Merge pull request #167 from JuliaGPU/tb/curand

Use an AbstractRNG to control dispatch of rand!, remove rand

Author: maleadt

.gitlab-ci.yml

@@ -7,7 +7,6 @@ stages:
 
 include:
   - 'https://raw.githubusercontent.com/JuliaGPU/gitlab-ci/master/templates/v0/common.yml'
-  - 'https://raw.githubusercontent.com/JuliaGPU/gitlab-ci/master/templates/v0/test_v0.7.yml'
   - 'https://raw.githubusercontent.com/JuliaGPU/gitlab-ci/master/templates/v0/test_v1.0.yml'
   - 'https://raw.githubusercontent.com/JuliaGPU/gitlab-ci/master/templates/v0/test_dev.yml'
   - 'https://raw.githubusercontent.com/JuliaGPU/gitlab-ci/master/templates/v0/postprocess_coverage.yml'

.travis.yml

@@ -5,7 +5,6 @@ os:
   - osx
 dist: trusty
 julia:
-  - 0.7
   - 1.0
   - nightly
 matrix:

REQUIRE

@@ -1,4 +1,4 @@
-julia 0.7-alpha
+julia 1.0
 StaticArrays
 FFTW
 FillArrays

appveyor.yml

@@ -1,6 +1,5 @@
 environment:
   matrix:
-  - julia_version: 0.7
   - julia_version: 1.0
   - julia_version: latest
 

src/random.jl

@@ -1,3 +1,8 @@
+## device interface
+
+# hybrid Tausworthe and Linear Congruent generator from
+# https://developer.nvidia.com/gpugems/GPUGems3/gpugems3_ch37.html
+
 function TausStep(z::Unsigned, S1::Integer, S2::Integer, S3::Integer, M::Unsigned)
     b = (((z << S1) ⊻ z) >> S2)
     return (((z & M) << S3) ⊻ b)
@@ -8,7 +13,6 @@ LCGStep(z::Unsigned, A::Unsigned, C::Unsigned) = A * z + C
 make_rand_num(::Type{Float64}, tmp) = 2.3283064365387e-10 * Float64(tmp)
 make_rand_num(::Type{Float32}, tmp) = 2.3283064f-10 * Float32(tmp)
 
-
 function next_rand(::Type{FT}, state::NTuple{4, T}) where {FT, T <: Unsigned}
     state = (
         TausStep(state[1], Cint(13), Cint(19), Cint(12), T(4294967294)),
@@ -42,22 +46,31 @@ function gpu_rand(::Type{T}, state, randstate::AbstractVector{NTuple{4, UInt32}}
     return to_number_range(f, T)
 end
 
-let rand_state_dict = Dict()
-    global cached_state, clear_cache
-    clear_cache() = (empty!(rand_state_dict); return)
-    function cached_state(x)
-        dev = GPUArrays.device(x)
-        get!(rand_state_dict, dev) do
-            N = GPUArrays.threads(dev)
-            res = similar(x, NTuple{4, UInt32}, N)
-            copyto!(res, [ntuple(i-> rand(UInt32), 4) for i=1:N])
-            res
-        end
+
+## host interface
+
+struct RNG <: AbstractRNG
+    state::GPUArray{NTuple{4,UInt32},1}
+
+    function RNG(A::GPUArray)
+        dev = GPUArrays.device(A)
+        N = GPUArrays.threads(dev)
+        state = similar(A, NTuple{4, UInt32}, N)
+        copyto!(state, [ntuple(i-> rand(UInt32), 4) for i=1:N])
+        new(state)
+    end
+end
+
+const GLOBAL_RNGS = Dict()
+function global_rng(A::GPUArray)
+    dev = GPUArrays.device(A)
+    get!(GLOBAL_RNGS, dev) do
+        RNG(A)
     end
 end
-function Random.rand!(A::GPUArray{T}) where T <: Number
-    rstates = cached_state(A)
-    gpu_call(A, (rstates, A,)) do state, randstates, a
+
+function Random.rand!(rng::RNG, A::GPUArray{T}) where T <: Number
+    gpu_call(A, (rng.state, A,)) do state, randstates, a
         idx = linear_index(state)
         idx > length(a) && return
         @inbounds a[idx] = gpu_rand(T, state, randstates)
@@ -66,14 +79,4 @@ function Random.rand!(A::GPUArray{T}) where T <: Number
     A
 end
 
-Random.rand(X::Type{<: GPUArray}, i::Integer...) = rand(X, Float32, i...)
-Random.rand(X::Type{<: GPUArray}, size::NTuple{N, Int}) where N = rand(X, Float32, size...)
-Random.rand(X::Type{<: GPUArray{T}}, i::Integer...) where T = rand(X, T, i...)
-Random.rand(X::Type{<: GPUArray{T}}, size::NTuple{N, Int}) where {T, N} = rand(X, T, size...)
-Random.rand(X::Type{<: GPUArray{T, N}}, size::NTuple{N, Integer}) where {T, N} = rand(X, T, size...)
-Random.rand(X::Type{<: GPUArray{T, N}}, size::NTuple{N, Int}) where {T, N} = rand(X, T, size...)
-
-function Random.rand(X::Type{<: GPUArray}, ::Type{ET}, size::Integer...) where ET
-    A = similar(X, ET, size)
-    rand!(A)
-end
+Random.rand!(A::GPUArray) = rand!(global_rng(A), A)

src/testsuite/indexing.jl

@@ -13,7 +13,8 @@ function test_indexing(AT)
             end
             @testset "multi dim, sliced setindex" begin
                 x = fill(AT{T}, T(0), (10, 10, 10, 10))
-                y = rand(AT{T}, 5, 5, 10, 10)
+                y = AT{T}(5, 5, 10, 10)
+                rand!(y)
                 x[2:6, 2:6, :, :] = y
                 x[2:6, 2:6, :, :] == y
            end

src/testsuite/random.jl

@@ -1,17 +1,10 @@
 function test_random(AT)
     @testset "Random" begin
         @testset "rand" begin  # uniform
-            for T in (Float32, Float64, Int64, Int32)
-                @test length(rand(AT{T,1}, (4,))) == 4
-                @test length(rand(AT{T}, (4,))) == 4
-                @test length(rand(AT{T}, 4)) == 4
-                @test eltype(rand(AT, 4)) == Float32
-                @test length(rand(AT, T, 4)) == 4
-                @test length(rand(AT{T,2}, (4,5))) == 20
-                @test length(rand(AT, T, 4, 5)) == 20
-                A = rand(AT{T,2}, (2,2))
+            for T in (Float32, Float64, Int64, Int32), d in (2, (2,2))
+                A = AT{T}(d)
                 B = copy(A)
-                @test all(A .== B)
+                rand!(A)
                 rand!(B)
                 @test !any(A .== B)
             end