Add image GPU loading in preprocimage

src/imfilter.jl

@@ -7,6 +7,10 @@ function imfilter_cpu(img, krn)
 end
 
 function imfilter_gpu(img, krn)
+  imfilter_gpu(img |> CuArray, krn)
+end
+
+function imfilter_gpu(img::CuArray, krn)
   # retrieve basic info
   N = ndims(img)
   T = eltype(img)
@@ -16,7 +20,7 @@ function imfilter_gpu(img, krn)
   padkrn  = padarray(krn, Fill(zero(T), ntuple(i->0, N), padsize))
 
   # perform ifft(fft(img) .* conj.(fft(krn)))
-  fftimg = img |> CuArray |> CUFFT.fft
+  fftimg = img |> CUFFT.fft
   fftkrn = padkrn |> CuArray |> CUFFT.fft
   result = (fftimg .* conj.(fftkrn)) |> CUFFT.ifft
 

src/iqsim.jl

@@ -226,7 +226,7 @@ function iqsim(trainimg::AbstractArray{T,N}, tilesize::Dims{N},
       rtile  = CartesianIndex(start):CartesianIndex(finish)
 
       # selected training image dataevent
-      TIdev = view(TI, rtile)
+      TIdev = view_kernel(TI, rtile)
 
       # boundary cut mask
       cutmask .= false

src/utils.jl

@@ -54,11 +54,32 @@ function activation(hard, tile)
   buff
 end
 
+function array_cpu(array)
+  array
+end
+
+function array_gpu(array)
+  array |> CuArray
+end
+
+const array_kernel = CUDA.functional() ? array_gpu : array_cpu
+
+function view_cpu(array, I)
+  view(array, I)
+end
+
+function view_gpu(array, I)
+  array[I] |> Array
+end
+
+const view_kernel = CUDA.functional() ? view_gpu : view_cpu
+
 function imagepreproc(trainimg, soft, geoconfig)
   padsize = geoconfig.padsize
 
   TI = Float64.(trainimg)
   replace!(TI, NaN => 0.)
+  TI_kernel = TI |> array_kernel
 
   SOFT = map(soft) do (aux, auxTI)
     prepend = ntuple(i->0, ndims(TI))
@@ -70,10 +91,12 @@ function imagepreproc(trainimg, soft, geoconfig)
     replace!(AUX, NaN => 0.)
     replace!(AUXTI, NaN => 0.)
 
-    AUX, AUXTI
+    AUXTI_kernel = AUXTI |> array_kernel
+
+    AUX, AUXTI_kernel
   end
 
-  TI, SOFT
+  TI_kernel, SOFT
 end
 
 function finddisabled(trainimg, geoconfig)