towards working tests

Project.toml

@@ -4,31 +4,21 @@ authors = ["Bartłomiej Gardas <[email protected]>", "Dariusz Kurzyk <dkur
 version = "1.0.0"
 
 [deps]
-BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
 Bits = "1654ce90-6ed3-553a-957f-9452c3a40996"
 CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
+DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
 Graphs = "86223c79-3864-5bf0-83f7-82e725a168b6"
 LabelledGraphs = "605abd48-4d17-4660-b914-d4df33194460"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
-LinearMaps = "7a12625a-238d-50fd-b39a-03d52299707e"
-MetaGraphs = "626554b9-1ddb-594c-aa3c-2596fe9399a5"
-SortingAlgorithms = "a2af1166-a08f-5f64-846c-94a0d3cef48c"
-SpinGlassEngine = "0563570f-ea1b-4080-8a64-041ac6565a4e"
 SpinGlassNetworks = "b7f6bd3e-55dc-4da6-96a9-ef9dbec6ac19"
 
 [compat]
-BenchmarkTools = "1.3"
 Bits = "0.2"
-CUDA = "3.12"
-Distributions = "0.25"
-DocStringExtensions = "0.8.6"
-Flux = "0.13"
-LabelledGraphs = "0.3"
-LightGraphs = "1.3"
-LinearMaps = "3.8"
-MetaGraphs = "0.6, 0.7"
-SortingAlgorithms = "1.0"
-julia = "1.8, 1.9"
+CUDA = "4"
+DocStringExtensions = "0.8"
+Graphs = "1.9"
+LabelledGraphs = "0.4"
+julia = "1.9, 1.10"
 
 [extras]
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"

docs/src/man/details.md

@@ -27,10 +27,10 @@ k = 2
 
 energies = CUDA.zeros(2^N);
 
-threadsPerBlock = 2^k
-blocksPerGrid = 2^(N-k)
+threads = 512
+blocks = cld(N, threads)
 
-@cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) SpinGlassExhaustive.kernel_qubo(cu_qubo, energies);
+@cuda blocks=blocks threads=threads SpinGlassExhaustive.kernel_qubo(cu_qubo, energies);
 
 states = sortperm(energies);
 energies[states];
@@ -49,10 +49,10 @@ k = 2
 
 energies = CUDA.zeros(2^N);
 
-threadsPerBlock = 2^k
-blocksPerGrid = 2^(N-k)
+threads = 512
+blocks = cld(N, threads)
 
-@cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) SpinGlassExhaustive.kernel(cu_graph, energies);
+@cuda blocks=blocks threads=threads SpinGlassExhaustive.kernel(cu_graph, energies);
 
 states = sortperm(energies);
 energies[states]
@@ -70,10 +70,10 @@ energies = CUDA.zeros(2^N);
 part_st = CUDA.zeros(2^(N-k));
 part_lst = CUDA.zeros(2^(N-k));
 
-threadsPerBlock = 2^k
-blocksPerGrid = 2^(N-k)
+threads = 512
+blocks = cld(N, threads)
 
-@cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) SpinGlassExhaustive.kernel_part(cu_graph, energies, part_lst, part_st);
+@cuda blocks=blocks threads=threads SpinGlassExhaustive.kernel_part(cu_graph, energies, part_lst, part_st);
 
 idx = sortperm(part_lst);
 states = part_st[idx];
@@ -109,16 +109,16 @@ lowest_d = CUDA.zeros(Float64, how_many*2);
 lowest_states_d = CUDA.zeros(Int64, how_many*2);
 
 k = 2
-threadsPerBlock = 2^k
-blocksPerGrid = 2^(chunk_size-k)
+threads = 512
+blocks = cld(chunk_size, threads)
 
 
 
 for i in 1:2^(N-chunk_size)
 
     idx = (i-1)*(2^chunk_size) + 1 
     
-    @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) SpinGlassExhaustive.kernel_bucket(J_dev, energies_d, idx)
+    @cuda blocks=blocks threads=threads SpinGlassExhaustive.kernel_bucket(J_dev, energies_d, idx)
 
     states_d = sortperm(energies_d)[1:how_many]
 

examples/experiment.jl

@@ -5,7 +5,7 @@ function exp_exhaustive_search()
   graph = generate_random_graph(N)
   cu_graph = graph |> cu 
 
