Merge a477f17 into e5b8196

examples/examplescommon.lua

@@ -374,6 +374,28 @@ C.removeMSBs = memoize(function(A,bits)
   return mod
 end)
 
+-----------------------------
+C.removeLSBs = memoize(function(A,bits)
+  err( types.isType(A),"C.removeLSBs: type should be rigel type")
+  err( type(bits)=="number","C.removeLSBs: bits should be number or value")
+
+  local otype
+  if A:isUint() then
+    J.err(A.precision>bits,"removeLSBs: can't remove all the bits! attempting to remove "..bits.." bits from "..tostring(A))
+    otype = types.uint(A.precision-bits)
+  elseif A:isInt() then
+    J.err(A.precision>bits,"removeLSBs: can't remove all the bits! attempting to remove "..bits.." bits from "..tostring(A))
+    otype = types.int(A.precision-bits)
+    assert(false) -- NYI
+  else
+    J.err( A:isUint() or A:isInt(), "generators.removeLSBs: type should be int or uint, but is: "..tostring(A) )
+  end
+
+  local mod = RM.lift(J.sanitize("generators_removeLSBs_"..tostring(bits).."_"..tostring(A)), A,nil,nil,
+                      function(inp) return S.cast(S.rshift(inp,S.constant(bits,A)),otype) end)
+  return mod
+end)
+
 -----------------------------
 C.select = memoize(function(ty)
   err(types.isType(ty), "C.select error: input must be type")
@@ -1003,7 +1025,7 @@ C.fifo = memoize(function(ty,size,nostall,csimOnly,X)
   local regs = {R.instantiateRegistered("f1",RM.fifo(ty,size,nostall,nil,nil,nil,csimOnly))}
   local st = R.applyMethod("s1",regs[1],"store",inp)
   local ld = R.applyMethod("l1",regs[1],"load")
-  return RM.lambda("C_FIFO_"..tostring(ty).."_size"..tostring(size), inp, R.statements{ld,st}, regs, "C.fifo", {size=size} )
+  return RM.lambda("C_FIFO_"..tostring(ty).."_size"..tostring(size).."_nostall"..tostring(nostall), inp, R.statements{ld,st}, regs, "C.fifo", {size=size} )
 end)
 
 -------------
@@ -1078,6 +1100,7 @@ C.downsampleSeq = memoize(function( A, W, H, T, scaleX, scaleY, framed, X )
   err( type(W)=="number", "C.downsampleSeq: W must be number")
   err( type(H)=="number", "C.downsampleSeq: H must be number")
   err( type(T)=="number", "C.downsampleSeq: T must be number")
+  err( T>0, "C.downsampleSeq: T must be >0")
   err( type(scaleX)=="number", "C.downsampleSeq: scaleX must be number")
   err( type(scaleY)=="number", "C.downsampleSeq: scaleY must be number")
   err( scaleX>=1, "C.downsampleSeq: scaleX must be >=1")
@@ -1818,4 +1841,27 @@ C.AXIReadPar = memoize(
     return res
   end)
 
+-- collect the number of tokens seen, and write it to a global
+C.tokenCounterReg = memoize(
+  function( A, regGlobal, N)
+    J.err( types.isType(A),"tokenCounterReg: a must be type" )
+    J.err( A:is("Handshake"),"tokenCounterReg: input should be handshaked, but is: "..tostring(A))
+    J.err( N==nil or type(N)=="number","tokenCounterReg: N must be number" )
+
+    if N==nil then
+      N = math.pow(2,32)-1
+    end
+
+    local G = require "generators"
+
+    return G.Module{A,
+      function(inp)
+        local inpb = G.FanOut{2}(inp)
+        local ct = G.HS{C.valueToTrigger(A.params.A)}(inpb[1])
+        local cnt = G.HS{RM.counter(types.uint(32),N)}(ct)
+        cnt = G.HS{G.Add{1}}(cnt)
+        return R.statements{inpb[0],R.writeGlobal("wg",regGlobal,cnt)}
+      end}
+  end)
+
 return C

