Improvement to compute graph

New cyclo static scheduler
New code generation mode using switch / case to decrease code size.

ComputeGraph/examples/References.md

@@ -0,0 +1,36 @@
+# Reference statistics
+
+The different examples should return following schedule statistics:
+
+
+## Example 1
+    Schedule length = 17
+    Memory usage 64 bytes
+
+## Example 2
+    Schedule length = 302
+    Memory usage 10720 bytes
+
+## Example 3 
+    Schedule length = 25
+    Memory usage 11264 bytes
+
+## Example 4
+    Schedule length = 25
+    Memory usage 11264 bytes
+
+## Example 5 
+    Schedule length = 292
+    Memory usage 6614 bytes
+
+## Example 6 
+    Schedule length = 17
+    Memory usage 2204 bytes
+
+## Example 7 
+    Schedule length = 3
+    Memory usage 512 bytes
+
+## Example 8 
+    Schedule length = 37
+    Memory usage 288 bytes

ComputeGraph/examples/example1/generated/scheduler.cpp

@@ -14,6 +14,17 @@ The support classes and code is covered by CMSIS-DSP license.
 #include "AppNodes.h"
 #include "scheduler.h"
 
+/*
+
+Description of the scheduling. It is a list of nodes to call.
+The values are indexes in the previous array.
+
+*/
+static unsigned int schedule[17]=
+{ 
+2,2,0,1,2,0,1,2,2,0,1,2,0,1,2,0,1,
+};
+
 /***********
 
 FIFO buffers
@@ -51,45 +62,40 @@ uint32_t scheduler(int *error,int someVariable)
     /* Run several schedule iterations */
     while((cgStaticError==0) && (debugCounter > 0))
     {
-       /* Run a schedule iteration */
+        /* Run a schedule iteration */
-       cgStaticError = source.run();
+        for(unsigned long id=0 ; id < 17; id++)
-       CHECKERROR;
+        {
-       cgStaticError = source.run();
+            switch(schedule[id])
-       CHECKERROR;
+            {
-       cgStaticError = filter.run();
+                case 0:
-       CHECKERROR;
+                {
-       cgStaticError = sink.run();
+                   cgStaticError = filter.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = source.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = filter.run();
+
-       CHECKERROR;
+                case 1:
-       cgStaticError = sink.run();
+                {
-       CHECKERROR;
+                   cgStaticError = sink.run();
-       cgStaticError = source.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = source.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = filter.run();
+                case 2:
-       CHECKERROR;
+                {
-       cgStaticError = sink.run();
+                   cgStaticError = source.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = source.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = filter.run();
+
-       CHECKERROR;
+                default:
-       cgStaticError = sink.run();
+                break;
-       CHECKERROR;
+            }
-       cgStaticError = source.run();
+        }
-       CHECKERROR;
-       cgStaticError = filter.run();
-       CHECKERROR;
-       cgStaticError = sink.run();
-       CHECKERROR;
-
        debugCounter--;
        nbSchedule++;
     }
+
     *error=cgStaticError;
     return(nbSchedule);
 }

ComputeGraph/examples/example3/generated/scheduler.cpp

@@ -14,6 +14,17 @@ The support classes and code is covered by CMSIS-DSP license.
 #include "AppNodes.h"
 #include "scheduler.h"
 
+/*
+
+Description of the scheduling. It is a list of nodes to call.
+The values are indexes in the previous array.
+
+*/
+static unsigned int schedule[25]=
+{ 
+6,2,0,7,3,4,8,1,6,2,0,7,3,4,8,2,0,7,3,4,1,5,8,1,5,
+};
+
 /***********
 
