Xsmm dialect cleanup

include/TPP/Dialect/Xsmm/CMakeLists.txt

@@ -1,8 +1,5 @@
-add_mlir_dialect(XsmmOps xsmm)
-add_mlir_doc(XsmmDialect XsmmDialect TPP/ -gen-dialect-doc)
-add_mlir_doc(XsmmOps XsmmOps TPP/ -gen-op-doc)
-
 set(LLVM_TARGET_DEFINITIONS XsmmEnum.td)
 mlir_tablegen(XsmmEnum.h.inc -gen-enum-decls)
 mlir_tablegen(XsmmEnum.cpp.inc -gen-enum-defs)
 add_public_tablegen_target(MLIRXsmmAttrDefIncGen)
+

include/TPP/Dialect/Xsmm/XsmmEnum.td

@@ -8,7 +8,6 @@
 
 include "mlir/IR/AttrTypeBase.td"
 include "mlir/IR/EnumAttr.td"
-include "TPP/Dialect/Xsmm/XsmmDialect.td"
 
 def Xsmm_DataType: I64EnumAttr<
     "DataType", "see: libxsmm_datatype",

include/TPP/Dialect/Xsmm/XsmmUtils.h

@@ -10,7 +10,6 @@
 #define TPP_DIALECT_XSMM_XSMMUTILS_H
 
 #include "TPP/Dialect/Xsmm/XsmmEnum.h"
-#include "TPP/Dialect/Xsmm/XsmmOps.h"
 #include "TPP/IR/StructuredOpMatcher.h"
 #include "mlir/Dialect/Func/IR/FuncOps.h"
 #include "mlir/Dialect/Linalg/IR/LinalgInterfaces.h"
@@ -65,22 +64,6 @@ struct BinaryInfo {
   int64_t ldo;
 };
 
-/// Represents a chain of XSMM ops that can be fused. All broadcast ops
-/// should have already been converted to flags. All stray allocations
-/// should have already been converted to in-place reuse.
-struct FusedMatch {
-  // This is the (optional) zero op that precedes the GEMM op
-  UnaryOp zeroOp;
-  // This is the BRGEMM op
-  BrgemmOp brgemmOp;
-  // This is the (optional) binary op that follows the GEMM
-  BinaryOp binaryOp;
-  BinaryKind binaryKind;
-  // This is the (optional) unary op that follows the GEMM/Binary
-  UnaryOp unaryOp;
-  UnaryKind unaryKind;
-};
-
 namespace utils {
 
 DataTypeAttr getDataType(RewriterBase &rewriter, Type type);
@@ -114,23 +97,12 @@ FailureOr<BinaryFlags> getBinaryFlagsVectorType(Type operandType,
 
 FailureOr<int64_t> getLeadingDim(Type type, size_t pos = 0);
 
-FailureOr<FusedMatch> getFusedBrgemmSequenceFromProducer(Operation *op);
-
-ArrayAttr getUnaryDispatchFlags(UnaryOp op);
-
-ArrayAttr getBinaryDispatchFlags(BinaryOp op);
-
 int64_t getOredFlags(ArrayAttr flags);
 
 SmallVector<Type> extractInvokeOperandTypes(OpBuilder &builder,
                                             ValueRange operands);
 SmallVector<Value> getOperands(OpBuilder &builder, Location loc,
                                ValueRange operands, IntegerAttr dataTypeAttr);
-template <typename DispatchOpTy>
-FailureOr<SmallVector<Attribute>> getBrgemmFlags(PatternRewriter &rewriter,
-                                                 DispatchOpTy dispatchOpTy,
-                                                 bool returnNone);
-
 FailureOr<BrgemmInfo> isMappableToBrgemm(PatternRewriter &rewriter,
                                          vector::ContractionOp contractOp,
                                          SmallVector<Value> &inputs,

include/TPP/PassBundles.td

@@ -61,21 +61,13 @@ def TppMapping : Pass<"tpp-mapping", "ModuleOp"> {
   ];
 }
 
-def LinalgLowering : Pass<"linalg-lowering", "func::FuncOp"> {
-  let summary = "Lower Linalg operations to XSMM operations.";
-  let dependentDialects = ["xsmm::XsmmDialect",
-                           "scf::SCFDialect",
-                           "memref::MemRefDialect"];
-}
-
 def LowLevelParallelization : Pass<"low-level-parallel", "ModuleOp"> {
   let summary = "Low level parallelization (multi-threading, AMX config).";
   let dependentDialects = ["affine::AffineDialect",
                            "arith::ArithDialect",
                            "func::FuncDialect",
                            "memref::MemRefDialect",
                            "scf::SCFDialect",
-                           "xsmm::XsmmDialect",
                            "LLVM::LLVMDialect"];
   let options = [
     ListOption<"parallelTaskGrid", "parallel-task-grid",
@@ -94,7 +86,6 @@ def LocalDialectsLowering : Pass<"lower-local-dialects", "ModuleOp"> {
                            "perf::PerfDialect",
                            "scf::SCFDialect",
                            "tensor::TensorDialect",
-                           "xsmm::XsmmDialect",
                            "LLVM::LLVMDialect"];
 }
 

include/TPP/Passes.h

@@ -84,10 +84,6 @@ namespace vector {
 class VectorDialect;
 } // namespace vector
 
