SpMV-RVV

Benchmarking sparse matrix-vector multiplication (SpMV) using scalar and RISC-V Vector Extension (RVV) implementations. Targets EPAC and Orange Pi RV2 RISC-V hardware, with x86 support for development and testing.

The benchmark iterates SpMV and reports timing, cycle counts, memory footprint, and FLOP counts for each sparse format.

Code organisation

• ./src — C source files (formats, SpMV kernels, driver, main)
• ./build — compiled executables and run scripts
• ./data — test Matrix Market (.mtx) files
• ./tests — Google Test unit tests (C++)

Sparse matrix formats

Scalar formats

Format	Description
COO	Coordinate format
CSR	Compressed Sparse Row
ELL	ELLPACK format
DIA	Diagonal format

Vectorised (RVV) formats

Format	Description
vCSRVG	Vectorised CSR — vector-permute
vCSRMG	Vectorised CSR — index-load
vELLVG	Vectorised ELL — vector-permute
vELLRVG	Vectorised ELL — vector-permute (roofline)
vELLMG	Vectorised ELL — index-load
vDIAVG	Vectorised DIA — vector-slide
vDIARVG	Vectorised DIA — vector-slide (roofline)

Format group aliases

Group	Included formats
SCALAR	COO, CSR, ELL, DIA
VECTOR	vCSRVG, vCSRMG, vELLVG, vELLMG, vDIAVG
VCSR	vCSRVG, vCSRMG
VELL	vELLVG, vELLMG
VDIA	vDIAVG
ROOFLINE	vELLRVG, vDIARVG
ALL	All formats

Build

Configuration

Edit the top of Makefile to configure the build before compiling:

Variable	Values	Description
FLOAT_TYPE	USE_FLOAT / USE_DOUBLE	Floating-point precision (default: USE_FLOAT)
rvv	1 / 0	Enable RVV intrinsics (default: 1)
DEBUG	1 / 0	Enable debug output (default: 0)
ARCH	e.g. rv64gcv	RISC-V march string

The compiler and flags sections contain pre-configured profiles for:

• gcc / clang — Orange Pi (default active)
• clang — EPAC (comment/uncomment as needed)
• spike simulator (for software emulation)

Floating-point precision can also be switched per-build without editing the Makefile:

$ make riscv64 FLOAT_TYPE=USE_DOUBLE

Build targets

# Build for x86 (development / correctness testing)
$ make x86

# Build for RISC-V with RVV
$ make riscv64

# Build unit tests (requires Google Test)
$ make test

# Remove all build artefacts
$ make clean

Executables are placed in ./build/:

• spmv_x86
• spmv_rvv

Run

Executable usage

./build/spmv_rvv <matrix.mtx> <format> <block_size_row> <block_size_col> <iters> <print_result>

Argument	Description
matrix.mtx	Path to a Matrix Market file
format	Format or group name (see tables above)
block_size_row	Row block size (use 0 to disable blocking)
block_size_col	Column block size
iters	Number of SpMV iterations to time
print_result	1 to print the output vector, 0 to suppress

Example — run all formats with block size 8 for 10 iterations:

$ ./build/spmv_rvv ./data/dwt_512.mtx ALL 8 8 10 0

Example — run scalar formats only, print result vector:

$ ./build/spmv_x86 ./data/test.mtx SCALAR 16 16 1 1

Convenience scripts

Script	Description
./build/run_x86.sh	Run a quick x86 test on matrix_128k_full.mtx with COO
./build/run_rvv.sh [block_size]	Run vDIAVG on dwt_512.mtx with the given block size
./build/run_epac.sh	Run vELLVG on dwt_512.mtx (EPAC configuration)
./build/run_orangepi.sh [block_size]	Run ALL formats on pwtk.mtx for Orange Pi
./build/run_benchmark.sh	Full benchmark over all .mtx files in ../benchmark_v2/

Binary matrix cache

On the first run the program parses the .mtx file and writes a binary cache alongside it (e.g. matrix_64_float.bin). Subsequent runs with the same block size and precision load the binary directly, skipping the text parser.

Output

Each benchmark line is printed by print_stats in the following format:

format     rows       cols     blocks    total_elem   nnz       mem(MB)   flops    iters    time(s)    cycles   result

• format — sparse format name
• rows / cols — matrix dimensions
• blocks — number of matrix blocks
• total_elem — allocated storage elements (including padding)
• nnz — number of non-zero elements
• mem(MB) — memory footprint in megabytes
• flops — floating-point operations per SpMV
• iters — iteration count
• time(s) — average wall-clock time per iteration
• cycles — hardware cycle count (via rdcycle, RISC-V only)
• result — running mean of the output vector (for correctness cross-checking)

Tests

Unit tests use Google Test. Build and run with:

$ make test
$ ./tests/csr_test
$ ./tests/coo_test

Adding a new SpMV format

Follow the six-step checklist at the top of src/spMv.c:

1. Implement the kernel function in src/spMv.c
2. Add a dispatch case in spMv() in src/spMv.c
3. Add the FLOP count case in get_flops_spMv() in src/spMv.c
4. Add the enum value to E_SPMV_FORMAT in src/mtx_sparse.h
5. Add the format string to str_spmv_format[] in src/mtx_sparse.c (same position as the enum)
6. Update the argument parser and format groups in src/main.c