Skip to content

Fixes for CUDAapi and CuArrays #14

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.

Sign up for GitHub

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

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

Fixes for CUDAapi and CuArrays #14

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9056420
Fixes for latest CUDAapi
maleadt Jan 11, 2018
48d992e
Convert objects for ccall to pointers using unsafe_convert.
maleadt Jan 11, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion REQUIRE
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
julia 0.6
NNlib
CuArrays
CUDAapi
CUDAapi 0.3.0
23 changes: 12 additions & 11 deletions src/batchnorm.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,14 +4,15 @@
mutable struct BND
ptr::Ptr{Void}
end
Base.unsafe_convert(::Type{Cptr}, bnd::BND) = bnd.ptr


function BND(xtd::TD, mode::Cuint)
d = Ref{Cptr}(0)
@cuda(cudnn, cudnnCreateTensorDescriptor, (Ptr{Cptr},), d)
@cuda(cudnn, cudnnDeriveBNTensorDescriptor,
(Ref{Void}, Cptr, Cuint),
d[], xtd.ptr, mode)
d[], xtd, mode)
return BND(d[])
end

Expand Down Expand Up @@ -63,10 +64,10 @@ function batchnorm_train!(y::CuArray{T,4}, x::CuArray{T,4}, s::BatchNormState;
Cptr, Cptr, Cptr, Cdouble,
Cptr, Cptr, Cdouble,
Cptr, Cptr),
handle, mode, Ref(T(alpha)), Ref(T(beta)), xtd.ptr, x.ptr, ytd.ptr, y.ptr,
bnScaleBiasMeanVarDesc.ptr, s.bnScale.ptr, s.bnBias.ptr, exponentialAverageFactor,
s.resultRunningMean.ptr, s.resultRunningVariance.ptr, epsilon,
s.resultSaveMean.ptr, s.resultSaveInvVariance.ptr)
handle, mode, Ref(T(alpha)), Ref(T(beta)), xtd, x, ytd, y,
bnScaleBiasMeanVarDesc, s.bnScale, s.bnBias, exponentialAverageFactor,
s.resultRunningMean, s.resultRunningVariance, epsilon,
s.resultSaveMean, s.resultSaveInvVariance)

end

Expand All @@ -92,9 +93,9 @@ function batchnorm_infer!(y::CuArray{T,4}, x::CuArray{T,4}, s::BatchNormState;
(Cptr, UInt32, Cptr, Cptr, Cptr, Cptr, Cptr, Cptr,
Cptr, Cptr, Cptr,
Cptr, Cptr, Cdouble),
handle, mode, Ref(T(alpha)), Ref(T(beta)), xtd.ptr, x.ptr, ytd.ptr, y.ptr,
bnScaleBiasMeanVarDesc.ptr, s.bnScale.ptr, s.bnBias.ptr,
estimatedMean.ptr, estimatedVariance.ptr, epsilon)
handle, mode, Ref(T(alpha)), Ref(T(beta)), xtd, x, ytd, y,
bnScaleBiasMeanVarDesc, s.bnScale, s.bnBias,
estimatedMean, estimatedVariance, epsilon)
end

function batchnorm_infer(x::CuArray{T,4}, s::BatchNormState; opts...) where T
Expand Down Expand Up @@ -126,9 +127,9 @@ function batchnorm_grad!(dx::CuArray{T,4}, x::CuArray{T,4}, dy::CuArray{T,4}, s:
Cptr, Cptr, Cptr, Cptr,
Cdouble, Cptr, Cptr),
handle, mode, Ref(T(alpha_data)), Ref(T(beta_data)), Ref(T(alpha_param)), Ref(T(beta_param)),
xtd.ptr, x.ptr, dytd.ptr, dy.ptr, dxtd.ptr, dx.ptr,
bnScaleBiasMeanVarDesc.ptr, s.bnScale.ptr, resultBnScaleDiff.ptr, resultBnBiasDiff.ptr,
epsilon, savedMean.ptr, savedInvVariance.ptr)
xtd, x, dytd, dy, dxtd, dx,
bnScaleBiasMeanVarDesc, s.bnScale, resultBnScaleDiff, resultBnBiasDiff,
epsilon, savedMean, savedInvVariance)
end