-namespace xsmm {
-class XsmmDialect;
-} // namespace xsmm
-
 namespace xegpu {
 class XeGPUDialect;
 } // namespace xegpu

include/TPP/Passes.td

@@ -11,18 +11,6 @@
 
 include "mlir/Pass/PassBase.td"
 
-def ConvertLinalgToXsmm : Pass<"convert-linalg-to-xsmm", "func::FuncOp"> {
-  let summary = "Convert linalg to xsmm";
-  let description = [{
-    Convert linalg operations to XSMM operations.
-  }];
-  let dependentDialects = ["func::FuncDialect",
-                           "memref::MemRefDialect",
-                           "linalg::LinalgDialect",
-                           "xsmm::XsmmDialect",
-                           "tensor::TensorDialect"];
-}
-
 def ConvertVectorToXsmm : Pass<"convert-vector-to-xsmm", "func::FuncOp"> {
   let summary = "Convert vector to xsmm";
   let description = [{
@@ -36,15 +24,6 @@ def ConvertVectorToXsmm : Pass<"convert-vector-to-xsmm", "func::FuncOp"> {
 }
 
 
-def VerifyXsmmCalls : Pass<"verify-xsmm-calls", "func::FuncOp"> {
-  let summary = "Verify XSMM calls (dispatch and invoke)";
-  let description = [{
-    Make sure XSMM dispatch and invoke call are in a consistent
-    state and they do not contradict each others.
-  }];
-  let dependentDialects = [ "xsmm::XsmmDialect" ];
-}
-
 def ConvertLinalgToFunc : Pass<"convert-linalg-to-func", "ModuleOp"> {
   let summary = "Convert linalg to func";
   let description = [{
@@ -67,17 +46,6 @@ def VectorizationPass : Pass<"vectorization-pass",
 }
 
 
-def ConvertXsmmToFunc : Pass<"convert-xsmm-to-func", "ModuleOp"> {
-  let summary = "Convert xsmm to func";
-  let description = [{
-    Convert XSMM operations to libXSMM function calls.
-  }];
-  let dependentDialects = ["func::FuncDialect",
-                           "memref::MemRefDialect",
-                           "xsmm::XsmmDialect",
-                           "LLVM::LLVMDialect"];
-}
-
 def ConvertCheckToLoops : Pass<"convert-check-to-loops", "func::FuncOp"> {
   let summary = "Convert check to loops";
   let description = [{
@@ -215,15 +183,6 @@ def RewriteBatchMatmulToMatmul : Pass<"rewrite-batch-matmul-to-matmul",
   let dependentDialects = ["scf::SCFDialect", "linalg::LinalgDialect"];
 }
 
-def CombineXsmmOpPass : Pass<"combine-xsmm-op-optimization", "func::FuncOp"> {
-  let summary = "Fuse brgemm-add-relu ops into a fused brgemm op";
-  let description =
-      [{Fuse brgemm-add-relu ops into a fused brgemm op}];
-
-  let dependentDialects = ["xsmm::XsmmDialect"];
-
-}
-
 def PropagatePackUnPack : Pass<"propagate-pack-and-unpack", "func::FuncOp"> {
   let summary = "Propagate tensor.pack and tensor.unpack";
   let description = [{
@@ -353,22 +312,6 @@ def GpuDataTransfer : Pass<"gpu-data-transfer", "func::FuncOp"> {
                            "gpu::GPUDialect"];
 }
 
-def FoldXsmmFlags : Pass<"fold-xsmm-flags", "func::FuncOp"> {
-  let summary = "Attempt to fold dispatch op as flags in XSMM.";
-  let description = [{
-    Attempt to fold dispatch operations as flags in consumer dispatch
-    operations, for example:
-    ```mlir
-      %alloc = memref.alloc
-      xsmm.unary zero (%alloc)
-      xsmm.gemm.dispatch (%alloc)
-    ```
-    the zero is folded as `beta_0` in `xsmm.gemm.dispatch`.
-  }];
-  let dependentDialects = [ "memref::MemRefDialect", "xsmm::XsmmDialect" ];
-}
-
-
 def SCFParallelLoopTiling : Pass<"scf-parallel-loop-tiling-pass"> {
   let summary = "Tile parallel loops";
   let options = [

lib/TPP/Conversion/CMakeLists.txt

@@ -1,7 +1,5 @@
 add_subdirectory(ConvertCheckToLoops)
 add_subdirectory(ConvertLinalgToFunc)
-add_subdirectory(ConvertLinalgToXsmm)
 add_subdirectory(ConvertPerfToFunc)
 add_subdirectory(ConvertPerfToLoops)
-add_subdirectory(ConvertXsmmToFunc)
 add_subdirectory(ConvertVectorToXsmm)