselThreadsPerCore;
+
+uint32_t InterleavedPartitionOffset=0x80000000;
+
+int main(void)
+{
+    // The output binary
+    std::ofstream dataBin;
+
+    dataBin.open("threadCtxInit_data.v");
+    if(dataBin.fail()) // Check that the file opened
+    {                 // if it didn't, barf
+        std::cout << "***ERROR*** FAILED TO OPEN DATA BINARY\n" << std::endl;
+        return -1;
+    }
+
+    // Populate the bin
+    uint32_t i = 0;
+    for(int coreNum = 0; coreNum<CoresPerDRAM; coreNum++)
+    {
+        for(int threadNum = 1; threadNum<=TinselThreadsPerCore; threadNum++)
+        {
+            uint32_t threadAddr, threadOffset;
+
+            threadOffset = ((coreNum << TinselLogThreadsPerCore) + threadNum);
+            threadOffset *= BytesPerDRAMPartition;
+
+            threadAddr = TinselBytesPerDRAM - threadOffset;
+
+            threadAddr |= InterleavedPartitionOffset;
+
+            // Write the address
+            dataBin << std::hex << std::uppercase << "@" << threadAddr << "\n";
+
+            // Write out the initialiser
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n";  
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n";  
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n"; 
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n"; 
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n"; 
+            dataBin << "00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 \n"; 
+            dataBin << "00 00 00 00 00 00 00 00 " << std::endl;
+
+            i++;
+        }
+    }
+
+    // And we are done
+    dataBin.close();
+
+    return 0;
+}
+