Motivation
The L1 array layer (element-wise arithmetic + sum) is in place, but the headline
use case — spectral analysis on the edge — needs an FFT.
Concretely: streaming a live vibration/audio spectrum from a Raspberry Pi 5 to a Rails
"control plane" (a Kaigi on Rails demo) currently has to fall back to a GPU-synthesised
spectrum, because there is no FFT yet. Adding FFT turns the same pipeline into a real
spectral analyzer.
Current state
GPU kernels today (shader/): add / adds / sub / mul / div / scale / sum
— element-wise arithmetic + a sum reduction. There is no fft method or shader.
Proposal
Add an FFT on Vulkan compute, exposed as a Ruby method, following the existing
shader + method pattern (*.comp → *.spv + mrb_define_method in src/gpu_narray.c):
- API sketch
GPU::SFloat#rfft → complex spectrum for real input (length n power of two)#power_spectrum / #magnitude → GPU::SFloat of length n/2 (ready to quantise/plot)
- Backend options
- Hand-written radix-2 Cooley–Tukey compute shader (GLSL) — fits the current
architecture, ~150 LoC, high learning/demo value.
- Integrate VkFFT — production-grade Vulkan FFT
(larger C++ integration).
Data already lives in a VkBuffer; the transform should run in-place on the GPU and
only copy back via #to_a. This is the natural L2 (DSP) layer.
Acceptance criteria
Context
Toward 名古屋Ruby会議05 (2026-09-19): "the code you prototype with is the code you
deploy to the edge." An FFT is the first DSP primitive that makes that story concrete.
Motivation
The L1 array layer (element-wise arithmetic +
sum) is in place, but the headlineuse case — spectral analysis on the edge — needs an FFT.
Concretely: streaming a live vibration/audio spectrum from a Raspberry Pi 5 to a Rails
"control plane" (a Kaigi on Rails demo) currently has to fall back to a GPU-synthesised
spectrum, because there is no FFT yet. Adding FFT turns the same pipeline into a real
spectral analyzer.
Current state
GPU kernels today (
shader/):add / adds / sub / mul / div / scale / sum— element-wise arithmetic + a sum reduction. There is no
fftmethod or shader.Proposal
Add an FFT on Vulkan compute, exposed as a Ruby method, following the existing
shader + method pattern (
*.comp → *.spv+mrb_define_methodinsrc/gpu_narray.c):GPU::SFloat#rfft→ complex spectrum for real input (lengthnpower of two)#power_spectrum/#magnitude→GPU::SFloatof lengthn/2(ready to quantise/plot)architecture, ~150 LoC, high learning/demo value.
(larger C++ integration).
Data already lives in a
VkBuffer; the transform should run in-place on the GPU andonly copy back via
#to_a. This is the natural L2 (DSP) layer.Acceptance criteria
n-point (power-of-two) real→complex FFT as a Vulkan compute dispatchGPU::SFloatexamples/fft_spectrum.rb— FFT of a known multi-tone signal, peaks at expected binstest/Context
Toward 名古屋Ruby会議05 (2026-09-19): "the code you prototype with is the code you
deploy to the edge." An FFT is the first DSP primitive that makes that story concrete.