JuliaGPU/CUDA.jl

CUDNN convolution incorrect for small images

Open

#848 opened on Apr 19, 2021

View on GitHub
 (4 comments) (0 reactions) (1 assignee)Julia (274 forks)batch import
good first issueneeds tests

Repository metrics

Stars
 (1,408 stars)
PR merge metrics
 (Avg merge 5d 5h) (16 merged PRs in 30d)

Description

The CUDNN convolution is giving incorrect results for small images, e.g.

julia> f(k, x, padding) = CUDA.CUDNN.cudnnConvolutionForward(CuArray(Float32.(k)),
                                                             CuArray(Float32.(x)),
                                                             padding=padding)
f (generic function with 1 method)

julia> size(f(rand(7, 1, 1, 1), rand(5, 5, 1, 1), 5))
(9, 15, 1, 1)  # Correct

julia> size(f(rand(7, 1, 1, 1), rand(3, 5, 1, 1), 5))
(3, 15, 1, 1)  # Should have been (7, 15, 1, 1)

julia> size(f(rand(7, 1, 1, 1), rand(1, 5, 1, 1), 5))
ERROR: CUDNNError: CUDNN_STATUS_BAD_PARAM (code 3)  # Should have been (5, 15, 1, 1)

The cause of the problem can be located to

https://github.com/JuliaGPU/CUDA.jl/blob/445081cf78a250941b2ecdef1a3a2ecead3fec44/lib/cudnn/convolution.jl#L157-L162

where the min silently changes the requested padding.

Cc: @denizyuret

Contributor guide