Embench: Integer matrix multiplication¶
Compiled workload catalogue · Execution contract
Purpose¶
Multiply two fixed 20 by 20 integer matrices.
Setup¶
| Property | Value |
|---|---|
| Source workload | matmult-int |
| Pinned source | src/matmult-int/matmult-int.c |
| Revision | Embench 1.0, commit 0466a18e4f6b47e19598d7c6ba72916d54b68f65 |
| Build selection | bash tools/rv32i-benchmarks/build-embench.sh matmult-int |
| Resulting image | build/rv32i-embench/artifacts/matmult-int/matmult-int.elf |
| Completion | Upstream verifier, main return, then committed tohost verdict |
These are the upstream fixed inputs and repetition rules, not the native generator's size/iterations options. This page documents source behavior; a successful build or full current-core run is not claimed. See port limitations.
Implementation¶
Copy the reference inputs into working arrays. A three-level row/column/dot-product loop computes the result matrix, using the upstream repetition factor 46.
Expected result¶
Compare the entire 400-element result matrix with the stored expected matrix, not only a checksum.
The shared upstream main converts verifier failure into a nonzero return. The CircuitSim startup bridge publishes tohost=1 on success or 3 on failure. A checked instruction prefix cannot establish this final result.
Measurement and running¶
Useful as a future CPU/accelerator comparison workload, but current execution is ordinary compiled RV32I. Multiplication is software-expanded; there is no M-extension or accelerator instruction.
Build this source explicitly, inspect its ELF size/instruction requirements, then follow the bounded prefix and campaign procedure. Full-run timing includes startup, warmup, benchmark body, verifier, and verdict unless a separate valid measurement window is selected.
Source and tests¶
The pinned upstream source contains the benchmark body and verifier. Other files in that source directory may provide helper implementations. Local port files are under tools/rv32i-benchmarks/embench/.
The ordinary repository regression suite tests the ELF loader/runner using small fixtures; it does not download and run this workload automatically. Keep its external verifier result, runner/oracle result, and any structural/behavioral comparison as separate recorded checks.