examples/gold/soc_tokencounter.regout.lua

@@ -0,0 +1 @@
+return {startCnt=16384,endCnt=4096}

examples/gold/soc_tokencounter.terra.cycles.txt

@@ -0,0 +1 @@
+8074

examples/makefile

@@ -1,7 +1,7 @@
 BUILDDIR ?= out
 
 # soc_flipWrite.lua 
-SRCS_SOC = soc_simple.lua soc_2in.lua soc_convgen.lua soc_convgenTaps.lua soc_flip.lua soc_15x15.lua soc_15x15x15.lua soc_flipWrite.lua soc_regin.lua soc_regout.lua soc_convtest.lua soc_read.lua soc_redu1024.lua soc_redu2048.lua soc_redu4096.lua soc_redu8192.lua soc_redu16384.lua soc_redu32768.lua soc_sort.lua soc_filterseq.lua soc_filterseq8.lua soc_unaligned.lua soc_underflow.lua soc_parread.lua
+SRCS_SOC = soc_simple.lua soc_2in.lua soc_convgen.lua soc_convgenTaps.lua soc_flip.lua soc_15x15.lua soc_15x15x15.lua soc_flipWrite.lua soc_regin.lua soc_regout.lua soc_convtest.lua soc_read.lua soc_redu1024.lua soc_redu2048.lua soc_redu4096.lua soc_redu8192.lua soc_redu16384.lua soc_redu32768.lua soc_sort.lua soc_filterseq.lua soc_filterseq8.lua soc_unaligned.lua soc_underflow.lua soc_parread.lua soc_tokencounter.lua
 
 VERILATOR_SOC = $(patsubst %.lua,$(BUILDDIR)/%.verilatorSOC.bit,$(SRCS_SOC))
 VERILATOR_SOC += $(patsubst %.lua,$(BUILDDIR)/%.verilatorSOC.raw,$(SRCS_SOC))
@@ -11,6 +11,12 @@ VERILATOR_SOC += $(patsubst %.lua,$(BUILDDIR)/%.verilatorSOC.regcorrect.txt,$(SR
 
 ZU9VIVADOSOCBITS = $(patsubst %.lua,$(BUILDDIR)/%.zu9vivadoSOC.bit,$(SRCS_SOC))
 
+ZU9VIVADOSOC = $(ZU9VIVADOSOCBITS)
+ZU9VIVADOSOC += $(patsubst %.lua,$(BUILDDIR)/%.zu9vivadoSOC.raw,$(SRCS_SOC))
+ZU9VIVADOSOC += $(patsubst %.lua,$(BUILDDIR)/%.zu9vivadoSOC.bmp,$(SRCS_SOC))
+ZU9VIVADOSOC += $(patsubst %.lua,$(BUILDDIR)/%.zu9vivadoSOC.correct.txt,$(SRCS_SOC))
+ZU9VIVADOSOC += $(patsubst %.lua,$(BUILDDIR)/%.zu9vivadoSOC.regcorrect.txt,$(SRCS_SOC))
+
 SRCS = $(wildcard *.lua)
 SRCS := $(filter-out harness.lua examplescommon.lua pyramid_core.lua harris_core.lua sift_core.lua sift_core_hw.lua campipe_core.lua descriptor_core.lua  stereo_core.lua stereo_tr_core.lua  lk_core.lua lk_tr_core.lua $(SRCS_SOC) harnessSOC.lua,$(SRCS))
 METADATA = $(patsubst %.lua,$(BUILDDIR)/%.metadata.lua,$(SRCS))
@@ -261,6 +267,8 @@ zu9vivadobits: $(ZU9VIVADOBITS)
 
 zu9vivadoSOCbits: $(ZU9VIVADOSOCBITS)
 
+zu9vivadoSOC: $(ZU9VIVADOSOC)
+
 axiverilog: $(AXIVERILOG)
 	touch $(BUILDDIR)/axiverilog_done.txt
 

examples/soc_tokencounter.lua

@@ -0,0 +1,24 @@
+local R = require "rigel"
+local SOC = require "soc"
+local C = require "examplescommon"
+local harness = require "harnessSOC"
+local G = require "generators"
+local RS = require "rigelSimple"
+local types = require "types"
+local SDF = require "sdf"
+types.export()
+
+local Regs = SOC.axiRegs{startCnt={u32,0,"out"},endCnt={u32,0,"out"}}
+regs = Regs:instantiate(SDF{1,8192})
+
+local RegInOut = G.Module{
+  function(i)
+    local o = G.AXIReadBurstSeq{"frame_128.raw",{128,64},u(8),1}(i)
+    o = C.tokenCounterReg(o.type,Regs.startCnt,128*64*2048)(o)
+    o = G.HS{C.downsampleSeq( u8, 128, 64, 1, 2, 2)}(o)
+    o = C.tokenCounterReg(o.type,Regs.endCnt,128*64*2048)(o)
+    print("OT",o.type)
+    return G.AXIWriteBurstSeq{"out/soc_tokencounter",{64,32},1}(o)
+  end}
+
+harness{regs.start,RegInOut,regs.done}