-  ig = SpinGlassEngine.ising_graph(graph_to_dict(cu_graph))    
+  ig = ising_graph(graph_to_dict(cu_graph))    
 
   exhaustive_search(ig)
 end
@@ -15,7 +15,7 @@ function exp_bucket_exhaustive_search()
   graph = generate_random_graph(N)
   cu_graph = graph |> cu 
 
-  ig = SpinGlassEngine.ising_graph(graph_to_dict(cu_graph))    
+  ig = ising_graph(graph_to_dict(cu_graph))    
 
   exhaustive_search_bucket(ig)
 end
@@ -25,7 +25,7 @@ function exp_partial_exhaustive_search()
   graph = generate_random_graph(N)
   cu_graph = graph |> cu 
 
-  ig = SpinGlassEngine.ising_graph(graph_to_dict(cu_graph))    
+  ig = ising_graph(graph_to_dict(cu_graph))    
 
   partial_exhaustive_search(ig)
 end
@@ -44,7 +44,7 @@ function exp_qubo()
   threadsPerBlock::Int64 = 2^k
   blocksPerGrid::Int64 = 2^(N-k)
 
-  @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel_qubo(cu_qubo, energies)
+  @cuda blocks=blocks threads=threads kernel_qubo(cu_qubo, energies)
 
   # sort!(energies)
   # or
@@ -67,7 +67,7 @@ function exp_ising()
   threadsPerBlock::Int64 = 2^k
   blocksPerGrid::Int64 = 2^(N-k)
 
-  @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel(cu_graph, energies)
+  @cuda blocks=blocks threads=threads kernel(cu_graph, energies)
 
   sort!(energies)
 
@@ -91,7 +91,7 @@ function exp_ising_part()
   threadsPerBlock::Int64 = 2^k
   blocksPerGrid::Int64 = 2^(N-k)
 
-  @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel_part(cu_graph, energies, part_lst, part_st)
+  @cuda blocks=blocks threads=threads kernel_part(cu_graph, energies, part_lst, part_st)
 
   sort!(part_lst)
 

src/SpinGlassExhaustive.jl

@@ -8,6 +8,10 @@ module SpinGlassExhaustive
     using LabelledGraphs
     using Bits
     using LinearAlgebra
+    using DocStringExtensions
+    using CUDA
 
     include("naive.jl")
+    include("utils.jl")
+    include("ising.jl")
 end

src/ising.jl

@@ -1,14 +1,7 @@
-using LabelledGraphs
-using SpinGlassNetworks
-using Bits
-using DocStringExtensions
-
 export energy_qubo, energy, kernel, kernel_qubo, kernel_part, exhaustive_search, partial_exhaustive_search
 
-const IsingGraph = LabelledGraph{MetaGraph{Int64, Float64}, Int64}
-
 struct Spectrum
-    energies::Vector{Float64}
+    energies::AbstractVector
     states::Vector{Int}
 end
 
@@ -75,7 +68,9 @@ function kernel(graph, energies)
 
     F = energy(state_code, graph) |> Float32
 
-    @inbounds energies[state_code] = F
+    if state_code <= length(energies)
+        @inbounds energies[state_code] = F
+    end
 
     return
 end
@@ -148,22 +143,19 @@ Returns energies and states for provided graph by brute-forece alorithm based on
 function exhaustive_search(ig::IsingGraph)
     L = SpinGlassNetworks.nv(ig)
 
-    σ = CUDA.fill(Int32(-1), L, 2^L)
-    J = couplings(ig) + SpinGlassNetworks.Diagonal(biases(ig))
+    J = couplings(ig) + Diagonal(biases(ig))
     J_dev = CUDA.CuArray(J)
 
     energies = CUDA.zeros(L)
 
     N = size(J_dev,1)
-    k = 2
 
     energies = CUDA.zeros(2^N)
 
-    threadsPerBlock::Int64 = 2^k
-    blocksPerGrid::Int64 = 2^(N-k)
+    threads = 512
+    blocks = cld(N, threads)
 
-    @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel(J_dev, energies)
-
+    @cuda blocks=blocks threads=threads kernel(J_dev, energies)
 
     states = sortperm(energies)
 