Expand Down
26 changes: 13 additions & 13 deletions src/conv.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,8 +6,8 @@ function conv2d!{T}(y::CuArray{T}, x::CuArray{T}, w::CuArray{T};
alpha=1, beta=0, o...) # padding=0, stride=1, upscale=1, mode=0
@cuda(cudnn, cudnnConvolutionForward,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,UInt32,Cptr,Csize_t,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),TD(x),x.ptr,FD(w),w.ptr,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(y),y.ptr)
handle,Ref(T(alpha)),TD(x),x,FD(w),w,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(y),y)
return y
end

Expand All @@ -29,18 +29,18 @@ function conv2_grad_x!{T}(dx::CuArray{T}, x::CuArray{T}, w::CuArray{T}, dy::CuAr
if cudnnVersion >= 4000
@cuda(cudnn,cudnnConvolutionBackwardData,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,UInt32,Cptr,Csize_t,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),FD(w),w.ptr,TD(dy),dy.ptr,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(dx),dx.ptr)
handle,Ref(T(alpha)),FD(w),w,TD(dy),dy,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(dx),dx)
elseif cudnnVersion >= 3000
@cuda(cudnn,cudnnConvolutionBackwardData_v3,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,UInt32,Cptr,Csize_t,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),FD(w),w.ptr,TD(dy),dy.ptr,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(dx),dx.ptr)
handle,Ref(T(alpha)),FD(w),w,TD(dy),dy,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),TD(dx),dx)
else
@cuda(cudnn,cudnnConvolutionBackwardData,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),FD(w),w.ptr,TD(dy),dy.ptr,CD(w,x;o...),
Ref(T(beta)),TD(dx),dx.ptr)
handle,Ref(T(alpha)),FD(w),w,TD(dy),dy,CD(w,x;o...),
Ref(T(beta)),TD(dx),dx)
end
return dx
end
Expand All @@ -64,17 +64,17 @@ function conv2d_grad_w!{T}(dw::CuArray{T}, x::CuArray{T}, w::CuArray{T}, dy::CuA
if cudnnVersion >= 4000
@cuda(cudnn,cudnnConvolutionBackwardFilter,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,UInt32,Cptr,Csize_t,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),TD(x),x.ptr,TD(dy),dy.ptr,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),FD(dw),dw.ptr)
handle,Ref(T(alpha)),TD(x),x,TD(dy),dy,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),FD(dw),dw)
elseif cudnnVersion >= 3000
@cuda(cudnn,cudnnConvolutionBackwardFilter_v3,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr,UInt32,Cptr,Csize_t,Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),TD(x),x.ptr,TD(dy),dy.ptr,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),FD(dw),dw.ptr)
handle,Ref(T(alpha)),TD(x),x,TD(dy),dy,CD(w,x;o...),algo,workSpace,
workSpaceSizeInBytes,Ref(T(beta)),FD(dw),dw)
else
@cuda(cudnn,cudnnConvolutionBackwardFilter,
(Cptr,Ptr{T},Cptr,Ptr{T},Cptr,Ptr{T},Cptr, Ptr{T},Cptr,Ptr{T}),
handle,Ref(T(alpha)),TD(x),x.ptr,TD(dy),dy.ptr,CD(w,x;o...),Ref(T(beta)),FD(dw),dw.ptr)
handle,Ref(T(alpha)),TD(x),x,TD(dy),dy,CD(w,x;o...),Ref(T(beta)),FD(dw),dw)
end
return dw
end
Expand Down
29 changes: 9 additions & 20 deletions src/descriptors.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,3 @@

mutable struct TD; ptr
function TD(a::CuArray)
d = Cptr[0]
Expand All @@ -10,7 +9,7 @@ mutable struct TD; ptr
(Cptr,UInt32,Cint,Ptr{Cint},Ptr{Cint}),
d[1], DT(a), n, sz, st)
td = new(d[1])
finalizer(td, x->@cuda(cudnn,cudnnDestroyTensorDescriptor,(Cptr,),x.ptr))
finalizer(td, x->@cuda(cudnn,cudnnDestroyTensorDescriptor,(Cptr,),x))
return td
end
end
Expand All @@ -36,7 +35,7 @@ mutable struct FD; ptr
d[1], DT(a), n, sz)
end
fd = new(d[1])
finalizer(fd, x->@cuda(cudnn,cudnnDestroyFilterDescriptor,(Cptr,),x.ptr))
finalizer(fd, x->@cuda(cudnn,cudnnDestroyFilterDescriptor,(Cptr,),x))
return fd
end
end
Expand All @@ -61,7 +60,7 @@ mutable struct CD; ptr
d[1],nd,cdsize(padding,nd),cdsize(stride,nd),cdsize(upscale,nd),mode)
end
cd = new(d[1])
finalizer(cd, x->@cuda(cudnn,cudnnDestroyConvolutionDescriptor,(Cptr,),x.ptr))
finalizer(cd, x->@cuda(cudnn,cudnnDestroyConvolutionDescriptor,(Cptr,),x))
return cd
end
end
Expand All @@ -86,12 +85,11 @@ mutable struct PD; ptr
d[1],mode,nd,cdsize(window,nd),cdsize(padding,nd),cdsize(stride,nd))
end
pd = new(d[1])
finalizer(pd, x->@cuda(cudnn,cudnnDestroyPoolingDescriptor,(Cptr,),x.ptr))
finalizer(pd, x->@cuda(cudnn,cudnnDestroyPoolingDescriptor,(Cptr,),x))
return pd
end
end


