# Unreasonably fast FFT on CUDA

**URL:** <https://discourse.julialang.org/t/unreasonably-fast-fft-on-cuda/109398>\
**Category:** Performance\
**Tags:** question\
**Created:** [January 29, 2024, 2:12pm UTC](https://discourse.julialang.org/t/unreasonably-fast-fft-on-cuda/109398 "2024-01-29T14:12:22Z")\
**Posts on this page:** 1\
**Showing post:** 8

<div class="post-metadata">

**Author:** ![roflmaostc](https://sea2.discourse-cdn.com/julialang/user_avatar/discourse.julialang.org/roflmaostc/32/30123_2.png) [@roflmaostc](https://discourse.julialang.org/u/roflmaostc)\
**Post date:** [January 29, 2024, 7:00pm UTC](https://discourse.julialang.org/t/unreasonably-fast-fft-on-cuda/109398/8 "2024-01-29T19:00:43Z")

</div>

Usually 3D convolutions are much faster with FFTs, if the kernel is not like (2,2,2).

Also just dumping my numbers.

```julia
julia> using FFTW, CUDA, BenchmarkTools

julia> function try_FFT_on_cuda()
           values = rand(353, 353, 353)
           value_complex = ComplexF32.(values)
           cvalues = similar(cu(value_complex), ComplexF32)
           copyto!(cvalues, values)
           cy = similar(cvalues)
           cF = plan_fft!(cvalues)
           @btime CUDA.@sync a = ($cF*$cy)
           return nothing
       end
try_FFT_on_cuda (generic function with 1 method)

julia> try_FFT_on_cuda()
  20.319 ms (6 allocations: 304 bytes)

julia> function try_FFT_on_cpu()
           values = rand(353, 353, 353)
           value_complex = ComplexF32.(values)
           cvalues = similar((value_complex), ComplexF32)
           copyto!(cvalues, values)
           cy = similar(cvalues)
           cF = plan_fft!(cvalues, flags=FFTW.MEASURE)
           @btime a = ($cF*$cy)
           return nothing
       end
try_FFT_on_cpu (generic function with 1 method)

julia> try_FFT_on_cpu()
  100.503 ms (314 allocations: 27.09 KiB)

# power of 2
julia> function try_FFT_on_cpu()
           values = rand(256, 256, 256)
           value_complex = ComplexF32.(values)
           cvalues = similar((value_complex), ComplexF32)
           copyto!(cvalues, values)
           cy = similar(cvalues)
           cF = plan_fft!(cvalues, flags=FFTW.MEASURE)
           @btime a = ($cF*$cy)
           return nothing
       end
try_FFT_on_cpu (generic function with 1 method)

julia> function try_FFT_on_cuda()
           values = rand(256, 256, 256)
           value_complex = ComplexF32.(values)
           cvalues = similar(cu(value_complex), ComplexF32)
           copyto!(cvalues, values)
           cy = similar(cvalues)
           cF = plan_fft!(cvalues)
           @btime CUDA.@sync a = ($cF*$cy)
           return nothing
       end
try_FFT_on_cuda (generic function with 1 method)

julia> try_FFT_on_cpu()
  18.883 ms (314 allocations: 27.09 KiB)

julia> try_FFT_on_cuda()
  2.483 ms (6 allocations: 304 bytes)

```

---

_[View the full topic](https://discourse.julialang.org/t/unreasonably-fast-fft-on-cuda/109398)._