Replace the per-sample ring-buffer FIR convolution with block-level
overlap-save convolution using rustfft. For a block of M samples and
N taps the old approach costs O(N·M); the new one costs O(M log M),
with rustfft using SIMD (AVX2/SSE4) internally.
Key changes:
- Add rustfft = "6" dependency
- Add BlockFirFilter: overlap-save filter with pre-computed H(f) and
a single forward+inverse FFT pair per block (no per-sample multiply)
- ChannelDsp.process_block() now:
1. Batch-mixes entire block to baseband in one vectorisable loop
2. Applies BlockFirFilter to I and Q (one FFT pair each)
3. Decimates and demodulates as before
- Keep the old FirFilter for unit tests (sample-by-sample interface)
- Add BlockFirFilter unit tests (DC passthrough, length preservation)
- IQ_BLOCK_SIZE promoted to pub const for use in filter sizing
For the default config (4096-sample blocks, 64 taps, decim=40):
Old: ~262144 multiply-adds per FIR × 2 components = ~524k per block
New: ~2 × (3 × 8192 × log2(8192)) ops, all SIMD-vectorised by rustfft
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Stan Grams <sjg@haxx.space>
sjg
600257a7c4