platform/zu9vivadoSOC/processimage_mpsoc_SOC.c

@@ -182,19 +182,23 @@ int setRegisters(int curArg, char *argv[], void* ptr, unsigned int GPIO_BASE){
   fflush(stdout);
 
   curArg++; // for "--registers"
-  while(strcmp(argv[curArg],"--outputs")!=0){
+  while(strcmp(argv[curArg],"--registersOut")!=0){
     printf("PARSE Reg %s %s\n",argv[curArg],argv[curArg+1]);
     fflush(stdout);
 
     unsigned int addr = strtol(argv[curArg],NULL,16);
+    unsigned int dat = strtol(argv[curArg+1],NULL,16);
     int bytes = strlen(argv[curArg+1])/2;
-    printf("Set Register %x withNBytes %d, GPIO_BASE %x\n",addr,bytes,GPIO_BASE);
+    printf("Set Register %x withNBytes %d to value 0x%s/%d, GPIO_BASE %x\n",addr,bytes,argv[curArg+1],dat,GPIO_BASE);
     fflush(stdout);
 
     //unsigned int* data = (unsigned int*)argv[curArg+1];
     char tmp[9]="";
 
-    for(int i=0; i<bytes/4; i++){
+    unsigned int num32bitchunks = bytes/4;
+    if(num32bitchunks<=0){num32bitchunks=1;}
+
+    for(int i=0; i<num32bitchunks; i++){
       for(int j=0; j<8; j++){
         tmp[j] = argv[curArg+1][i*8+j];
       }
@@ -210,6 +214,9 @@ int setRegisters(int curArg, char *argv[], void* ptr, unsigned int GPIO_BASE){
       fflush(stdout);
 
       *(unsigned int*)(ptr+offset) = data;
+
+      unsigned int regDat = *(unsigned int*)(ptr+offset);
+      printf("Readback Reg %d\n",regDat);
       //setReg( top, verbose, addr+i*4, data);
     }
 
@@ -219,6 +226,34 @@ int setRegisters(int curArg, char *argv[], void* ptr, unsigned int GPIO_BASE){
   return curArg;
 }
 
+int getRegisters(int curArg, char *argv[], void* ptr, unsigned int GPIO_BASE ){
+  curArg++; // for '--registersOut'
+
+  FILE* regFile = fopen("regout.lua","w");
+  fprintf(regFile,"return {");
+
+  bool first = true;
+  while(strcmp(argv[curArg],"--outputs")!=0){
+    unsigned int addr = strtol(argv[curArg+1],NULL,16);
+
+
+    unsigned int offsetBytes = addr-GPIO_BASE;
+    unsigned int regOut = *(unsigned int*)(ptr+offsetBytes);
+
+    printf("Read Reg %s (addr:%x) value:%d\n",argv[curArg],addr,regOut);
+
+    if(!first){fprintf(regFile,",");}
+    first=false;
+    fprintf(regFile,"%s=%d",argv[curArg],regOut);
+    curArg+=2;
+  }
+
+  fprintf(regFile,"}");
+  fclose(regFile);
+
+  return curArg;
+}
+
 int writeOutputs(int curArg, int argc, char *argv[], void* ptr){
   printf("Write Outputs\n");
   fflush(stdout);
@@ -380,7 +415,8 @@ int main(int argc, char *argv[]) {
       printf("Done after %f seconds, %f cycles (@ %f MHZ)\n", len, (MHZ*HZ)*len,MHZ);
     }
   }
-
+
+  curArg = getRegisters( curArg, argv, gpioptr, gpio_addr );
   writeOutputs(curArg,argc,argv,ptr);
 
   printf("conf: start: %d, done: %d\n", conf->start, conf->done);

platform/zu9vivadoSOC/run

@@ -10,7 +10,7 @@ BITFILE=$1
 BITFILE_BASE=$(basename $1)
 METAFILE=$2
 OUTFILE=$3
-#OUTPATH=$4
+OUTPATH=$4
 
 while [[ -z `ip -6 neigh show | grep $ZYNQ_ETH | cut -d " " -f1` ]]
 do
@@ -32,6 +32,7 @@ MEMEND=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE memoryEnd)
 INPUTS=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __INPUTS_ZYNQ)
 OUTPUTS=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __OUTPUTS_ZYNQ)
 REGISTERS=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __REGISTERS)
+REGISTERSOUT=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __REGISTERSOUT)
 
