Simulation-focused multicycle RV32 subset CPU with scalar MUL and a blocking, single-lane digital matvec accelerator. Arithmetic is signed 32-bit with wraparound. One shared RAM port, one outstanding read, no CPU data cache. The accelerator's private vector buffer defaults to 16 words; benchmark configurations provision 128. This is a functional digital model, without physical energy/area characterization.
Linux/WSL requires Verilator 5.x, GNU Make, Bash, Python 3 and a C++ compiler.
make verify # Historical RTL, protocol, reset and benchmark regression
make gemm-study # Square matrix multiplication: 16 and 32
make gemm-large-study # Square matrix multiplication: 64 and 128Both GEMM studies have a matching -check target to replay recorded evidence.
GEMM uses streaming CPU, four-output CPU tiling and one cold offload per A row.
Complete-job timing includes setup, launches, waits, status and completion.
Operands start in RAM: A/C row-major, B column-major. Host placement and
layout conversion are excluded consistently.
- 16×16 / 32×32 GEMM
- 64×64 / 128×128 GEMM
- Larger-vector tiling
- Output-count scaling
- Instruction-cache study
rtl/ contains CPU, memory, accelerator and system modules. tb/ contains
self-checking tests. programs/ stores generated images; scripts/ generates,
runs and checks them. Production defaults and historical images remain separate
from benchmark overrides.
See the documentation index, architecture, accelerator contract and roadmap.