# mutable struct RNN_D; ptr
# function RNN_D(h_size::Int, num_layers::Int; handle=cudnnhandle())
# d = Cptr[0]
Expand All @@ -101,21 +99,12 @@ end


# rnn_d = new(d[1])
# finalizer(rnn_d, x->@cuda(cudnn,cudnnDestroyRNNDescriptor,(Cptr,),x.ptr))
# finalizer(rnn_d, x->@cuda(cudnn,cudnnDestroyRNNDescriptor,(Cptr,),x))
# return rnn_d
# end
# end









import Base: unsafe_convert
unsafe_convert(::Type{Cptr}, td::TD)=td.ptr
unsafe_convert(::Type{Cptr}, fd::FD)=fd.ptr
unsafe_convert(::Type{Cptr}, cd::CD)=cd.ptr
unsafe_convert(::Type{Cptr}, pd::PD)=pd.ptr
Base.unsafe_convert(::Type{Cptr}, td::TD)=td.ptr
Base.unsafe_convert(::Type{Cptr}, fd::FD)=fd.ptr
Base.unsafe_convert(::Type{Cptr}, cd::CD)=cd.ptr
Base.unsafe_convert(::Type{Cptr}, pd::PD)=pd.ptr
12 changes: 6 additions & 6 deletions src/dropout.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,10 +23,10 @@ function DOD(dropout::Float32; handle=cudnnhandle(), states=C_NULL,
# should we return it in DOD?
@cuda(cudnn, cudnnSetDropoutDescriptor,
(Ref{Cptr}, Cptr, Cfloat, Cptr, Cuint, Culong),
d, handle, dropout, states.ptr, statesSizeInBytes[], seed)
d, handle, dropout, states, statesSizeInBytes[], seed)
dod = DOD(d[], states)
finalizer(dod, x->@cuda(cudnn, cudnnDestroyDropoutDescriptor, (Cptr,), x.ptr))
# finalizer(dod, x -> cudnnDestroyDropoutDescriptor(x.ptr))
finalizer(dod, x->@cuda(cudnn, cudnnDestroyDropoutDescriptor, (Cptr,), x))
# finalizer(dod, x -> cudnnDestroyDropoutDescriptor(x))
return dod
end

Expand All @@ -48,7 +48,7 @@ end
function get_reserve_space_size(x::CuArray{T}) where T
td = TD(x)
sz = Ref{Csize_t}(0)
@cuda(cudnn, cudnnDropoutGetReserveSpaceSize, (Cptr, Ref{Csize_t}), td.ptr, sz)
@cuda(cudnn, cudnnDropoutGetReserveSpaceSize, (Cptr, Ref{Csize_t}), td, sz)
return sz[]
end

Expand All @@ -75,7 +75,7 @@ function dropout_forward!(y::CuArray{T}, x::CuArray{T}, dropout::Float64;
@cuda(cudnn, cudnnDropoutForward,
# (Cptr, Cptr, Cptr, Cptr, Cptr, Cptr, Cptr, Csize_t),
(Cptr, Cptr, Cptr, Cptr, Cptr, Cptr, Cptr, Csize_t),
handle, dod.ptr, xtd.ptr, x.ptr, ytd.ptr, y.ptr, rs.ptr, rs_sz)
handle, dod, xtd, x, ytd, y, rs, rs_sz)

end

Expand All @@ -99,6 +99,6 @@ function main()
reserveSizeInBytes = Cint(get_reserve_space_size(x))