 FIFO buffers
@@ -86,12 +97,14 @@ uint32_t scheduler(int *error)
     /* Run several schedule iterations */
     while((cgStaticError==0) && (debugCounter > 0))
     {
-       /* Run a schedule iteration */
+        /* Run a schedule iteration */
-       cgStaticError = src.run();
+        for(unsigned long id=0 ; id < 25; id++)
-       CHECKERROR;
+        {
-       cgStaticError = audioWin.run();
+            switch(schedule[id])
-       CHECKERROR;
+            {
-       {
+                case 0:
+                {
+                   {
          float32_t* i0;
          float32_t* o2;
          i0=fifo1.getReadBuffer(256);
@@ -99,69 +112,74 @@ uint32_t scheduler(int *error)
          arm_mult_f32(i0,HANN,o2,256);
          cgStaticError = 0;
        }
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = toCmplx.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = cfft.run();
+
-       CHECKERROR;
+                case 1:
-       cgStaticError = icfft.run();
+                {
-       CHECKERROR;
+                   cgStaticError = audioOverlap.run();
-       cgStaticError = toReal.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = audioOverlap.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = src.run();
+                case 2:
-       CHECKERROR;
+                {
-       cgStaticError = audioWin.run();
+                   cgStaticError = audioWin.run();
-       CHECKERROR;
+                   CHECKERROR;
-       {
+                }
-         float32_t* i0;
+                break;
-         float32_t* o2;
+
-         i0=fifo1.getReadBuffer(256);
+                case 3:
-         o2=fifo2.getWriteBuffer(256);
+                {
-         arm_mult_f32(i0,HANN,o2,256);
+                   cgStaticError = cfft.run();
-         cgStaticError = 0;
+                   CHECKERROR;
-       }
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = toCmplx.run();
+
-       CHECKERROR;
+                case 4:
-       cgStaticError = cfft.run();
+                {
-       CHECKERROR;
+                   cgStaticError = icfft.run();
-       cgStaticError = icfft.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = toReal.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = audioWin.run();
+                case 5:
-       CHECKERROR;
+                {
-       {
+                   cgStaticError = sink.run();
-         float32_t* i0;
+                   CHECKERROR;
-         float32_t* o2;
+                }
-         i0=fifo1.getReadBuffer(256);
+                break;
-         o2=fifo2.getWriteBuffer(256);
+
-         arm_mult_f32(i0,HANN,o2,256);
+                case 6:
-         cgStaticError = 0;
+                {
-       }
+                   cgStaticError = src.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = toCmplx.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = cfft.run();
+
-       CHECKERROR;
+                case 7:
-       cgStaticError = icfft.run();
+                {
-       CHECKERROR;
+                   cgStaticError = toCmplx.run();
-       cgStaticError = audioOverlap.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = sink.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = toReal.run();
+                case 8:
-       CHECKERROR;
+                {
-       cgStaticError = audioOverlap.run();
+                   cgStaticError = toReal.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = sink.run();
+                }
-       CHECKERROR;
+                break;
-
+
+                default:
+                break;
+            }
+        }
        debugCounter--;
        nbSchedule++;
     }
+
     *error=cgStaticError;
     return(nbSchedule);
 }

ComputeGraph/examples/example6/generated/scheduler.cpp

@@ -14,6 +14,17 @@ The support classes and code is covered by CMSIS-DSP license.
 #include "AppNodes.h"
 #include "scheduler.h"
 
+/*
+
+Description of the scheduling. It is a list of nodes to call.
+The values are indexes in the previous array.
+
+*/
+static unsigned int schedule[17]=
+{ 
+4,0,1,2,3,3,4,0,1,2,3,3,0,1,2,3,3,
+};
+
 /***********
 
 FIFO buffers
@@ -63,45 +74,54 @@ uint32_t scheduler(int *error,arm_mfcc_instance_f32 *mfccConfig)
     /* Run several schedule iterations */
     while((cgStaticError==0) && (debugCounter > 0))
     {
-       /* Run a schedule iteration */
+        /* Run a schedule iteration */
-       cgStaticError = src.run();
+        for(unsigned long id=0 ; id < 17; id++)
-       CHECKERROR;
+        {
-       cgStaticError = audioWin.run();
+            switch(schedule[id])
-       CHECKERROR;
+            {
-       cgStaticError = mfcc.run();
+                case 0:
-       CHECKERROR;
+                {
-       cgStaticError = mfccWin.run();
+                   cgStaticError = audioWin.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = sink.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = sink.run();
+
-       CHECKERROR;
+                case 1:
-       cgStaticError = src.run();
+                {
-       CHECKERROR;
+                   cgStaticError = mfcc.run();
-       cgStaticError = audioWin.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = mfcc.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = mfccWin.run();
+                case 2:
-       CHECKERROR;
+                {
-       cgStaticError = sink.run();
+                   cgStaticError = mfccWin.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = sink.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = audioWin.run();
+
-       CHECKERROR;
+                case 3:
-       cgStaticError = mfcc.run();
+                {
-       CHECKERROR;
+                   cgStaticError = sink.run();
-       cgStaticError = mfccWin.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = sink.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = sink.run();
+                case 4:
-       CHECKERROR;
+                {
-
+                   cgStaticError = src.run();
+                   CHECKERROR;
+                }
+                break;
+
+                default:
+                break;
+            }
+        }
        debugCounter--;
        nbSchedule++;
     }
+
     *error=cgStaticError;
     return(nbSchedule);
 }

ComputeGraph/examples/example8/generated/scheduler.cpp

@@ -14,6 +14,17 @@ The support classes and code is covered by CMSIS-DSP license.
 #include "AppNodes.h"
 #include "scheduler.h"
 
