rm float64 methods

Author: MikeInnes

src/nnpack/interface.jl

@@ -6,27 +6,15 @@ function check_support(x, k, pad, stride, dilation = 1)
     return pad_, stride_, fallback
 end
 
-function softmax!(x::A) where A<:AbstractVecOrMat{Float64}
-    x = Float32.(x)
-    softmax!(x)
-end
-
 softmax!(x::A) where A<:AbstractVecOrMat{Float32} =
     nnp_softmax_output(x, x)
 
-softmax!(y::A, x::A) where A<:AbstractVecOrMat{Float64} = softmax!(Float32.(y), Float32.(x))
-
 softmax!(y::A, x::A) where A<:AbstractVecOrMat{Float32} =
     nnp_softmax_output(x, y)
 
-softmax(x::A) where A<:AbstractVecOrMat{Float64} = softmax(Float32.(x))
-
 softmax(x::A) where A<:AbstractVecOrMat{Float32} =
     nnp_softmax_output(x, similar(x))
 
-maxpool(x::A, k; pad = map(_->0,k), stride = k) where A<:AbstractArray{Float64, 4} =
-    maxpool(Float32.(x), k, pad = pad, stride = stride)
-
 function maxpool(x::A, k; pad = map(_->0,k), stride = k) where A<:AbstractArray{Float32, 4}
     pad_, stride_, fallback = check_support(x, k, pad, stride)
     if fallback
@@ -36,15 +24,9 @@ function maxpool(x::A, k; pad = map(_->0,k), stride = k) where A<:AbstractArray{
     end
 end
 
-maxpool!(y::A, x::A, k; pad = map(_->0,k), stride = k) where A<:AbstractArray{Float64, 4} =
-    maxpool!(Float32.(y), Float32.(x), k, pad = pad, stride = stride)
-
 maxpool!(y::A, x::A, k; pad = map(_->0,k), stride = k) where A<:AbstractArray{Float32, 4} =
     nnp_max_pooling_output(x, y, k, padding = expand(Val{length(k)}, pad), stride = expand(Val{length(k)}, stride))
 
-conv(x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where A<:AbstractArray{Float64, 4} =
-    conv(Float32.(x), Float32.(w), pad = pad, stride = stride, dilation = dilation, algo = algo)
-
 function conv(x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where A<:AbstractArray{Float32, 4}
     pad_, stride_, fallback = check_support(x, (size(w, 1), size(w, 2)), pad, stride, dilation)
     y = similar(x, cdims(size(x), dilation_dims(w, dilation), pad_, stride_))
@@ -55,9 +37,6 @@ function conv(x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) w
     end
 end
 
-conv(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float64, 4}, A2<:AbstractArray{Float64, 1}} =
-    conv(Float32.(x), Float32.(w), Float32.(b), pad = pad, stride = stride, dilation = dilation, algo = algo)
-
 function conv(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float32, 4}, A2<:AbstractArray{Float32, 1}}
     pad_, stride_, fallback = check_support(x, (size(w, 1), size(w, 2)), pad, stride, dilation)
     y = similar(x, cdims(size(x), dilation_dims(w, dilation), pad_, stride_))
@@ -68,9 +47,6 @@ function conv(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UIn
     end
 end
 
-crosscor(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float64, 4}, A2<:AbstractArray{Float64, 1}} =
-    crosscor(Float32.(x), Float32.(w), Float32.(b), pad = pad, stride = stride, dilation = dilation, algo = algo)
-
 function crosscor(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float32, 4}, A2<:AbstractArray{Float32, 1}}
     pad_, stride_, fallback = check_support(x, (size(w, 1), size(w, 2)), pad, stride, dilation)
     y = similar(x, cdims(size(x), dilation_dims(w, dilation), pad_, stride_))
@@ -81,19 +57,13 @@ function crosscor(x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo =
     end
 end
 
-conv!(y::A1, x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where {A1<:AbstractArray{Float64, 4}, A2<:AbstractArray{Float64, 1}} =
-    conv!(Float32.(y), Float32.(x), Float32.(w), Float32.(b), pad = pad, stride = stride, dilation = dilation, algo = algo, flipkernel = flipkernel)
-
 function conv!(y::A1, x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where {A1<:AbstractArray{Float32, 4}, A2<:AbstractArray{Float32, 1}}
     if flipkernel == 0
         w = reverse(reverse(w, dims=1), dims=2)
     end
     nnp_convolution_output(y, x, w, b, algo = algo, padding = pad, stride = stride)
 end
 
-crosscor!(y::A1, x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float64, 4}, A2<:AbstractArray{Float64, 1}} =
-    conv!(Float32.(y), Float32.(x), Float32.(w), Float32.(b), pad = pad, stride = stride, dilation = dilation, algo = algo, flipkernel = 1)
-
 crosscor!(y::A1, x::A1, w::A1, b::A2; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where {A1<:AbstractArray{Float32, 4}, A2<:AbstractArray{Float32, 1}} =
     conv!(y, x, w, b, pad = pad, stride = stride, dilation = dilation, algo = algo, flipkernel = 1)
 
@@ -109,17 +79,11 @@ function ∇conv_data(dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo
     end
 end
 
-∇conv_data!(dx::A, dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where A<:AbstractArray{Float64, 4} =
-    ∇conv_data!(Float32.(dx), Float32.(dy), Float32.(x), Float32.(w), pad = pad, stride = stride, dilation = dilation, algo = algo, flipkernel = flipkernel)
-
 function ∇conv_data!(dx::A, dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where A<:AbstractArray{Float32, 4}
     flipkernel == 0 && (w = reverse(reverse(w, dims=1), dims=2))
     nnp_convolution_input_gradient(dx, x, dy, w, padding = pad, stride = stride, algo = algo)
 end
 
-∇conv_filter(dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where A<:AbstractArray{Float64, 4} =
-    ∇conv_filter(Float32.(dy), Float32.(x), Float32.(w), pad = pad, stride = stride, dilation = dilation, algo = algo)
-
 function ∇conv_filter(dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0)) where A<:AbstractArray{Float32, 4}
     pad_, stride_, fallback = check_support(x, (size(w, 1), size(w, 2)), pad, stride, dilation)
     if fallback
@@ -129,9 +93,6 @@ function ∇conv_filter(dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, al
     end
 end
 
-∇conv_filter!(dw::A, dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where A<:AbstractArray{Float64, 4} =
-    ∇conv_filter!(Float32.(dw), Float32.(dy), Float32.(x), Float32.(w), pad = pad, stride = stride, dilation = dilation, algo = algo, flipkernel = flipkernel)
-
 function ∇conv_filter!(dw::A, dy::A, x::A, w::A; pad = 0, stride = 1, dilation = 1, algo = UInt32(0), flipkernel = 0) where A<:AbstractArray{Float32, 4}
     flipkernel == 0 && (w = reverse(reverse(w, dims=1), dims=2))
     dw .= nnp_convolution_kernel_gradient(dw, x, dy, w, padding = pad, stride = stride, algo = algo)