Pass options in Default pipeline

Author: Kavitha Madhu

include/TPP/PassBundles.td

@@ -37,10 +37,7 @@ def DefaultTppPasses : Pass<"default-tpp-passes", "ModuleOp"> {
            "unsigned", "Grid-sizes for parallel tasks.">,
     Option<"linalgToVector", "linalg-to-vector",
            "bool", /*default=*/"true",
-           "Lower linalg directly to vector.">,
-    Option<"vectorToXsmm", "vector-to-xsmm",
-           "bool", /*default=*/"true",
-           "Lower vector to xsmm.">
+           "Lower linalg directly to vector.">
   ];
 }
 

lib/TPP/DefaultPipeline.cpp

@@ -53,6 +53,10 @@ llvm::cl::list<unsigned>
                      llvm::cl::list_init<unsigned>(SmallVector<unsigned>{2, 8}),
                      llvm::cl::CommaSeparated);
 
+llvm::cl::opt<bool> linalgToVector("linalg-to-vector",
+                                   llvm::cl::desc("Lower linalg to vector"),
+                                   llvm::cl::init(false));
+
 namespace mlir {
 namespace tpp {
 #define GEN_PASS_DEF_DEFAULTPIPELINE
@@ -124,8 +128,8 @@ struct DefaultPipeline : public tpp::impl::DefaultPipelineBase<DefaultPipeline>,
       pm.addPass(createGpuPipeline(GpuPipelineOptions{gpuBackend}));
     } else {
       // Apply the default preprocessing pass
-      DefaultTppPassesOptions tppDefaultOptions{linalgToLoops,
-                                                parallelTaskGrid};
+      DefaultTppPassesOptions tppDefaultOptions{linalgToLoops, parallelTaskGrid,
+                                                linalgToVector};
       pm.addPass(createDefaultTppPasses(tppDefaultOptions));
     }
 

lib/TPP/DefaultTppPasses.cpp

@@ -21,6 +21,7 @@
 #include "TPP/Dialect/Perf/PerfDialect.h"
 #include "TPP/Dialect/Xsmm/XsmmDialect.h"
 #include "TPP/PassUtils.h"
+#include "mlir/Transforms/Passes.h"
 
 using namespace mlir;
 using namespace mlir::tpp;
@@ -97,8 +98,14 @@ struct DefaultTppPasses
       // Bufferize: tensor->memref.
       pm.addPass(createBufferize());
 
-      // Lower all Tile operations.
-      pm.addNestedPass<func::FuncOp>(createLinalgLowering());
+      if (linalgToVector) {
+        pm.addNestedPass<func::FuncOp>(createVectorizationPass());
+        pm.addNestedPass<func::FuncOp>(createVectorContractPass());
+        pm.addNestedPass<func::FuncOp>(createCanonicalizerPass());
+      } else {
+        // Lower all Tile operations.
+        pm.addNestedPass<func::FuncOp>(createLinalgLowering());
+      }
       pm.addPass(createCleanup());
     }
 
@@ -109,10 +116,18 @@ struct DefaultTppPasses
     // Low leve parallelization passes.
     LowLevelParallelizationOptions LowLevelParallelization{parallelTaskGrid};
     pm.addPass(createLowLevelParallelization(LowLevelParallelization));
-
-    // Covert all local TPP-related dialects.
-    pm.addPass(createLocalDialectsLowering());
-
+    if (linalgToVector) {
+      pm.addPass(createConvertVectorToSCFPass());
+    } else {
+      // Covert all local TPP-related dialects.
+      pm.addPass(createLocalDialectsLowering());
+
+      pm.addNestedPass<func::FuncOp>(createIntelAMXTileConfigInsertionPass());
+      pm.addNestedPass<func::FuncOp>(createCanonicalizerPass());
+      pm.addNestedPass<func::FuncOp>(createLoopInvariantCodeMotionPass());
+      pm.addNestedPass<func::FuncOp>(createCanonicalizerPass());
+      pm.addNestedPass<func::FuncOp>(createIntelAMXTileConfigHoistingPass());
+    }
     // Clean up after the default pipeline.
     pm.addNestedPass<func::FuncOp>(createPostprocessing());
   }

lib/TPP/PassBundles/LowLevelParallelization.cpp

@@ -65,11 +65,5 @@ struct LowLevelParallelization
     mlir::tpp::SCFParallelLoopTilingOptions tilingOptions;
     tilingOptions.tileSizes = parallelTaskGrid;
     pm.addPass(createSCFParallelLoopTiling(tilingOptions));
-
-    pm.addNestedPass<func::FuncOp>(createIntelAMXTileConfigInsertionPass());
-    pm.addNestedPass<func::FuncOp>(createCanonicalizerPass());
-    pm.addNestedPass<func::FuncOp>(createLoopInvariantCodeMotionPass());
-    pm.addNestedPass<func::FuncOp>(createCanonicalizerPass());
-    pm.addNestedPass<func::FuncOp>(createIntelAMXTileConfigHoistingPass());
   }
 };

lib/TPP/Transforms/Vectorization.cpp

@@ -44,7 +44,7 @@ struct LinalgGenericToVector : OpRewritePattern<linalg::GenericOp> {
       SmallVector<int64_t> shape;
       SmallVector<ReassociationIndices> indices;
       int index = 0;
-      for (int i = 0; i < dyn_cast<ShapedType>(linalgOp.getOperand(0).getType())
+      for (size_t i = 0; i < dyn_cast<ShapedType>(linalgOp.getOperand(0).getType())
                               .getShape()
                               .size();
            i++) {