RVV Mel Spectrogram

An audio front end (FFT, power spectrum, mel filter bank) that I vectorised by hand with RISC-V Vector Extension (RVV 1.0) intrinsics. It is benchmarked on the Spike simulator with the cycle counter (rdcycle, Zicntr) and checked against the course's NumPy reference, scripts/mel_spectrogram.py.

_{Mel spectrogram of librosa's trumpet example (n_fft 512, hop 160, 40 mel bands), supplied with the assignment. The melspec_trumpet benchmark runs the kernels on the first second of this clip and compares against data/mel_spectrogram.txt.}

Lab 2 of NCKU Computer Organization (CSIE, Spring 2026). See the lab series below.

Kernels

All three are in src/main.c.

fft is a radix-2 Cooley-Tukey. Bit-reversal is in place with a reversed counter (j ^= bit). Then log2 n butterfly stages, with twiddle factors precomputed per stage. Butterflies run vl at a time with LMUL = 8, and the complex multiply is vfmul + vfnmsac / vfmacc.

power_spectrum uses two strided loads (vlse32, stride 8 bytes) to split the interleaved re, im pairs, then re*re and vfmacc(im, im). One loop over the flat frames x 257 array.

mel_filter_bank computes out[f][m] = sum_k power[f][k] * bank[m][k]. I process four mel rows at once so each power-spectrum load is reused four times, then reduce with vfredusum. A scalar tail handles n_mels % 4.

All vector loops are strip-mined with vsetvl, so nothing depends on VLEN.

Constant	Value	Meaning
N_FFT	512	FFT frame size
HOP_LENGTH	160	Samples between frames (16 kHz audio, 10 ms hop)
N_FREQ_BINS	257	N_FFT / 2 + 1
N_MELS	40	Mel bands
MAX_FFT_N	4096	Largest FFT the kernel must support

Layout

src/main.c              my RVV implementation (fft, power_spectrum, mel_filter_bank)
src/utils.c             provided: hann_window, stft, melspectrogram pipeline
src/bench.c             provided: correctness + cycle benchmark harness
include/mel_spectrogram.h
scripts/                NumPy ground truth, librosa cross-check, scoring script
data/                   reference inputs and expected outputs
assets/                 provided: spectrogram illustration
Makefile, pyproject.toml, uv.lock

Build and run

Needs the RISC-V GNU toolchain, Spike, pk and uv, all preinstalled in the course image:

docker run -it --rm -v "$(pwd)":/workspace -w /workspace docker.io/asrlab/comp-org:pa2

make compile   # riscv64-unknown-linux-gnu-gcc -O3 -fno-tree-vectorize -march=rv64gcv ...
make run       # runs build/bench on Spike (RV64GCV_Zicntr), writes output/results.csv
make judge     # compile + run + score correctness and speed-up vs. the scalar baseline

-fno-tree-vectorize turns off auto-vectorisation, so the speed-up is from the intrinsics.

Not done yet

• mel_filter_bank at LMUL = 8 has only four register groups, but the 4-row block loads five values (power + 4 rows) before using them. A smaller LMUL should fix that.
• Keep a vfmacc accumulator and call vfredusum once per dot product instead of once per strip.
• Compute twiddle factors once for the largest stage and index with a stride (W_len^k = W_N^{k·N/len}) to drop the per-stage sin/cos calls.

Lab series

Lab	Repository	Topic
1	riscv-inline-asm-algorithms	RV64IF inline assembly
2	rvv-mel-spectrogram (this repo)	RISC-V Vector (RVV) intrinsics, FFT, DSP
3	cache-aware-riscv-optimization	Tree-PLRU cache simulator, cache-blocked transpose, RVV GEMM

---

_{The benchmark harness, pipeline glue (utils.c, bench.c), Python reference, data and the spectrogram image in assets/ were provided by the course staff. The three kernels in src/main.c are my own work.}