FFT Spectrum Analyser
Signals & AudioA 32-point
radix-2 DIT FFT
runs every frame on a Super Nintendo, transforming a synthesised sine tone into 16 frequency
bars. The tone sweeps from bin 1 to 15 — the tallest bar follows it. Each of 5 butterfly
stages runs 16 complex multiplies: 320 __mulsi3 calls per frame. Compiled with
the llvm-mos-based 65816 toolchain
(+mos-a16).
Self-running — the tone sweeps through 15 frequency bins automatically. (Boot holds while the gate CRC computes.)
Source fft.cfft.hspiro.h in wbniv/llvm-mos-65816
What it is
Each frame the demo generates 32 samples of a sine tone, runs the full DIT FFT, computes 16 magnitude bars (Manhattan approximation: |re| + |im|), and repaints the canvas. The canvas clears and redraws every frame; the dirty-tile DMA streams the update across ~4 v-blanks, giving smooth progressive animation.
// DIT butterfly — 4× __mulsi3 per butterfly:
int32_t t_re = ((int32_t)TW_RE[j*w] * xr[bot]
- (int32_t)TW_IM[j*w] * xi[bot]) >> 8;
int32_t t_im = ((int32_t)TW_RE[j*w] * xi[bot]
+ (int32_t)TW_IM[j*w] * xr[bot]) >> 8;
xr[bot] = xr[top] - t_re; xi[bot] = xi[top] - t_im;
xr[top] += t_re; xi[top] += t_im; Compiler stress-test #25 — FFT butterfly + bit-reversal
A Round 2 demo covering two new corners: the bit-reversal
permutation (index shuffling never seen in prior demos) and the
butterfly complex multiply — the same __mulsi3
corner as Round 1, but in a structurally different context (staged butterfly vs
per-pixel iteration) that exercises different register-allocation pressure.
| Corner | What it exercises |
|---|---|
| Butterfly twiddle multiply | 4× __mulsi3 per butterfly: t_re = (TW_RE[j*w] * xr[bot] - TW_IM[j*w] * xi[bot]) >> 8 and the imaginary counterpart. Each 16×16→32-bit product is one __mulsi3 libcall. 5 stages × 16 butterflies × 4 __mulsi3 = 320 multiply calls per FFT frame. The disasm gate confirmed __mulsi3=4 (folded by -Os loop optimisation — the hot multiply is the dominant operation). |
| Bit-reversal permutation | Before the butterfly passes, samples are rearranged in bit-reversed index order: sample at index 3 (binary 00011) moves to index 24 (binary 11000). The precomputed FFT_BITREV[32] table drives this swap — a pure bit-manipulation corner distinct from all prior demos. |
| Q8.8 fixed-point complex multiply | Twiddle factors W_32^k = cos(-2πk/32) × 256 + i·sin(-2πk/32) × 256 are stored as int16_t pairs. The butterfly multiplies them by the data words (also int16_t Q8.8), right-shifts 8 to re-normalise. The product range is 256 × 256 = 65536, which fits in int32_t — the multiplication MUST be 32-bit or the result overflows, making this a genuine __mulsi3 workload. |
| Sweeping tone demonstration | The signal is a pure sine at frequency bin cur_bin, generated from the SPIRO_SIN_LUT. cur_bin steps from 1 to 15 every ~1 second. The tallest bar should be at cur_bin; neighbouring bars show FFT windowing leakage. Correct butterfly arithmetic → one tall bar; a miscompile → wrong bar heights or the wrong bin is tallest. |
Written in C with the llvm-mos 65816 toolchain. Hit Verify fidelity to
reproduce the build gate's WRAM assert (gate CRC 0x6D7A) live in this tab.
Disasm gate: __mulsi3=4, rep/sep=63.
No compiler bug found — FFT butterfly codegen correct across all modes.