+/*
+
+Description of the scheduling. It is a list of nodes to call.
+The values are indexes in the previous array.
+
+*/
+static unsigned int schedule[37]=
+{ 
+6,6,1,5,0,4,3,2,6,1,6,5,0,4,3,2,6,1,5,0,4,3,2,6,1,6,5,0,4,3,2,1,5,0,4,3,2,
+};
+
 /***********
 
 FIFO buffers
@@ -75,85 +86,68 @@ uint32_t scheduler(int *error,int someVariable)
     /* Run several schedule iterations */
     while((cgStaticError==0) && (debugCounter > 0))
     {
-       /* Run a schedule iteration */
+        /* Run a schedule iteration */
-       cgStaticError = source.run();
+        for(unsigned long id=0 ; id < 37; id++)
-       CHECKERROR;
+        {
-       cgStaticError = source.run();
+            switch(schedule[id])
-       CHECKERROR;
+            {
-       cgStaticError = filter.run();
+                case 0:
-       CHECKERROR;
+                {
-       cgStaticError = sd.run();
+                   cgStaticError = dup0.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = dup0.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = sc.run();
+
-       CHECKERROR;
+                case 1:
-       cgStaticError = sb.run();
+                {
-       CHECKERROR;
+                   cgStaticError = filter.run();
-       cgStaticError = sa.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = source.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = filter.run();
+                case 2:
-       CHECKERROR;
+                {
-       cgStaticError = source.run();
+                   cgStaticError = sa.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = sd.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = dup0.run();
+
-       CHECKERROR;
+                case 3:
-       cgStaticError = sc.run();
+                {
-       CHECKERROR;
+                   cgStaticError = sb.run();
-       cgStaticError = sb.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = sa.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = source.run();
+                case 4:
-       CHECKERROR;
+                {
-       cgStaticError = filter.run();
+                   cgStaticError = sc.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = sd.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = dup0.run();
+
-       CHECKERROR;
+                case 5:
-       cgStaticError = sc.run();
+                {
-       CHECKERROR;
+                   cgStaticError = sd.run();
-       cgStaticError = sb.run();
+                   CHECKERROR;
-       CHECKERROR;
+                }
-       cgStaticError = sa.run();
+                break;
-       CHECKERROR;
+
-       cgStaticError = source.run();
+                case 6:
-       CHECKERROR;
+                {
-       cgStaticError = filter.run();
+                   cgStaticError = source.run();
-       CHECKERROR;
+                   CHECKERROR;
-       cgStaticError = source.run();
+                }
-       CHECKERROR;
+                break;
-       cgStaticError = sd.run();
+
-       CHECKERROR;
+                default:
-       cgStaticError = dup0.run();
+                break;
-       CHECKERROR;
+            }
-       cgStaticError = sc.run();
+        }
-       CHECKERROR;
-       cgStaticError = sb.run();
-       CHECKERROR;
-       cgStaticError = sa.run();
-       CHECKERROR;
-       cgStaticError = filter.run();
-       CHECKERROR;
-       cgStaticError = sd.run();
-       CHECKERROR;
-       cgStaticError = dup0.run();
-       CHECKERROR;
-       cgStaticError = sc.run();
-       CHECKERROR;
-       cgStaticError = sb.run();
-       CHECKERROR;
-       cgStaticError = sa.run();
-       CHECKERROR;
-
        debugCounter--;
        nbSchedule++;
     }
+
     *error=cgStaticError;
     return(nbSchedule);
 }

cmsisdsp/cg/static/scheduler/ccode.py

@@ -42,7 +42,10 @@ def gencode(sched,directory,config):
     schedDescription=""
 
     if config.codeArray:
-       ctemplate = env.get_template("codeArray.cpp")
+       if config.switchCase:
+          ctemplate = env.get_template("codeSwitch.cpp")
+       else:
+          ctemplate = env.get_template("codeArray.cpp")
        nb = 0
        for s in sched.schedule:
          schedDescription = schedDescription + ("%d," % sched.nodes[s].codeID)

cmsisdsp/cg/static/scheduler/config.py

@@ -63,7 +63,10 @@ class Configuration:
         # When codeArray  is true, instead of using
         # function calls we parse un array giving
         # the index of functions to call in another array
-        self.codeArray = False
+        self.codeArray = True
+        # If users do not want to use function pointers,
+        # we can generate a switch/case instead
+        self.switchCase = True
 
