Remove NNlib from tests (#636)

Author: pxl-th

test/Project.toml

@@ -10,7 +10,6 @@ InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
-NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
 PrettyTables = "08abe8d2-0d0c-5749-adfa-8a2ac140af0d"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

test/dnn/activations.jl

@@ -1,35 +1,23 @@
-@testset "NNlib comparison" begin
+@testset "Activation functions" begin
     for (T, atol) in ((Float16, 1f-3), (Float32, 1f-6))
         x, dy = randn(T, 16), randn(T, 16)
         xd, dyd = ROCArray(x), ROCArray(dy)
 
-        y = NNlib.relu(x)
         yd = MIOpen.relu(xd)
-        @test all(isapprox.(Array(yd), y; atol))
 
-        y = NNlib.leakyrelu(x, 0.1)
         yd = MIOpen.leakyrelu(xd, 0.1)
-        @test all(isapprox.(Array(yd), y; atol))
 
-        y = NNlib.softplus(x)
         yd = MIOpen.softrelu(xd)
-        @test all(isapprox.(Array(yd), y; atol))
 
-        y = NNlib.relu6(x)
         yd = MIOpen.clippedrelu(xd, 6)
-        @test all(isapprox.(Array(yd), y; atol))
 
-        y = NNlib.elu(x, 0.1)
         yd = MIOpen.elu(xd, 0.1)
-        @test all(isapprox.(Array(yd), y; atol))
 
         y = abs.(x)
         yd = MIOpen.abs(xd)
         @test all(isapprox.(Array(yd), y; atol))
 
-        y = NNlib.sigmoid(x)
         yd = MIOpen.sigmoid(xd)
-        @test all(isapprox.(Array(yd), y; atol))
 
         y = tanh.(x)
         yd = MIOpen.tanh(xd)

test/dnn/conv.jl

@@ -37,41 +37,29 @@ end
     wh2 = rand(Float32, 3, 4, 3, 16)
     x, w1, w2 = ROCArray.((xh, wh1, wh2))
 
-    yh = NNlib.conv(xh, wh1; pad=(0, 0), stride=(1, 1), dilation=(1, 1), flipped=true)
     y = MIOpen.convolution(x, w1; padding=(0, 0), stride=(1, 1), dilation=(1, 1), groups=1)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (9, 9, 16, 10)
 
-    yh = NNlib.conv(xh, wh1; pad=(2, 2), stride=(2, 2), dilation=(1, 1), flipped=true)
     y = MIOpen.convolution(x, w1; padding=(2, 2), stride=(2, 2), dilation=(1, 1), groups=1)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (7, 7, 16, 10)
 
-    yh = NNlib.conv(xh, wh2; pad=(2, 3), stride=(1, 2), dilation=(1, 1), flipped=true)
     y = MIOpen.convolution(x, w2; padding=(2, 3), stride=(1, 2), dilation=(1, 1), groups=1)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (12, 7, 16, 10)
 
-    yh = NNlib.conv(xh, wh1; pad=(2, 3), stride=(1, 2), dilation=(2, 2), flipped=true)
     y = MIOpen.convolution(x, w1; padding=(2, 3), stride=(1, 2), dilation=(2, 2), groups=1)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (12, 7, 16, 10)
 
     # Depthwise convolution.
     wdh1 = rand(Float32, 2, 2, 1, 3)
     wd1 = ROCArray(wdh1)
-    yh = NNlib.depthwiseconv(xh, wdh1; pad=(0, 0), stride=(1, 1), dilation=(1, 1), flipped=true)
     y = MIOpen.convolution(x, wd1; padding=(0, 0), stride=(1, 1), dilation=(1, 1), groups=3)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (9, 9, 3, 10)
 
     # Grouped convolution.
     xh = ones(Float32, 10, 10, 4, 10)
     wdh2 = ones(Float32, 2, 2, 2, 4)
     x, wd2 = ROCArray.((xh, wdh2))
-    yh = NNlib.conv(xh, wdh2; pad=(0, 0), stride=(1, 1), dilation=(1, 1), groups=2, flipped=true)
     y = MIOpen.convolution(x, wd2; padding=(0, 0), stride=(1, 1), dilation=(1, 1), groups=2)
-    @test y ≈ ROCArray(yh)
     @test size(y) == (9, 9, 4, 10)
 end
 

test/dnn/miopen.jl

@@ -1,6 +1,5 @@
 @testset "MIOpen" begin
 
-using NNlib
 using AMDGPU.MIOpen
 
 @testset "Tensor descriptors" begin

test/dnn/pool.jl

@@ -1,44 +1,29 @@
-@testset "NNlib comparison" begin
+@testset "Maxpool" begin
     x = rand(Float32, 16, 16, 3, 2)
     xd = ROCArray(x)
 
     for k in (2, 3, 4), stride in (1, 2, 3), padding in (0, 1, 2)
-        pdims = NNlib.PoolDims(x, k; stride, padding)
         pkwargs = (; dims=(k, k), padding=(padding, padding), stride=(stride, stride))
 
         # Max pooling.
 
-        y = NNlib.maxpool(x, pdims)
         yd, workspace = MIOpen.maxpool(xd; pkwargs...)
         yd, workspace = MIOpen.maxpool!(yd, xd; pkwargs...)
-        @test Array(yd) ≈ y
-        @test Array(yd) ≈ y
 
-        dy = ones(Float32, size(y))
+        dy = ones(Float32, size(yd))
         dyd = ROCArray(dy)
-
-        dx = NNlib.∇maxpool(dy, y, x, pdims)
         dxd = MIOpen.∇maxpool(dyd, yd, xd; workspace, pkwargs...)
-        @test Array(dxd) ≈ dx
-
         dxd = MIOpen.∇maxpool!(dxd, dyd, yd, xd; workspace, pkwargs...)
-        @test Array(dxd) ≈ dx
 
         # Mean pooling.
 
-        y = NNlib.meanpool(x, pdims)
         yd, workspace = MIOpen.meanpool(xd; pkwargs...)
-        @test Array(yd) ≈ y
         yd, workspace = MIOpen.meanpool!(yd, xd; pkwargs...)
-        @test Array(yd) ≈ y
 
-        dy = ones(Float32, size(y))
+        dy = ones(Float32, size(yd))
         dyd = ROCArray(dy)
 
-        dx = NNlib.∇meanpool(dy, y, x, pdims)
         dxd = MIOpen.∇meanpool(dyd, yd, xd; workspace, pkwargs...)
-        @test Array(dxd) ≈ dx
         dxd = MIOpen.∇meanpool!(dxd, dyd, yd, xd; workspace, pkwargs...)
-        @test Array(dxd) ≈ dx
     end
 end

test/external/forwarddiff.jl

@@ -79,15 +79,3 @@ end
         @test ForwardDiff.gradient(_x -> sum(_x .^ p), x) ≈ p .* (x .^ (p - 1))
     end
 end
-
-#= FIXME
-@testset "Broadcast Fix" begin
-    if AMDGPU.libmiopen !== nothing
-        using NNlib
-
-        f(x) = logσ.(x)
-        ds = Dual.(rand(5),1)
-        @test f(ds) ≈ collect(f(ROCArray(ds)))
-    end
-end
-=#