Source kernel compilation

Tensile/Common.py

@@ -441,19 +441,6 @@ def supportedCompiler(compiler: str) -> bool:
 validMFMA["I8"] = validMFMA["H"] + validMFMA["F8"]
 validWMMA = [[16,16,16,1], ]
 validTT = 64
-validMFMA["_format9"] = []
-
-for MFMA in [validMFMA["H"], validMFMA["S"], validMFMA["B"], validMFMA["D"], validMFMA["X"], validMFMA["F8"], validWMMA]:
-  for MI in MFMA:
-    for bm in range(int(math.log(MI[3],2))+1):
-      for tt0 in range(1,validTT+1):
-        for tt1 in range(1,validTT+1):
-          for wave_m in range (3):
-            for wave_n in range(3):
-              validMFMA["_format9"].append([MI[0],MI[1],MI[2],MI[3],2**bm,tt0,tt1,2**wave_m, 2**wave_n])
-
-validMatrixInstructions = [[], [-1]] + validMFMA["H"] + validMFMA["S"] + validMFMA["B"] + validMFMA["D"] + validMFMA["X"] + validMFMA["F8"]
-validMatrixInstructions = validMatrixInstructions + validMFMA["_format9"]
 
 # The supported typed GEMM, each entry is (Ti, To, Tc).
 # DataType (Ti)        = The data-type of the input matrices: A/B
@@ -1060,7 +1047,7 @@ def supportedCompiler(compiler: str) -> bool:
     #      MatrixInst  BlkM   WT    Wave
     #  - means (32x64) per MI * (4x1) per wave * (2x2) per workgroup = (32*4*2)x(64*1*2) = 256x128 macro tile
     # Tensile will ignore the parameters ThreadTile and WorkGroup when the alternative format is used
-    "MatrixInstruction":          validMatrixInstructions,
+    "MatrixInstruction":          -1,
 
     # StoreRemap: Optimize MatrixInstruction store patterns to enhance performance.
     #             MI output data between each threads are along N dims.
@@ -2407,6 +2394,22 @@ def assignGlobalParameters( config, capabilitiesCache: Optional[dict] = None ):
   """
   global globalParameters
 
+  if False:
+    validMFMA["_format9"] = []
+
+    for MFMA in [validMFMA["H"], validMFMA["S"], validMFMA["B"], validMFMA["D"], validMFMA["X"], validMFMA["F8"], validWMMA]:
+      for MI in MFMA:
+        for bm in range(int(math.log(MI[3],2))+1):
+          for tt0 in range(1,validTT+1):
+            for tt1 in range(1,validTT+1):
+              for wave_m in range (3):
+                for wave_n in range(3):
+                  validMFMA["_format9"].append([MI[0],MI[1],MI[2],MI[3],2**bm,tt0,tt1,2**wave_m, 2**wave_n])
+
+    validMatrixInstructions = [[], [-1]] + validMFMA["H"] + validMFMA["S"] + validMFMA["B"] + validMFMA["D"] + validMFMA["X"] + validMFMA["F8"]
+    validMatrixInstructions = validMatrixInstructions + validMFMA["_format9"]
+    validParameters["MatrixInstruction"] = validMatrixInstructions
+
   # Minimum Required Version
   if "MinimumRequiredVersion" in config:
     if not versionIsCompatible(config["MinimumRequiredVersion"]):

Tensile/SolutionLibrary.py

@@ -246,11 +246,20 @@ def remapSolutionIndices(self, indexMap):
         for row in self.rows:
             row["library"].remapSolutionIndices(indexMap)
 
+    def __repr__(self) -> str:
+        return str(self.rows)
+
 
 class MasterSolutionLibrary:
     StateKeys = ["solutions", "library"]
     ArchitectureSet = set()
 
+    def __init__(self, solutions, library, version=None):
+        self.lazyLibraries = {}
+        self.solutions = solutions
+        self.library = library
+        self.version = version
+
     @classmethod
     def ArchitectureIndexMap(cls, architectureName: str) -> int:
         """Maps hex characters from gfx name to an index.
@@ -275,12 +284,19 @@ def ArchitectureIndexMap(cls, architectureName: str) -> int:
             if archString is not None:
                 archLiteral = archString.group(0)
                 archval = (int(archLiteral, 16) << 18)
+
+        # TODO(@bstefanuk) this section of code needs to be removed when dealing with
+        # parallel processing since each process may operate on multiple fallbacks that
+        # match a problem type.
+
         # Check for duplicate architecture values
-        if archval >= 0 and not archval in cls.ArchitectureSet:
-            cls.ArchitectureSet.add(archval)
-        else:
-            tPrint(1, f"ERROR: Duplicate architecture value {archval} for {architectureName}, with arch set {cls.ArchitectureSet}")
-            raise RuntimeError("ERROR in architecture solution index mapping.")
+        # if archval >= 0 and not archval in cls.ArchitectureSet:
+        #     cls.ArchitectureSet.add(archval)
+        # else:
+        #     tPrint(1, f"ERROR: Duplicate architecture value {archval} for {architectureName}, with arch set {cls.ArchitectureSet}")
+        #     raise RuntimeError("ERROR in architecture solution index mapping.")
+
+        cls.ArchitectureSet.add(archval)
         return archval
 
     @classmethod
@@ -455,10 +471,10 @@ def selection(d, problemType, solutions, library, placeholderName):
         library = None
         placeholderName = "TensileLibrary"
         placeholderLibrary = None
-        for libName in reversed(libraryOrder):
-            library, placeholderName = libName(origData, problemType, allSolutions, library,
+        for updateNameFunc in reversed(libraryOrder):
+            library, placeholderName = updateNameFunc(origData, problemType, allSolutions, library,
                                                placeholderName)
-            if libName == placeholder:
+            if updateNameFunc == placeholder:
                 placeholderLibrary = library
 
         solutions = {s.index: s for s in allSolutions}
@@ -486,12 +502,6 @@ def BenchmarkingLibrary(cls, solutions):
 
         return cls(solutionMap, library)
 
-    def __init__(self, solutions, library, version=None):
-        self.lazyLibraries = {}
-        self.solutions = solutions
-        self.library = library
-        self.version = version
-
     def state(self):
         rv = {
             "solutions": state(iter(list(self.solutions.values()))),
@@ -599,6 +609,9 @@ def merge(self, other, startIndex=0):
 
         return curIndex  #Next unused index
 
+    def __repr__(self) -> str:
+        return str(self.library)
+
     @property
     def cpp_base_class(self):
         return "SolutionLibrary<ContractionProblem, ContractionSolution>"