@@ -181,7 +173,7 @@ function partial_exhaustive_search(ig::IsingGraph)
 
 
     σ = CUDA.fill(Int32(-1), L, N)
-    J = couplings(ig) + SpinGlassNetworks.Diagonal(biases(ig))
+    J = couplings(ig) + Diagonal(biases(ig))
     J_dev = CUDA.CuArray(J)
 
     k = 2
@@ -190,10 +182,10 @@ function partial_exhaustive_search(ig::IsingGraph)
     part_st = CUDA.zeros(2^(L-k))
     part_lst = CUDA.zeros(2^(L-k))
 
-    threadsPerBlock::Int64 = 2^k
-    blocksPerGrid::Int64 = 2^(L-k)
+    threads = 512
+    blocks = cld(L, threads)
 
-    @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel_part(J_dev, energies, part_lst, part_st)
+    @cuda blocks=blocks threads=threads kernel_part(J_dev, energies, part_lst, part_st)
 
 
     idx = sortperm(part_lst)
@@ -255,7 +247,7 @@ function exhaustive_search_bucket(ig::IsingGraph, how_many::Int = 8)
     L = SpinGlassNetworks.nv(ig)
 
     σ = CUDA.fill(Int32(-1), L, 2^L)
-    J = couplings(ig) + SpinGlassNetworks.Diagonal(biases(ig))
+    J = couplings(ig) + Diagonal(biases(ig))
     J_dev = CUDA.CuArray(J)
 
     N = size(J_dev,1)
@@ -271,17 +263,16 @@ function exhaustive_search_bucket(ig::IsingGraph, how_many::Int = 8)
     lowest_d = CUDA.zeros(Float64, how_many*2)
     lowest_states_d = CUDA.zeros(Int64, how_many*2)
 
-    k = 2
-    threadsPerBlock = 2^k
-    blocksPerGrid = 2^(chunk_size-k)
+    threads = 512
+    blocks = cld(N, threads)
 
 
 
     for i in 1:2^(N-chunk_size)
 
         idx = (i-1)*(2^chunk_size) + 1 
 
-        @cuda blocks=(blocksPerGrid) threads=(threadsPerBlock) kernel_bucket(J_dev, energies_d, idx)
+        @cuda blocks=blocks threads=threads kernel_bucket(J_dev, energies_d, idx)
 
         states_d = sortperm(energies_d)[1:how_many]
 

src/naive.jl

@@ -34,7 +34,7 @@ function SpinGlassNetworks.brute_force(
     ::Val{:GPU};
     num_states::Int=1
 )
-    L = nv(ig)
+    L = SpinGlassNetworks.nv(ig)
     N = 2^L
     σ = CUDA.fill(Int32(-1), L, N)
     JJ = couplings(ig)

src/utils.jl

@@ -1,12 +1,11 @@
-using Distributions
-export generate_random_graph
+export generate_random_graph, graph_to_dict, graph_to_qubo
 """
 $(SIGNATURES)
 - `d::Int`: size of random graph.
 Returns random array of size d.
 """
 function generate_random_graph(d::Int)
-    graph = rand(Uniform(-5,5),d, d)
+    graph = 10*(rand(d, d) .- 0.5)
     graph = graph * graph'
 
     graph
@@ -40,17 +39,18 @@ $(SIGNATURES)
 - `graph`: graph of the ising model.
 Converts Ising model graph to Dict.
 """
-function graph_to_dict(graph)
+function graph_to_dict(graph::Array)
     N = size(graph,1)
     graph_dict = Dict()
     for i in 1:N
-        graph_dict[(i, i)] = graph[i,i] #rand(Uniform(-5,5))
+        graph_dict[(i, i)] = graph[i, i]
         for j in i:N
-            graph_dict[(i, j)] = graph[i,j] #rand(Uniform(-5,5))
+            graph_dict[(i, j)] = graph[i,j]
         end
     end
     graph_dict
 end
+graph_to_dict(graph::CuArray) = graph_to_dict(Array(graph))
 
 """
 $(SIGNATURES)

test/brute_force.jl

@@ -1,7 +1,3 @@
-
-using SpinGlassNetworks
-using SpinGlassExhaustive
-
 function bench_cpu(instance::String, max_states::Int=100)
     m = 2
     n = 2