Fix MBSK, used WavefrontSize to cacualte required synchornizer size.

Author: vin-huang

tensilelite/Tensile/Contractions.py

@@ -469,7 +469,7 @@ def CompoundPredicates(cls, state, problemType):
                 valuepredicates.append(int((state["NumElementsPerThread"])/state["NumElementsPerBatchStore"]))
             else:
                 valuepredicates.append(1)
-            valuepredicates.append(state["NumThreads"])
+            valuepredicates.append(ceil(state["NumThreads"] / state["WavefrontSize"]))
             rv += [cls('SynchronizerSizeCheck', index=0, value=valuepredicates)]
 
         if state["InternalSupportParams"]["KernArgsVersion"] >= 1 and \
@@ -605,8 +605,8 @@ def FromOriginalState(cls, d):
             globalAccum = 4
         pgr = int(d['PrefetchGlobalRead'])
         synchronizerSizePerWG = ceil((d['MIWaveTile'][0]*d['MIWaveTile'][1] if d['EnableMatrixInstruction'] else d['ThreadTile0']*d['ThreadTile1']        \
-                                    * ceil((d['NumElementsPerThread'])/d['NumElementsPerBatchStore']) if d['NumElementsPerBatchStore'] != 0 else 1            \
-                                    * ceil(d["NumThreads"] / 64)))
+                                    * ceil((d['NumElementsPerThread'])/d['NumElementsPerBatchStore']) if d['NumElementsPerBatchStore'] != 0 else 1        \
+                                    * ceil(d["NumThreads"] / d["WavefrontSize"])))
 
         return cls(waveNum                  = d['NumThreads'] // d['WavefrontSize'],
                    workGroup                = d['WorkGroup'],

tensilelite/Tensile/Source/lib/include/Tensile/ContractionProblemPredicates.hpp

@@ -230,10 +230,10 @@ namespace TensileLite
                 virtual bool operator()(ContractionProblemGemm const& problem) const override
                 {
                     // WorkGroup numbers x number of global write instruction x Wave numbers
-                    // M/MT0 x N/MT1 x NumElementsPerThread/StoreVectorWidth x x Wavenumbers
+                    // M/MT0 x N/MT1 x NumElementsPerThread/StoreVectorWidth x x Wavenumbers x batch
                     bool ret = (std::ceil(static_cast<float>(problem.freeSizeA(0)) / value[0])
                                 * std::ceil(static_cast<float>(problem.freeSizeB(0)) / value[1]))
-                                   * (value[2]) * (value[4] / 64) * value[3]
+                                   * value[2] * value[4] * value[3] * problem.d().sizes()[2]
                                <= 409600;
                     if(problem.groupedGemm())
                         ret = ret && (problem.groupedGemmCount() <= 16);