[Tuner][Codegen] Make default tuning be available per-SKU instead of per-arch

[kuhar]

We have multiple GPUs that share the same ISA architecture, for example gfx940 is implemented by MI300X, MI300A, and MI308X. These targets have different number of workgroup processors and different power caps. Because of this, they require dedicated tuning specs.

We should update our tuning specs to be per-SKU first, and then per-architecture as a fallback. We can start with MI300X and MI308X.

The default tuning specs are stored in https://github.com/iree-org/iree/tree/main/compiler/plugins/target/ROCM/builtins/tuning while the logic to decide which file to pick up is in

iree/compiler/src/iree/compiler/Codegen/Common/MaterializeTuningSpecsPass.cpp

Lines 126 to 133 in 3e8c81c

	// Try to look up the default tuning spec for this architecture, if any.
	StringRef arch = gpuTarget.getArch();
	std::string defaultTuningSpecName =
	llvm::formatv("iree_default_tuning_spec_{}.mlir", arch);
	std::optional<StringRef> defaultTuningSpecSource;
	EmbeddedDataDirectory::withGlobal([&](EmbeddedDataDirectory &dir) {
	defaultTuningSpecSource = dir.getFile(defaultTuningSpecName);
	});

.

To test this, we can modify the existing test to contain these RUN lines: https://github.com/iree-org/iree/blob/main/compiler/plugins/target/ROCM/builtins/tuning/test/spec_gfx942.mlir

• one for --iree-gpu-test-target=mi300x
• one for --iree-gpu-test-target=mi300a -- check that without a spec for this SKU, we load the spec for gfx942
• one for --iree-gpu-test-target=gfx942 -- we can still support sku-agnostic targets like we do today

[kuhar]

cc: @MaheshRavishankar