cudnnDropoutForward(handle, dropoutDesc.ptr, xdesc, x.ptr, ydesc, y.ptr, reserveSpace,
cudnnDropoutForward(handle, dropoutDesc, xdesc, x, ydesc, y, reserveSpace,
reserveSizeInBytes)
end
3 changes: 2 additions & 1 deletion src/init.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,8 @@


const toolkit = CUDAapi.find_toolkit()
const libcudnn = CUDAapi.find_library("cudnn", toolkit)
const libcudnn = CUDAapi.find_cuda_library("cudnn", toolkit)
libcudnn == nothing && error("Could not find libcudnn")
const Cptr = Ptr{Void}

macro cuda(lib,fun,x...) # give an error if library missing, or if error code!=0
Expand Down
3 changes: 2 additions & 1 deletion src/libcudnn.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,8 @@
# Automatically generated using Clang.jl wrap_c, version 0.0.0

const toolkit = CUDAapi.find_toolkit()
const libcudnn = CUDAapi.find_library("cudnn", toolkit)
const libcudnn = CUDAapi.find_cuda_library("cudnn", toolkit)
libcudnn == nothing && error("Could not find libcudnn")


function cudaDeviceReset()
Expand Down
8 changes: 4 additions & 4 deletions src/pool.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@ function pool{T}(x::CuArray{T}; handle=cudnnhandle(), alpha=1,
beta = 0
@cuda(cudnn, cudnnPoolingForward,
(Cptr, Cptr, Ptr{T}, Cptr, Ptr{T}, Ptr{T}, Cptr,Ptr{T}),
handle, PD(x;o...), Ref(T(alpha)), TD(x), x.ptr, Ref(T(beta)), TD(y), y.ptr)
handle, PD(x;o...), Ref(T(alpha)), TD(x), x, Ref(T(beta)), TD(y), y)
return y
end

Expand All @@ -17,8 +17,8 @@ function pool_grad{T}(x::CuArray{T}, y::CuArray{T}, dy::CuArray{T};
beta = 0
@cuda(cudnn,cudnnPoolingBackward,
(Cptr, Cptr, Ptr{T}, Cptr, Ptr{T}, Cptr, Ptr{T}, Cptr, Ptr{T}, Ptr{T}, Cptr, Ptr{T}),
handle, PD(x; mode=mode, o...), Ref(T(alpha)), TD(y), y.ptr,
TD(dy), dy.ptr, TD(x), x.ptr, Ref(T(beta)), TD(dx), dx.ptr)
handle, PD(x; mode=mode, o...), Ref(T(alpha)), TD(y), y,
TD(dy), dy, TD(x), x, Ref(T(beta)), TD(dx), dx)
return dx
end

Expand All @@ -32,5 +32,5 @@ end
function unpool_grad(dy; window=2, alpha=1, o...) # padding=0, stride=window, mode=0,
# maxpoolingNanOpt=0
w = prod(psize(window,dy))
pool(dy.ptr; o..., window=window, mode=1, alpha=1/alpha) * w
pool(dy; o..., window=window, mode=1, alpha=1/alpha) * w
end
8 changes: 4 additions & 4 deletions src/softmax.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,7 @@ function softmax4d!(y::CuArray{T}, x::CuArray{T};
alpha=1.0, beta=0.0) where T
@cuda(cudnn, cudnnSoftmaxForward,
(Cptr, Cuint, Cuint, Ptr{T}, Cptr, Ptr{T}, Ptr{T}, Cptr, Ptr{T}),
handle, algorithm, mode, Ref(T(alpha)), TD(x), x.ptr, Ref(T(beta)), TD(y), y.ptr)
handle, algorithm, mode, Ref(T(alpha)), TD(x), x, Ref(T(beta)), TD(y), y)
return y
end

Expand All @@ -24,9 +24,9 @@ function softmax4d_grad!(dx::CuArray{T}, y::CuArray{T}, dy::CuArray{T};
Cptr, Ptr{T},
Ptr{T}, Cptr, Ptr{T}),
handle, algorithm, mode,
Ref(T(alpha)), TD(y), y.ptr,
TD(dy), dy.ptr,
Ref(T(beta)), TD(dx), dx.ptr)
Ref(T(alpha)), TD(y), y,
TD(dy), dy,
Ref(T(beta)), TD(dx), dx)
return dx
end

Expand Down