NPU Support for shark-sept24-release

[kumardeepakamd]

OS Support: Windows 11 and Ubuntu 22.04 Devices: NPU(Phoenix, Strix) Owners track the progress, make sure tests are added to CI, and update the status every week. Multiple engineers will contribute to an item.

Item	Owner	July 1	July 8	Target (Sept30)
BFP16 support for Strix (https://github.com/nod-ai/iree-amd-aie/issues/437)	@erwei-xilinx	Metric	Metric	Vectorized performant support
conv	@newling	Metric	Metric	All shapes
matmul	@yzhang93	Metric	Metric	All shapes
fused-elementwise-ops	@yzhang93	Metric	Metric	All element-wise ops
Padding	@yzhang93	Metric	Metric	padding for all cases
Conversion of dispatches to objectfifo	@jtuyls	Metric	Metric	No dependency on mlir-air
Direct lowering of data movement	@makslevental	Metric	Metric	Directly generate config and BDs at performance parity with mlir-aie
Direct lowering of ops to llvm-aie	@jsetoain	Metric	Metric	IREE to llvm-aie without mlir-aie AIEVec dialect
[DONE] Vectorization support in objectfifo	@Abhishek-Varma	Metric	Metric	connect vectorization work to objectfifo pipeline
[DONE] Mixing of codegen and microkenel	@Abhishek-Varma	Metric	Metric	fused op calling one matmul chess compiled ukernel
Efficient Core ELF, routing and data config switching	@nirvedhmeshram	Metric	Metric	Generate incremental config for dispatches and use Control Packets to effect that
Heterogeneous device offloading-to-aie e2e flow	@daveliddell	Metric	Metric	Seamless --device=amd-aie working
model: vae	@newling	Metric	Metric	All dispatches for vae decoder working on NPU
model: sd3,sdxl,llama3	@MaheshRavishankar	Metric	Metric	sd3, sdxl, llama3 running inference with offload-to-aie
model: 3k onnx models	@kumardeepakamd	Metric	Metric	3k models running inference with offload-to-aie
pip install iree-amd-aie and run	@makslevental	Metric	Metric	Minimal user steps to be able to target aie