Add hasProperty method to targetModel

[erwei-xilinx]

The hasProperty method checks whether a target model has a given device property (implemented as enum).
List of initial properties added:

• UsesSemaphoreLocks
• IsNPU
• IsVirtualized
• UsesMultiDimensionalBDs

[fifield]

This is probably better than the current convention which is to test for !aie1. Should we update such cases in this PR or gradually?

[jgmelber]

Isn't it enough to check if the code is targeting AIE1. Otherwise it has semaphore locks?

[erwei-xilinx]

Isn't it enough to check if the code is targeting AIE1. Otherwise it has semaphore locks?

Functionality wise, yes we can definitely check for aie1. It's just that when other projects such as air use mlir-aie's target models, it isn't so intuitive for outside developers as there isn't a logical connection between "the device is not aie1" and "the device is using semaphore locks".

[erwei-xilinx]

This is probably better than the current convention which is to test for !aie1. Should we update such cases in this PR or gradually?

Sure, I updated some cases with the hasProperty method in this PR.

[fifield]

AIE2 does not imply IsNPU. It should get inherited from BaseNPUTargetModel or BaseNPU3TargetModel.

[erwei-xilinx]

Thanks for finding the bug. Fixed.

[fifield]

Suggested change

	virtual bool hasProperty(ModelProperty Prop) const = 0;
	bool hasProperty(ModelProperty Prop) const { return false; }

But it might also be worth looking at how rtti works in AIETargetModel. Using a similar scheme, each subclass would add its bitvector of properties via the call to the base class constructor. The base class would OR the bitvector with its own, so the subclass only needs to add new properties not present in the base class.

#include <iostream>
#include <cstdlib>

class ClassA {
private:
  const uint32_t features;
public:
  ClassA(uint32_t f) : features(f) {}
  uint32_t getFeatures() const { return features; };
};

class ClassB : public ClassA {
  static constexpr uint32_t class_b_features = 0x1;
public:
  ClassB(uint32_t f) : ClassA(f | class_b_features) {}
  ClassB() : ClassA(class_b_features) {}
};

class ClassC : public ClassB {
  static constexpr uint32_t class_c_features = 0x2;
public:
  ClassC() : ClassB(class_c_features) {}
};

class ClassD : public ClassB {
  static constexpr uint32_t class_d_features = 0x4;
public:
  ClassD() : ClassB(class_d_features ) {}
};

int main(int argc, char *argv[]) {
  std::cout << "B " << ClassB().getFeatures() << "\n";
  std::cout << "C " << ClassC().getFeatures() << "\n";
  std::cout << "D " << ClassD().getFeatures() << "\n";
}

$ ./a.out
B 1
C 3
D 5

A side benefit is that the vector of properties gets built once when the TargetModel objects are created, instead of each time hasProperty() is called.

[erwei-xilinx]

Good point. PR updated. Thanks!

[fifield]

This doesn't add anything and can be covered by the base class AIETargetModel ?

[erwei-xilinx]

Agreed. Now removed in PR.

[fifield]

This should be private

[erwei-xilinx]

Good point. Let me fix this in a separate PR.

[erwei-xilinx]

PR created: #1944