         # True for an horizontal graphviz layout
         self.horizontal = True

cmsisdsp/cg/static/scheduler/description.py

@@ -505,6 +505,10 @@ class Graph():
 
     def topologyMatrix(self):
         self.checkGraph()
+        # For cyclo static scheduling : compute the node periods
+        for n in self.nodes:
+            n.computeCyclePeriod()
+
         rows=[]
         # This is used in schedule generation
         # and all functions must use the same node ordering
@@ -522,21 +526,27 @@ class Graph():
             ib=self._sortedNodes.index(nb.owner)
 
             # Produced by na on the edge
-            currentRow[ia] = na.nbSamples
+            # for execution of one cycle of the na.owner node
+            totalProduced = na.cycleTotal * na.owner.cyclePeriod // na.cyclePeriod
+            currentRow[ia] = totalProduced
 
             # Consumed by nb on the edge
-            currentRow[ib] = -nb.nbSamples
+            # for execution of a full cycle of the node
+            totalProduced = nb.cycleTotal * nb.owner.cyclePeriod // nb.cyclePeriod
+            currentRow[ib] = -totalProduced
 
             rows.append(currentRow)
 
         return(np.array(rows))
 
-    def nullVector(self):
+    def nullVector(self,m):
-        m = self.topologyMatrix()
+        #print("Null vector")
+        # m is topology matrix computed with toplogyMatrix
         r=Matrix(m).nullspace()
         if len(r) != 1:
            raise NotSchedulableError
         result=list(r[0])
+        #print(result)
         denominators = [x.q for x in result]
         # Remove denominators
         ppcm = ilcm(*denominators)
@@ -544,17 +554,72 @@ class Graph():
         intValues = [x * ppcm for x in result]
         # Convert intValues to the smallest possible values
         gcd = igcd(*intValues)
-        return([x / gcd for x in intValues])
+        return([x // gcd for x in intValues])
 
     @property
     def initEvolutionVector(self):
         """Initial FIFO state taking into account delays"""
         return(np.array([self.getDelay(x) for x in self.edges]))
 
-    def evolutionVectorForNode(self,nodeID):
+    def evolutionVectorForNode(self,nodeID,test=True):
         """Return the evolution vector corresponding to a selected node"""
-        v = np.zeros(len(self._sortedNodes))
+        # For a simple static scheduling, the toplogy matrix T
-        v[nodeID] = 1 
+        # is enough. If we know which node is executed, we have
+        # a node vector V where there is a 1 at the position of the
+        # execute node.
+        # And the data generated on the fifos is:
+        # T . V so the new fifo vector B' is given with
+        # B' = T .V + B
+        # But in case of cyclo static scheduling the topology
+        # matrix is for a full period of the node
+        # so 1 or several periods of each IO.
+        # And we don't have the granularity corresponding to 
+        # one node execution.
+        # For one node execution, we need to know where we are
+        # in the cycle on each IO
+        # and where we are in the cycle for the node to
+        # track how many full eriods of the node execution have been
+        # done.
+        # So this function is not just returning the node vector
+        # and letting the main function updating the fifos.
+        # This function is returning the new fifo state
+        # Node vector would have been
+        # v = np.zeros(len(self._sortedNodes))
+        # v[nodeID] = 1 
+        # for cyclo static scheduling
+        n = self._sortedNodes[nodeID]
+        #print(n)
+        v = np.zeros(len(self._sortedEdges))
+
+        # In test mode we are testing several nodes
+        # to find the best one to schedule.
+        # So we should not advance the cycle
+        # of the IOs
+        for i in n._inputs:
+            io = n._inputs[i]
+            edge = io._fifo
+            # If 0, it is a connection to a constant node
+            # so there is no FIFO
+            if len(edge)>0:
+               pos = self._sortedEdges.index(edge)
+               v[pos] = -io.cycleValue
+               if not test:
+                  io.advanceCycle()
+            
+
+        for o in n._outputs:
+            io = n._outputs[o]
+            edge = io._fifo
+            # If 0 it is a connection to a constant edge
+            # so there is no FIFO
+            if len(edge)>0:
+               pos = self._sortedEdges.index(edge)
+               v[pos] = io.cycleValue
+               if not test:
+                  io.advanceCycle()
+            
+
+        #print(v)
         return(v)
 
     
@@ -565,9 +630,21 @@ class Graph():
         # After this transform, each output should be connected to
         # only one output.
         self.insertDuplicates()
+
+        networkMatrix = self.topologyMatrix()
+
         # Init values
         initB = self.initEvolutionVector
-        initN = self.nullVector()
+        initN = self.nullVector(networkMatrix)
+        #print(initN)
+
+        # nullVector is giving the number of repetitions
+        # for a node cycle.
+        # So the number of iteration of the node must be multiplied
+        # by the cycle period for this node.
+        #for nodeID in range(len(self._sortedNodes)):
+        #    initN[nodeID] = initN[nodeID] * self._sortedNodes[nodeID].cyclePeriod
+
 
