Add ELF output format support for NPU2 (AIE2P)

[erwei-xilinx]

Summary

• Add ELF output format as default for NPU2 (AIE2P/Strix), replacing the two-file xclbin+insts.bin workflow with a single self-contained aie.elf binary
• NPU1 (AIE2) continues using xclbin unchanged — no behavioral change for legacy devices
• New AMD_TRITON_NPU_OUTPUT_FORMAT env var allows overriding the auto-detected default ("elf" on npu2, "xclbin" on npu1)

What changed in driver.py

• _get_output_format(): auto-detects format based on NPU version, with env var override and npu1+elf validation
• _generate_elf_launcher(): C++ launcher template using xrt::elf, xrt::ext::kernel, xrt::ext::bo, and xrt::run APIs
• _extract_elf_kernel_name(): parses full_elf_config.json to discover the kernel name (e.g., main:vecadd) since ELF uses IR-derived names instead of MLIR_AIE
• compile_module(): conditional linker flags (ELF doesn't need test_utils/boost), aircc invocation (--output-format elf --elf-name), caching (2 files vs 3), and module loading
• NPULauncher.__init__(): dispatches to correct launcher generator based on format

Test plan

• Vec-add on NPU2 with ELF (default) — PASS
• Full test suite on NPU2 (14/15 pass, matvec fails in both ELF and xclbin — pre-existing)
• Matvec with AMD_TRITON_NPU_OUTPUT_FORMAT=xclbin on NPU2 — same failure (confirms pre-existing)
• NPU1 regression test (xclbin path unchanged)
• AMD_TRITON_NPU_OUTPUT_FORMAT=elf on NPU1 — should raise RuntimeError

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR adds ELF output support for NPU2 (AIE2P/Strix) to make aie.elf the default binary output, while keeping the existing xclbin workflow for NPU1 (AIE2). It also introduces an environment-variable override to force either format.

Changes:

• Add output-format auto-detection with AMD_TRITON_NPU_OUTPUT_FORMAT override (default: ELF on npu2, xclbin on npu1).
• Add ELF-specific C++ launcher generation and switch aircc invocation/caching to produce and load aie.elf (+ kernel name metadata).
• Add a new CLAUDE.md development guide covering repo structure and compilation pipeline.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 4 comments.

File	Description
amd_triton_npu/backend/driver.py	Implements format selection, ELF launcher path, aircc ELF build flags, and format-aware caching/loading.
CLAUDE.md	Adds a large development guide describing the project, build, CI, and compilation pipeline (including ELF on npu2).

Comments suppressed due to low confidence (1)

amd_triton_npu/backend/driver.py:10

• tempfile is imported twice (import tempfile and again in import os, subprocess, tempfile, platform). This is redundant and can confuse linting/formatting; remove the duplicate import and keep a single style for imports.

import hashlib
import json
import tempfile
import sys
import sysconfig

import os, subprocess, tempfile, platform

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.