 INPUT_FILES_ON_HOST=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __INPUT_HOST_FILES)
 INPUT_FILES_ON_DEVICE=$($LUA $DIR/../../misc/extractMetadata.lua $METAFILE __INPUT_DEVICE_FILES)
@@ -52,9 +53,10 @@ SSHOPT="-o UserKnownHostsFile=/dev/null -o StrictHostKeyChecking=no -6"
 sshpass -p 'root' scp $SSHOPT $DIR/processimage_mpsoc_SOC $BITFILE $INPUT_FILES_ON_HOST root@\[$ZYNQ_ADDR\]:$ZYNQ_WRITE_PATH
 sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "mv $ZYNQ_WRITE_PATH/$BITFILE_BASE $ZYNQ_BIN_FILE_PATH"
 sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "echo $BITFILE_BASE > $ZYNQ_FIRMWARE_PATH"
-sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "$ZYNQ_WRITE_PATH/processimage_mpsoc_SOC 0xA0000000 $MEMSTART $MEMEND --inputs $INPUTS --registers $REGISTERS --outputs $OUTPUTS"
+sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "$ZYNQ_WRITE_PATH/processimage_mpsoc_SOC 0xA0000000 $MEMSTART $MEMEND --inputs $INPUTS --registers $REGISTERS --registersOut $REGISTERSOUT --outputs $OUTPUTS"
 sshpass -p 'root' scp $SSHOPT root@\[$ZYNQ_ADDR\]:$ZYNQ_WRITE_PATH/out.raw $OUTFILE
-sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "rm $ZYNQ_WRITE_PATH/processimage_mpsoc_SOC $INPUT_FILES_ON_DEVICE $ZYNQ_WRITE_PATH/out.raw $ZYNQ_BIN_FILE_PATH/$BITFILE_BASE"
+sshpass -p 'root' scp $SSHOPT root@\[$ZYNQ_ADDR\]:$ZYNQ_WRITE_PATH/regout.lua $(echo $OUTPATH)zu9vivadoSOC.regout.lua
+sshpass -p 'root' ssh $SSHOPT root@$ZYNQ_ADDR "rm $ZYNQ_WRITE_PATH/processimage_mpsoc_SOC $INPUT_FILES_ON_DEVICE $ZYNQ_WRITE_PATH/out.raw $ZYNQ_WRITE_PATH/regout.lua $ZYNQ_BIN_FILE_PATH/$BITFILE_BASE"
 
 #rm -f /tmp/zynq10lock
 #	$(TERRA) ../misc/extractCycles.t out/$*.axi.raw > out/$*.axi.cycles.txt

src/fpgamodules.lua

@@ -175,20 +175,21 @@ return sliceRamCache[bits]
   return ram
 end
 
-modules.fifo = memoize(function(ty,items,verbose)
+modules.fifo = memoize(function(ty,items,verbose,nostall)
   err(types.isType(ty),"fifo type must be type")
   err(type(items)=="number","fifo items must be number")
   err(type(verbose)=="boolean","fifo verbose must be bool")
   err( items>1,"fifo: items must be >1")
 