         # Current values (copys)
         b = np.array(initB)
@@ -579,7 +656,9 @@ class Graph():
            print(b)
 
         # Topology matrix
-        t = np.array(self.topologyMatrix())
+        t = np.array(networkMatrix)
+        #print(t)
+        #print(n)
 
         # Define the list of FIFOs objects
         nbFIFOS = t.shape[0]
@@ -588,7 +667,18 @@ class Graph():
             allFIFOs.append(FIFODesc(i))
 
         # Normalization vector
-        normV =  1.0*np.apply_along_axis(abs,1,t).max(axis=1)
+        # For static scheduling it is
+        #normV =  1.0*np.apply_along_axis(abs,1,t).max(axis=1)
+        # For cyclo static scheduling we need the max per execution
+        # and not the accumulated value for the run of a node
+        # period which may be several periods of an IO
+        #print(normV)
+        normV = np.zeros(len(self._sortedEdges)) 
+        for e in range(len(self._sortedEdges)):
+            (src,dst) = self._sortedEdges[e]
+            normV[e] = max(src.cycleMax, dst.cycleMax)
+
+        #print(normV)
 
         # bMax below is used to track maximum FIFO size 
         # occuring during a run of the schedule
@@ -615,8 +705,11 @@ class Graph():
 
         zeroVec = np.zeros(len(self._sortedNodes))
         evolutionTime = 0
-        # While there are remaining nodes to schedule
+        #print(self._sortedNodes)
+        # While there are remaining node periods to schedule
         while (n != zeroVec).any():
+            #print("")
+            #print(n)
             # Look for the best node to schedule
             # which is the one giving the minimum FIFO increase
             
@@ -629,10 +722,14 @@ class Graph():
             for node in self._sortedNodes:
                 # If the node can be scheduled
                 if n[nodeID] > 0:
-                   # Evolution vector if this node is selected
+                   # Evolution vector for static scheduling
-                   v = self.evolutionVectorForNode(nodeID)
+                   # v = self.evolutionVectorForNode(nodeID)
+                   # for cyclo static we need the new fifo state
+                   newB = self.evolutionVectorForNode(nodeID) + b
                    # New fifos size after this evolution
-                   newB = np.dot(t,v) + b
+                   # For static scheduling, fifo update would have been
+                   # newB = np.dot(t,v) + b
+                   #print(newB)
 
                    # Check that there is no FIFO underflow:
                    if np.all(newB >= 0):
@@ -656,13 +753,28 @@ class Graph():
             # Now  we have evaluated all schedulable nodes for this run
             # and selected the one giving the smallest FIFO increase
 
+            # Implementation for static scheduling
             # Real evolution vector for selected node
-            evol =  self.evolutionVectorForNode(selected)
+            # evol =  self.evolutionVectorForNode(selected)
             # Keep track that this node has been schedule
-            n = n - evol
+            # n = n - evol
             # Compute new fifo state
-            fifoChange = np.dot(t,evol)
+            # fifoChange = np.dot(t,evol)
+            # b = fifoChange + b
+
+            # Implementation for cyclo static scheduling
+            #print("selected")
+            fifoChange = self.evolutionVectorForNode(selected,test=False)
             b = fifoChange + b
+            # For cyclo static, we create an evolution vector 
+            # which contains a 1
+            # at a node position only if this node has executed
+            # its period.
+            # Otherwise the null vector is not decreased
+            v = np.zeros(len(self._sortedNodes))
+            v[selected] = self._sortedNodes[selected].executeNode() 
+            n = n - v
+
 
             if config.displayFIFOSizes:
                print(b)

cmsisdsp/cg/static/scheduler/node.py

@@ -29,6 +29,10 @@
 from jinja2 import Environment, FileSystemLoader, PackageLoader,select_autoescape
 import pathlib
 import os.path
+import numpy as np
+
+from sympy.core.numbers import ilcm
+
 
 class NoFunctionArrayInPython(Exception):
     pass
@@ -51,12 +55,69 @@ class IO:
     """Class of input / outputs"""
     def __init__(self,owner,name,theType,nbSamples):
         self._theType = theType
+        # For cycle static scheduling it may also be
+        # a list of samples for each iteration of a cycle
         self._nbSamples = nbSamples 
         self._owner = owner
         self._name = name
         self._fifo = [] 
         self.constantNode = None
 
