Fix the workaround for #4.

JuliaGPU · Mar 30, 2017 · 42bb5ac · 42bb5ac
1 parent 35d9f75
commit 42bb5ac
Show file tree

Hide file tree

Showing 3 changed files with 41 additions and 14 deletions.
diff --git a/examples/vadd.jl b/examples/vadd.jl
@@ -1,9 +1,12 @@
 using CUDAdrv, CUDAnative
 using Base.Test
 
-function kernel_vadd(a, b, c)
-    i = (blockIdx().x-1) * blockDim().x + threadIdx().x
-    c[i] = a[i] + b[i]
+const BLOCK_SIZE = 16
+
+function kernel_vadd(c)
+    shadow = @cuStaticSharedMem(Float32, (BLOCK_SIZE,))
+    tx = threadIdx().x
+    c[tx] = shadow[tx]
 
     return nothing
 end
@@ -16,12 +19,13 @@ a = round.(rand(Float32, dims) * 100)
 b = round.(rand(Float32, dims) * 100)
 
 d_a = CuArray(a)
+matrix = CuArray{Float32}((BLOCK_SIZE, BLOCK_SIZE))
 d_b = CuArray(b)
 d_c = similar(d_a)
 
 len = prod(dims)
-@cuda (1,len) kernel_vadd(d_a, d_b, d_c)
-c = Array(d_c)
-@test a+b ≈ c
+CUDAnative.@code_llvm @cuda (1,len) kernel_vadd(d_c)
+# c = Array(d_c)
+# @test a+b ≈ c
 
 destroy(ctx)
diff --git a/src/device/array.jl b/src/device/array.jl
@@ -31,14 +31,12 @@ Base.size(g::CuDeviceArray) = g.shape
 Base.length(g::CuDeviceArray) = prod(g.shape)
 
 @inline function Base.getindex{T}(A::CuDeviceArray{T}, index::Int)
-    # FIXME: disabled due to PTX assembler issue (see #4)
-    # @boundscheck checkbounds(A, index)
+    @boundscheck checkbounds(A, index)
     Base.pointerref(Base.unsafe_convert(Ptr{T}, A), index, 8)::T
 end
 
 @inline function Base.setindex!{T}(A::CuDeviceArray{T}, x, index::Int)
-    # FIXME: disabled due to PTX assembler issue (see #4)
-    # @boundscheck checkbounds(A, index)
+    @boundscheck checkbounds(A, index)
     Base.pointerset(Base.unsafe_convert(Ptr{T}, A), convert(T, x)::T, index, 8)
 end
 
@@ -50,15 +48,18 @@ Base.show{T,N}(io::IO, a::CuDeviceArray{T,N}) =
     print(io, "$(join(a.shape, '×')) device array at $(pointer(a))")
 
 
-## compatibility fixes
+## quirks
 
-# TODO: remove this hack as soon as immutables with heap references (such as BoundsError)
-#       can be stack-allocated
+# bounds checking is currently broken due to a PTX assembler issue (see #4)
+Base.checkbounds(::CuDeviceArray, I...) = nothing
+
+# replace boundserror-with-arguments to a non-allocating, argumentless version
+# TODO: can this be fixed by stack-allocating immutables with heap references?
 struct CuBoundsError <: Exception end
 @inline Base.throw_boundserror{T,N}(A::CuDeviceArray{T,N}, I) =
     (Base.@_noinline_meta; throw(CuBoundsError()))
 
-# TODO: same for SubArray, although it might be too complex to ever be non-allocating
+# idem
 function Base.unsafe_view{T}(A::CuDeviceArray{T,1}, I::Vararg{Base.ViewIndex,1})
     Base.@_inline_meta
     ptr = Base.unsafe_convert(Ptr{T}, A) + (I[1].start-1)*sizeof(T)

diff --git a/test/array.jl b/test/array.jl
@@ -44,6 +44,28 @@ end
 
 
 
+############################################################################################
+
+@testset "bounds checking" begin
+    @eval function array_oob_1d(array)
+        return array[1]
+    end
+
+    # NOTE: these tests verify that bounds checking is _disabled_ (see #4)
+
+    ir = sprint(io->CUDAnative.code_llvm(io, array_oob_1d, (CuDeviceArray{Int,1},)))
+    @test !contains(ir, "trap")
+
+    @eval function array_oob_2d(array)
+        return array[1, 1]
+    end
+
+    ir = sprint(io->CUDAnative.code_llvm(io, array_oob_2d, (CuDeviceArray{Int,2},)))
+    @test !contains(ir, "trap")
+end
+
+
+
 ############################################################################################
 
 @testset "views" begin