Mlir

Module
└── func.func @main()
    └── region
        └── block ^bb0:
            └── operations (ops) producing SSA values

// example.mlir
func.func @add(%a: memref<4xf32>, %b: memref<4xf32>, %c: memref<4xf32>) {
  %c0 = arith.constant 0 : index
  %c4 = arith.constant 4 : index
  scf.for %i = %c0 to %c4 step %c1 {
    %av = memref.load %a[%i] : memref<4xf32>
    %bv = memref.load %b[%i] : memref<4xf32>
    %sum = arith.addf %av, %bv : f32
    memref.store %sum, %c[%i] : memref<4xf32>
  }
  return
}

# Parse and print
mlir-opt example.mlir

# Run canonicalization
mlir-opt example.mlir -canonicalize

# Lower affine to scf
mlir-opt affine.mlir -lower-affine

# Full pipeline toward LLVM
mlir-opt input.mlir \
  --linalg-bufferize \
  --convert-linalg-to-loops \
  --convert-scf-to-cf \
  --convert-arith-to-llvm \
  --convert-memref-to-llvm \
  --convert-func-to-llvm \
  -o llvm.mlir

// MyOps.td
include "mlir/IR/OpBase.td"

def My_Dialect : Dialect {
    let name = "my";
    let summary = "My custom dialect";
}

class My_Op<string mnemonic, list<Trait> traits = []> :
    Op<My_Dialect, mnemonic, traits>;

def AddOp : My_Op<"add", [Pure]> {
    let summary = "Add two values";
    let arguments = (ins AnyType:$lhs, AnyType:$rhs);
    let results = (outs AnyType:$result);
    let assemblyFormat = "$lhs `,` $rhs attr-dict `:` type($result)";
}

# Generate C++ from TableGen
mlir-tblgen -gen-op-defs MyOps.td -I include/ -o MyOps.cpp.inc

#include "mlir/IR/DialectImplementation.h"
#include "MyDialect.h"

#include "MyOps.cpp.inc"

void MyDialect::initialize() {
    addOperations<
#define GET_OP_LIST
#include "MyOps.cpp.inc"
    >();
}

#define GET_OP_CLASSES
#include "MyOps.cpp.inc"

#include "mlir/Conversion/LLVMCommon/ConversionTarget.h"
#include "mlir/Transforms/DialectConversion.h"

struct AddOpLowering : OpConversionPattern<my::AddOp> {
    using OpConversionPattern::OpConversionPattern;

    LogicalResult matchAndRewrite(my::AddOp op, OpAdaptor adaptor,
                                  ConversionPatternRewriter &rewriter) const override {
        rewriter.replaceOpWithNewOp<arith::AddIOp>(op, adaptor.getLhs(), adaptor.getRhs());
        return success();
    }
};

void populateLoweringPatterns(RewritePatternSet &patterns) {
    patterns.add<AddOpLowering>(patterns.getContext());
}

// In pass:
mlir::ConversionTarget target(*context);
target.addIllegalDialect<my::MyDialect>();
target.addLegalDialect<arith::ArithDialect>();

if (failed(applyPartialConversion(module, target, std::move(patterns))))
    signalPassFailure();

%0 = linalg.matmul ins(%A, %B : tensor<128x256xf32>, tensor<256x64xf32>)
                   outs(%C : tensor<128x64xf32>) -> tensor<128x64xf32>

# Torch-MLIR: PyTorch → MLIR
python -m torch_mlir.tools.import-onnx --onnx-model model.onnx -o model.mlir

# IREE: MLIR → GPU/CPU executable
iree-compile --iree-hal-target-backends=llvm-cpu model.mlir -o model.vmfb
iree-run-module --module=model.vmfb --function=main