+        # For cyclo static scheduling nbSamples is a list
+        # and when simulating the schedule we need to know
+        # where we currently are in the list
+        self._cyclePosition = 0 
+
+    # For cyclo static scheduling we advance the position in the
+    # cycle
+    def advanceCycle(self):
+        if isinstance(self.nbSamples,int):
+            pass 
+        else:
+            self._cyclePosition = self._cyclePosition + 1 
+            if self._cyclePosition == len(self.nbSamples):
+               self._cyclePosition = 0
+
+    # For cyclo static scheduling, get the value in the current
+    # cycle position
+    @property
+    def cycleValue(self):
+        if isinstance(self.nbSamples,int):
+            return(self.nbSamples)
+        else:
+            return(self.nbSamples[self._cyclePosition])
+
+    # For cyclo static scheduling: we need
+    # the length of the cycle 
+    @property
+    def cyclePeriod(self):
+        if isinstance(self.nbSamples,int):
+            return(1)
+        else:
+            return(len(self.nbSamples))
+
+    # For cyclo static scheduling we need the total
+    # data generated by a run of the cycle
+    @property
+    def cycleTotal(self):
+        if isinstance(self.nbSamples,int):
+            return(self.nbSamples)
+        else:
+            return(np.sum(self.nbSamples))
+
+    # For cyclo static we need the max in a period top use
+    # as a normalization factor to identify the most filled
+    # FIFO 
+    @property
+    def cycleMax(self):
+        if isinstance(self.nbSamples,int):
+            return(self.nbSamples)
+        else:
+            return(np.max(self.nbSamples))
+    
+    
+
+
     @property
     def fifo(self):
         return self._fifo
@@ -182,6 +243,15 @@ class BaseNode:
         # Literal arguments
         self.schedArgs=None 
 
+        # For cyclo static scheduling
+        # It is the LCM of the cycles of all
+        # connected edges.
+        # After this period, the node and its edges
+        # should be back to the original state
+        self.cyclePeriod = 1
+
+        self._positionInCycle = 0
+
     def __getattr__(self,name):
         """Present inputs / outputs as attributes"""
         if name in self._inputs:
@@ -198,6 +268,33 @@ class BaseNode:
            return(self._outputs[name])
         raise IndexError 
 
+    # For cyclo static scheduling we need to compute
+    # the cycle period for this node based upon the
+    # period of the IOs 
+    def computeCyclePeriod(self):
+        allLengths = []
+        for k in self._inputs:
+            allLengths.append(self._inputs[k].cyclePeriod)
+        for k in self._outputs:
+            allLengths.append(self._outputs[k].cyclePeriod)
+        if len(allLengths)>1:
+           self.cyclePeriod = ilcm(*allLengths)
+        else:
+           self.cyclePeriod = allLengths[0]
+
+    # For cyclo static scheduling, we need to track the number
+    # of period execution 
+    # We return 1 when a period has been executed 
+    # and this is used to decrease the period count in
+    # the null vector of the toplogy matrix
+    def executeNode(self):
+        self._positionInCycle = self._positionInCycle + 1 
+        if (self._positionInCycle == self.cyclePeriod):
+            self._positionInCycle = 0 
+            return(1)
+        else:
+            return(0)
+
     def addLiteralArg(self,l):
         if self.schedArgs:
             self.schedArgs.append(ArgLiteral(l))
@@ -279,11 +376,24 @@ class BaseNode:
         # Use ordered io names
         for io in self.inputNames:
             x = self._inputs[io]
-            ios.append("%s,%d" % (x.ctype,x.nbSamples))
+            # For cyclo static scheduling, we may have a list
+            # of samples
+            if isinstance(x.nbSamples,int):
+               ios.append("%s,%d" % (x.ctype,x.nbSamples))
+            else:
+                # In case of a list of samples
+                # thne templaet is receiving only the
+                # max value
+                m = np.max(x.nbSamples)
+                ios.append("%s,%d" % (x.ctype,m))
 
         for io in self.outputNames:
             x = self._outputs[io]
-            ios.append("%s,%d" % (x.ctype,x.nbSamples))
+            if isinstance(x.nbSamples,int):
+               ios.append("%s,%d" % (x.ctype,x.nbSamples))
+            else:
+                m = np.max(x.nbSamples)
+                ios.append("%s,%d" % (x.ctype,m))
 
         
         return("".join(joinit(ios,",")))
@@ -297,13 +407,24 @@ class BaseNode:
         """
         ios=[] 
         # Use ordered io names
+        # For cyclo static scheduling, the nbSamples
+        # may be a list and we use the max value
+        # for the code generayion
         for io in self.inputNames:
             x = self._inputs[io]
-            ios.append("%d" % x.nbSamples)
+            if isinstance(x.nbSamples,int):
+               ios.append("%d" % x.nbSamples)
+            else:
+                m = np.max(x.nbSamples)
+                ios.append("%d" % m)
 