-  local fifo = Ssugar.moduleConstructor( J.sanitize("fifo_"..tostring(ty).."_"..items) )
+  local fifo = Ssugar.moduleConstructor( J.sanitize("fifo_"..tostring(ty).."_"..items.."_nostall"..tostring(nostall)) )
   -- writeAddr, readAddr hold the address we will read/write from NEXT time we do a read/write
   local addrBits = (math.ceil(math.log(items)/math.log(2)))+1 -- the +1 is so that we can disambiguate wraparoudn
   local writeAddr = fifo:add( systolic.module.regBy( types.uint(addrBits), modules.incIfWrap(types.uint(addrBits),items-1), true,nil,0 ):instantiate("writeAddr"))
   local readAddr = fifo:add( systolic.module.regBy( types.uint(addrBits), modules.incIfWrap(types.uint(addrBits),items-1), true,nil,0 ):instantiate("readAddr"))
   local bits = ty:verilogBits()
   local bytes = bits/8
-
+  bytes = math.ceil(bytes) -- bits may not be byte aligned
+
   local ram, rams
   if items <= 128 and J.isPowerOf2(bits) then
     ram = fifo:add( sliceRamGen(bits):instantiate("ram") )
@@ -225,7 +226,7 @@ modules.fifo = memoize(function(ty,items,verbose)
   local pushCE = S.CE("CE_push")
   local pushBack = fifo:addFunction( Ssugar.lambdaConstructor("pushBack",ty,"pushBack_input" ) )
   pushBack:setCE(pushCE)
-  local pushBackAssert = fifo:add( systolic.module.assert( "attempting to push to a full fifo", true ):instantiate("pushBackAssert") )
+  local pushBackAssert = fifo:add( systolic.module.assert( "attempting to push to a full fifo", true, nostall~=true ):instantiate("pushBackAssert") )
   local hasSpace = S.lt( fsize, S.constant(items-2,types.uint(addrBits)) ):disablePipelining() -- note that this is different than ready
   pushBack:addPipeline( pushBackAssert:process( hasSpace )  )
   pushBack:addPipeline( writeAddr:setBy(systolic.constant(true, types.bool()) ) )
@@ -237,7 +238,9 @@ modules.fifo = memoize(function(ty,items,verbose)
       pushBack:addPipeline( rams[b]:write( S.tuple{ S.cast(writeAddr:get(),types.uint(7)), S.bitSlice(pushBack:getInput(),b-1,b-1)} )  )
     end
   else
-    pushBack:addPipeline( ram:writeAndReturnOriginal( S.tuple{ S.cast(writeAddr:get(),types.uint(addrBits-1)), S.cast(pushBack:getInput(),types.bits(bits))} )  )
+    local tinp = S.cast(pushBack:getInput(),types.bits(bits))
+    tinp = S.cast(tinp,types.bits(bytes*8))
+    pushBack:addPipeline( ram:writeAndReturnOriginal( S.tuple{ S.cast(writeAddr:get(),types.uint(addrBits-1)), tinp} )  )
   end
 
   -- pushBackReset
@@ -267,7 +270,10 @@ modules.fifo = memoize(function(ty,items,verbose)
     popOutput = systolic.cast( systolic.tuple(bitfield), ty)
     popFront:setOutput( popOutput, "popFront" )
   else
-    popOutput = S.cast(ram:read(S.cast(readAddr:get(),types.uint(addrBits-1))),ty)
+    popOutput = S.cast(ram:read(S.cast(readAddr:get(),types.uint(addrBits-1))),types.uint(bytes*8))
+    popOutput = S.cast(popOutput,types.uint(bits))
+    popOutput = S.cast(popOutput,types.bits(bits))
+    popOutput = S.cast(popOutput,ty)
     popFront:setOutput( popOutput, "popFront" )
   end
 

src/generators.lua

@@ -49,6 +49,7 @@ generators.PartitionBits = R.newGenerator("generators","PartitionBits",{"type","
 generators.Rshift = R.newGenerator("generators","Rshift",{"type","rate"},{"number"}, function(args) return C.rshift(args.type,args.number) end)
 generators.AddMSBs = R.newGenerator("generators","AddMSBs",{"type","rate"},{"number"}, function(args) return C.addMSBs(args.type,args.number) end)
 generators.RemoveMSBs = R.newGenerator("generators","RemoveMSBs",{"type","rate"},{"number"}, function(args) return C.removeMSBs(args.type,args.number) end)
+generators.RemoveLSBs = R.newGenerator("generators","RemoveLSBs",{"type","rate"},{"number"}, function(args) return C.removeLSBs(args.type,args.number) end)
 generators.Index = R.newGenerator("generators","Index",{"type","rate"},{"number","size"},
                                   function(args)
                                     if args.size~=nil then
@@ -85,11 +86,11 @@ function(args)
   return RM.packTuple(args.type.params.list,args.bool)
 end)
 
-generators.FIFO = R.newGenerator("generators","FIFO",{"type","number","rate"},{},
+generators.FIFO = R.newGenerator("generators","FIFO",{"type","number","rate"},{"bool"},
 function(args)
   J.err( R.isHandshake(args.type) or R.isHandshakeTrigger(args.type),"FIFO: expected handshake input, but is: "..tostring(args.type))
   local ty = R.extractData(args.type)
-  return C.fifo( ty, args.number )
+  return C.fifo( ty, args.number, args.bool )
 end)
 
 generators.Add = R.newGenerator("generators","Add",{"type","rate"},{"bool","number"},
@@ -271,7 +272,7 @@ function(args)
     err(false,"generators.Downsample: unsupported input type: "..tostring(args.type))
   end
 end)
-
+                                       
 generators.PosSeq = R.newGenerator("generators","PosSeq",{"size","number"},{"type"},
 function(args)
   if args.type~=nil then