Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

add depthwise_conv* overloads for CUDA #22

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

Conversation

DhairyaLGandhi
Copy link
Member

No description provided.

Comment on lines +99 to +112
function depthwise_conv!(y::DenseCuArray{T}, x::DenseCuArray{T}, w::DenseCuArray{T}, cdims::DepthwiseConvDims;
alpha = 1, beta = 0, algo = -1) where T <: CUDNNFloat
conv!(y, x, w, cims; alpha, beta, algo)
end

function ∇depthwise_conv_filter!(dw::DenseCuArray{T}, x::DenseCuArray{T}, dy::DenseCuArray{T},
cdims::ConvDims; alpha = 1, beta = 0, algo = -1) where T <: CUDNNFloat
∇conv_filter!(dw, x, dy, cdims; alpha, beta, algo)
end

function ∇depthwise_conv_data!(dx::DenseCuArray{T}, dy::DenseCuArray{T}, w::DenseCuArray{T},
cdims::ConvDims; alpha = 1, beta = 0, algo = -1) where T <: CUDNNFloat
∇conv_data!(dx, dy, w, cdims; alpha, beta, algo)
end
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

these don't have to be cuda specific, we can add them to NNlib and remove the specific implementations (after a performance comparison)

Copy link
Member Author

@DhairyaLGandhi DhairyaLGandhi Jul 17, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add what to nnlib, sorry? This package is specific to GPU functionality.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

exactly these methods, with AbstractArray arguments, i.e. fallback on conv

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Umm, we probably want to retain the cpu kernels anyway. Without explicitly having and launching Julia with many threads, grouped convolutions would scale with the number of groups.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this would be true for any implementation, specialized or not

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

julia> x′ = rand(Float32, 28, 28, 4, 2);

julia> w′ = rand(Float32, 3, 3, 4, 30);

julia> cdims = DenseConvDims(x′, w′, groups = 4)

julia> @btime conv($x′, $w′, $cdims);
 362.792 μs (86 allocations: 736.36 KiB) # -t1
 236.368 μs (94 allocations: 831.89 KiB) # -t2
 232.137 μs (94 allocations: 831.89 KiB) # -t4

julia> @btime depthwiseconv($x′, $(permutedims(w′, (1,2,4,3))));
 348.914 μs (42 allocations: 731.03 KiB) # -t1
 156.558 μs (47 allocations: 826.53 KiB) # -t2
 161.059 μs (47 allocations: 826.53 KiB) # -t4

This is with https://github.com/DhairyaLGandhi/NNlib.jl#dg/g2 which has a couple of fixes pending a PR.

Copy link
Member

@ToucheSir ToucheSir left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Looks reasonable to me, just needs a couple tests in https://github.com/FluxML/NNlibCUDA.jl/blob/master/test/conv.jl (I know the implementation is technically covered indirectly now, but there's no guarantee these methods will forward to the conv ones forever).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants