Improve Vector->AIEVec lowering strategy for arith.addi/addf/muli/mulf

[jamestcl-amd]

• For input example (after vector canonicalization) like below, the arith.mulf and arith.addf will not be converted to aievec.mul_elem and aievec.add_elem. This is due to the FMA/Conv check that specifically leave these operations unconverted for the aievec.conv op lowering.

module {
  func.func @func1(%arg0: memref<1024xbf16>, %arg1: memref<1xbf16>, %arg2: memref<1024xbf16>) {
    %cst = arith.constant dense<0.000000e+00> : vector<16xbf16>
    %cst_0 = arith.constant 0.000000e+00 : bf16
    affine.for %arg3 = 0 to 1024 step 16 {
      %0 = vector.transfer_read %arg0[%arg3], %cst_0 {in_bounds = [true]} : memref<1024xbf16>, vector<16xbf16>
      %1 = vector.transfer_read %arg1[%arg3], %cst_0 : memref<1xbf16>, vector<16xbf16>
      %2 = vector.extract %1[0] : vector<16xbf16>
      %3 = vector.broadcast %2 : bf16 to vector<16xbf16>
      %4 = arith.minf %0, %cst : vector<16xbf16>
      %5 = arith.mulf %3, %4 : vector<16xbf16>
      %6 = arith.maxf %0, %cst : vector<16xbf16>
      %7 = arith.addf %6, %5 : vector<16xbf16>
      vector.transfer_write %7, %arg2[%arg3] {in_bounds = [true]} : vector<16xbf16>, memref<1024xbf16>
    }
    return
  }
}

• Here are the checks that blocks the lowering of the above example:

• mlir-aie/lib/Dialect/AIEVec/Transforms/VectorToAIEVecConversions.cpp

  Lines 838 to 842 in 1cf0528

  	if ((lhsDefOp && isa<arith::MulIOp>(lhsDefOp)) ||
  	(rhsDefOp && isa<arith::MulIOp>(rhsDefOp)) ||
  	(lhsDefOp && isa<arith::MulFOp>(lhsDefOp)) ||
  	(rhsDefOp && isa<arith::MulFOp>(rhsDefOp)))
  	return failure();

• mlir-aie/lib/Dialect/AIEVec/Transforms/VectorToAIEVecConversions.cpp

  Lines 526 to 529 in c49db4f

  	auto isAddOp = [&](Operation *op) { return isa<arith::AddFOp>(op); };
  	// Verify it is not a part of FMA
  	if (mulOp->hasOneUse() && llvm::any_of(mulOp->getUsers(), isAddOp))
  	return failure();

• mlir-aie/lib/Dialect/AIEVec/Transforms/VectorToAIEVecConversions.cpp

  Lines 1849 to 1852 in c49db4f

  	auto isAddOp = [&](Operation *op) { return isa<arith::AddFOp>(op); };
  	// Verify it is not a part of FMA
  	if (op->hasOneUse() && llvm::any_of(op->getUsers(), isAddOp))
  	return true;