         for io in self.outputNames:
             x = self._outputs[io]
-            ios.append("%d" % x.nbSamples)
+            if isinstance(x.nbSamples,int):
+               ios.append("%d" % x.nbSamples)
+            else:
+                m = np.max(x.nbSamples)
+                ios.append("%d" % m)
 
         
         return("".join(joinit(ios,",")))
@@ -589,7 +710,12 @@ class GenericFunction(GenericNode):
                tempgen.append("%s,input%dSize" % (x.ctype,inputid))
                constructortypes.append("FIFOBase<%s> &src%d" % (x.ctype,inputid))
                genericconstructorargs.append("src%d" % inputid)
-               ios.append("%s,%d" % (x.ctype,x.nbSamples))
+               # For cyclo static scheduling, nbSamples may be a list
+               if isinstance(x.nbSamples,int):
+                  ios.append("%s,%d" % (x.ctype,x.nbSamples))
+               else:
+                  m = np.max(x.nbSamples)
+                  ios.append("%s,%d" % (x.ctype,m))
                typenameID = typenameID + 1
 
         for io in self.outputNames:
@@ -602,7 +728,11 @@ class GenericFunction(GenericNode):
                tempgen.append("%s,output%dSize" % (x.ctype,outputid))
                constructortypes.append("FIFOBase<%s> &dst%d" % (x.ctype,outputid))
                genericconstructorargs.append("dst%d" % outputid)
-               ios.append("%s,%d" % (x.ctype,x.nbSamples))
+               if isinstance(x.nbSamples,int):
+                  ios.append("%s,%d" % (x.ctype,x.nbSamples))
+               else:
+                  m = np.max(x.nbSamples)
+                  ios.append("%s,%d" % (x.ctype,m))
                typenameID = typenameID + 1
 
         self._realInputs = inputid
@@ -697,7 +827,13 @@ class GenericFunction(GenericNode):
          theType=self._inputs[self.inputNames[0]].ctype
        else:
          theType=self._inputs[self.inputNames[0]].nptype
-       theLen = self._inputs[self.inputNames[0]].nbSamples
+       # For cyclo static scheduling, nbSamples may be a list
+       # and in this case we are uding the mac value
+       nbSamples = self._inputs[self.inputNames[0]].nbSamples
+       if isinstance(nbSamples,int):
+          theLen = self._inputs[self.inputNames[0]].nbSamples
+       else:
+          theLen = np.max(nbSamples)
        theId = 0
        # List of buffer and corresponding fifo to initialize buffers
        inputs=[]

cmsisdsp/cg/static/scheduler/templates/code.cpp

@@ -1,3 +1,4 @@
+{% extends "commonc.cpp" %}
 /*
 
 Generated with CMSIS-DSP Compute Graph Scripts.
@@ -21,50 +22,10 @@ The support classes and code is covered by CMSIS-DSP license.
 {% if config.cOptionalArgs %},{{config.cOptionalArgs}}{% endif %}
 {% endmacro -%}
 
-/***********
+{% block schedArray %}
+{% endblock %}
 
-FIFO buffers
+{% block scheduleLoop %}
-
-************/
-{% for fifo in fifos %}
-#define FIFOSIZE{{fifo.fifoID}} {{fifo.length}}
-{% endfor %}
-
-{% for buf in sched._graph._allBuffers %}
-#define BUFFERSIZE{{buf._bufferID}} {{buf._length}}
-{{buf._theType.ctype}} {{config.prefix}}buf{{buf._bufferID}}[BUFFERSIZE{{buf._bufferID}}]={0};
-
-{% endfor %}
-
-uint32_t {{config.schedName}}(int *error{{optionalargs()}})
-{
-    int cgStaticError=0;
-    uint32_t nbSchedule=0;
-{% if config.debug %}
-    int32_t debugCounter={{config.debugLimit}};
-{% endif %}
-
-    /*
-    Create FIFOs objects
-    */
-{% for id in range(nbFifos) %}
-{% if fifos[id].hasDelay %}
-    FIFO<{{fifos[id].theType.ctype}},FIFOSIZE{{id}},{{fifos[id].isArrayAsInt}}> fifo{{id}}({{config.prefix}}buf{{fifos[id].buffer._bufferID}},{{fifos[id].delay}});
-{% else %}
-    FIFO<{{fifos[id].theType.ctype}},FIFOSIZE{{id}},{{fifos[id].isArrayAsInt}}> fifo{{id}}({{config.prefix}}buf{{fifos[id].buffer._bufferID}});
-{% endif %}
-{% endfor %}
-
-    /* 
-    Create node objects
-    */
-{% for node in nodes %}
-{% if node.hasState %}
-    {{node.typeName}}<{{node.ioTemplate()}}> {{node.nodeName}}({{node.args}});
-{% endif %}
-{% endfor %}
-
-    /* Run several schedule iterations */
 {% if config.debug %}
     while((cgStaticError==0) && (debugCounter > 0))
 {% else %}
@@ -88,6 +49,5 @@ uint32_t {{config.schedName}}(int *error{{optionalargs()}})
 {% endif %}
        nbSchedule++;
     }
-    *error=cgStaticError;
+    
-    return(nbSchedule);
+{% endblock %}
-}

