Accordo

Accordo automatically validates GPU kernel correctness by capturing and comparing kernel outputs from reference and optimized implementations.

Features

• Automatic kernel extraction: Uses KernelDB to extract kernel signatures from binaries
• Snapshot-based validation: Capture once, compare against multiple optimizations
• Configurable tolerance: Set precision requirements for floating-point comparisons
• Performance tracking: Measure and compare execution times

Installation

pip install "git+https://github.com/AMDResearch/intellikit.git#subdirectory=accordo"

Quick start

from accordo import Accordo

# Create validator for a specific kernel
validator = Accordo(binary="./app_ref", kernel_name="reduce_sum")

# Capture snapshots from reference and optimized binaries
ref = validator.capture_snapshot(binary="./app_ref")
opt = validator.capture_snapshot(binary="./app_opt")

# Compare with specified tolerance
result = validator.compare_snapshots(ref, opt, tolerance=1e-6)

if result.is_valid:
    print(f"PASS: {result.num_arrays_validated} arrays matched")
else:
    print(result.summary())

Testing multiple optimizations

validator = Accordo(binary="./ref", kernel_name="matmul")
ref = validator.capture_snapshot(binary="./ref")

for opt_binary in ["./opt_v1", "./opt_v2", "./opt_v3"]:
    opt = validator.capture_snapshot(binary=opt_binary)
    result = validator.compare_snapshots(ref, opt, tolerance=1e-6)
    print(f"{opt_binary}: {'PASS' if result.is_valid else 'FAIL'}")

CLI

accordo validate \
  --kernel-name NAME \
  --ref-binary PATH_TO_EXECUTABLE \
  --opt-binary PATH_TO_EXECUTABLE \
  [--tolerance FLOAT]          # default: 1e-6
  [--timeout SECONDS]          # per snapshot, default: 30
  [--working-dir DIR]          # default: .
  [--kernel-args 'n1:t1,...']
  [--log-level DEBUG|INFO|WARNING|ERROR]

Example:

accordo validate --kernel-name reduce_sum --ref-binary ./app_ref --opt-binary ./app_opt

API reference

Accordo(binary, kernel_name, **options)

Parameters:

• binary (str | list) — binary path to extract kernel signature from
• kernel_name (str) — name of the kernel to validate
• working_directory (str) — working directory (default: ".")
• log_level (str) — logging level (default: "WARNING")

Methods:

• capture_snapshot(binary, timeout_seconds=30) → Snapshot
• compare_snapshots(reference, optimized, tolerance=1e-6) → ValidationResult

Snapshot

Attribute	Type	Description
arrays	list[np.ndarray]	Captured output arrays
execution_time_ms	float	Execution time
grid_size	dict | None	Kernel grid dimensions
block_size	dict | None	Kernel block dimensions

ValidationResult

Attribute	Type	Description
is_valid	bool	Whether validation passed
num_arrays_validated	int	Total arrays checked
num_mismatches	int	Failed comparisons
mismatches	list[ArrayMismatch]	Detailed mismatch info

Methods:

• summary() → str — human-readable validation summary

Requirements

• Python >= 3.8
• ROCm toolchain
• KernelDB (automatically installed)