Benchmark Guide

Benchmark guide

Repository boundaries

Path	Responsibility
00_open_graph	Pinned OpenGraph CPU/reference framework and graph structures
01_capi_precis	Pinned CAPI-Precis AFU-control dependency
02_capi_graph	AccelGraph host entry point, CAPI algorithms, tests, MMIO runtime
03_capi_integration	Graph compute-unit RTL, simulator design, synthesis, images
04_test_graphs	Small and suite-specific benchmark inputs
05_scripts	CU selection and version helpers
06_slides	Original papers, diagrams, and current brand assets

Execution stages

1. Parse graph, algorithm, direction, threading, reorder, cache, and CAPI options.
2. Load or convert the edge list and construct the selected graph structure.
3. Run preprocessing and optional relabeling.
4. Select the OpenMP or CAPI implementation.
5. For CAPI, map graph arrays into the WED and launch one or more CU rounds.
6. Record timing and algorithm statistics.
7. Compare results through the existing reference or report path.

CAPI coverage

The active layout manifest is authoritative:

Algorithm	Active layout
BFS	CSR PULL BottomUp
PageRank	CSR PULL FloatPoint, FixedPoint, Quantized
SPMV	CSR PULL FloatPoint, FixedPoint
Connected components	CSR ShiloachVishkin
Triangle counting	CSR BinaryIntersection

PageRank PUSH FloatPoint, FixedPoint, and Quantized remain exact expected failures until their missing RTL source sets are restored. Other research-stage source paths are not active accelerator support. CU selection is determined by data structure, algorithm, direction, and precision in 05_scripts/choose_algorithm_*.py.

Current portable real-top evidence limits PageRank PULL, SPMV PULL, and TriangleCount to -K1. ConnectedComponents passes its full topology; the multi-CU ring blocker for the other layouts is recorded in Verification infrastructure.

Common options

Option	Purpose
-f	Input graph
-z	Input format: text edge list, binary edge list, or binary CSR
-d	Graph data structure
-a	Algorithm
-p	Pull, push, or hybrid direction
-r	Root vertex for traversal algorithms
-i	Iteration count
-n, -N, -K	Preprocessing, algorithm, and accelerator-kernel thread/CU counts
-l, -L, -O	First, second, and third reorder stages
-C	Cache-size model
-M	Cache-mask mode
-t	Trial count
-e	Numeric tolerance
-b	Delta for delta-stepping workloads
-s	Symmetrize the graph
-w	Load or generate weights

Pass custom arguments through the existing Make targets, for example:

make run-openmp ARGS='-f <graph> -z 1 -d 0 -a 0 -p 0 -r 0 -N 8 -t 1'

The selected binary's --help output is the exact parser reference.

Evidence for a benchmark result

A reportable CAPI result needs all of the following:

• exact graph file and format;
• algorithm, direction, precision, reorder, cache, thread, and CU settings;
• host, submodule, and FPGA image revisions;
• bounded accelerator completion with no error register;
• completion reset before any next round;
• result match or algorithm-specific correctness evidence;
• elapsed time and trial count.

See accelerator verification for liveness and the deployment runbook for operational evidence.