cmsisdsp/cg/static/scheduler/templates/codeSwitch.cpp

@@ -0,0 +1,46 @@
+{% extends "commonc.cpp" %}
+
+{% block schedArray %}
+/*
+
+Description of the scheduling. It is a list of nodes to call.
+The values are indexes in the previous array.
+
+*/
+static unsigned int schedule[{{schedLen}}]=
+{ 
+{{schedDescription}}
+};
+{% endblock %}
+
+{% block scheduleLoop %}
+{% if config.debug %}
+    while((cgStaticError==0) && (debugCounter > 0))
+{% else %}
+    while(cgStaticError==0)
+{% endif %}
+    {
+        /* Run a schedule iteration */
+        for(unsigned long id=0 ; id < {{schedLen}}; id++)
+        {
+            switch(schedule[id])
+            {
+                {% for nodeID in range(nbNodes) -%}
+                case {{nodeID}}:
+                {
+                   {{nodes[nodeID].cRun()}}
+                   CHECKERROR;
+                }
+                break;
+
+                {% endfor %}default:
+                break;
+            }
+        }
+{% if config.debug %}
+       debugCounter--;
+{% endif %}
+       nbSchedule++;
+    }
+
+{% endblock %}

cmsisdsp/cg/static/scheduler/templates/commonc.cpp

@@ -0,0 +1,75 @@
+/*
+
+Generated with CMSIS-DSP Compute Graph Scripts.
+The generated code is not covered by CMSIS-DSP license.
+
+The support classes and code is covered by CMSIS-DSP license.
+
+*/
+
+{% if config.dumpFIFO %}
+#define DEBUGSCHED 1
+{% endif %}
+
+#include "arm_math.h"
+#include "custom.h"
+#include "GenericNodes.h"
+#include "AppNodes.h"
+#include "scheduler.h"
+
+{% macro optionalargs() -%}
+{% if config.cOptionalArgs %},{{config.cOptionalArgs}}{% endif %}
+{% endmacro -%}
+
+{% block schedArray %}
+{% endblock %}
+
+/***********
+
+FIFO buffers
+
+************/
+{% for fifo in fifos %}
+#define FIFOSIZE{{fifo.fifoID}} {{fifo.length}}
+{% endfor %}
+
+{% for buf in sched._graph._allBuffers %}
+#define BUFFERSIZE{{buf._bufferID}} {{buf._length}}
+{{buf._theType.ctype}} {{config.prefix}}buf{{buf._bufferID}}[BUFFERSIZE{{buf._bufferID}}]={0};
+
+{% endfor %}
+
+uint32_t {{config.schedName}}(int *error{{optionalargs()}})
+{
+    int cgStaticError=0;
+    uint32_t nbSchedule=0;
+{% if config.debug %}
+    int32_t debugCounter={{config.debugLimit}};
+{% endif %}
+
+    /*
+    Create FIFOs objects
+    */
+{% for id in range(nbFifos) %}
+{% if fifos[id].hasDelay %}
+    FIFO<{{fifos[id].theType.ctype}},FIFOSIZE{{id}},{{fifos[id].isArrayAsInt}}> fifo{{id}}({{config.prefix}}buf{{fifos[id].buffer._bufferID}},{{fifos[id].delay}});
+{% else %}
+    FIFO<{{fifos[id].theType.ctype}},FIFOSIZE{{id}},{{fifos[id].isArrayAsInt}}> fifo{{id}}({{config.prefix}}buf{{fifos[id].buffer._bufferID}});
+{% endif %}
+{% endfor %}
+
+    /* 
+    Create node objects
+    */
+{% for node in nodes %}
+{% if node.hasState %}
+    {{node.typeName}}<{{node.ioTemplate()}}> {{node.nodeName}}({{node.args}});
+{% endif %}
+{% endfor %}
+
+    /* Run several schedule iterations */
+{% block scheduleLoop %}
+{% endblock %}
+    *error=cgStaticError;
+    return(nbSchedule);
+}