Additional metadata for kernel args

[coleramos425]

This PR extends Accordo's Snapshot functionality to capture GPU kernel dispatch dimensions (grid size and block/workgroup size) alongside kernel argument data. This enhancement provides additional context about kernel execution configuration, which is valuable for validation and debugging, specifically for kernel extraction tools.

Changes

Python API:

• Extended Snapshot dataclass with optional grid_size and block_size fields
• Updated Snapshot.summary() to display grid/block dimensions when available
• Modified capture_snapshot() to pass metadata file path via environment variable

C++ Runtime Hook:

• Added dispatch metadata extraction in write_packets() when processing HSA_PACKET_TYPE_KERNEL_DISPATCH packets
• Metadata written as JSON to a temporary file specified by ACCORDO_METADATA_FILE environment variable
• Captures grid dimensions (x, y, z) and workgroup dimensions (x, y, z) for the traced kernel

Example Usage

from accordo import Accordo

# Configure validation
config = Accordo.Config(
    kernel_name="matrixTransposeShared_0",
    kernel_args=[
        Accordo.KernelArg(name="out", type="float*"),
        Accordo.KernelArg(name="in", type="const float*"),
        Accordo.KernelArg(name="width", type="int"),
        Accordo.KernelArg(name="height", type="int"),
    ],
    tolerance=1e-6
)

validator = Accordo(config)

# Capture snapshot - grid/block dimensions automatically included
ref_snapshot = validator.capture_snapshot(
    binary=["./b2b_matrix_transpose_ref"],
    working_directory=".",
    timeout_seconds=30
)

print(ref_snapshot.summary())

Output:

Snapshot Summary:
  Binary: ./b2b_matrix_transpose_ref
  Working Directory: /workspace/examples/bank_conflict/b2b_matrix_transpose
  Execution Time: 1353.43ms
  Number of Arrays: 1
  Grid Size: x=1024, y=1024, z=1
  Block Size: x=16, y=16, z=1
  Array 0: shape=(1048576,), dtype=float32

[mawad-amd]

Looks good. Thanks!