Make qos_fill and qos_stop local for each thread in sparsebytemap

Project.toml

@@ -63,4 +63,4 @@ julia = "1.10"
 HDF5 = "f67ccb44-e63f-5c2f-98bd-6dc0ccc4ba2f"
 NPZ = "15e1cf62-19b3-5cfa-8e77-841668bca605"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
-TensorMarket = "8b7d4fe7-0b45-4d0d-9dd8-5cc9b23b4b77"
+TensorMarket = "8b7d4fe7-0b45-4d0d-9dd8-5cc9b23b4b77"

src/architecture.jl

@@ -200,7 +200,7 @@
 
 struct CPUThread{Parent} <: AbstractTask
     tid::Int
-    dev::CPU
+    device::CPU
     parent::Parent
 end
 get_device(task::CPUThread) = task.device
@@ -220,8 +220,8 @@
 Base.ndims(::Type{CPULocalArray{A}}) where {A} = ndims(A)
 
 transfer(device::Union{CPUThread,CPUSharedMemory}, arr::AbstractArray) = arr
-function transfer(device::CPULocalMemory, arr::AbstractArray)
-    CPULocalArray{A}(mem.device, [copy(arr) for _ in 1:(mem.device.n)])
+function transfer(mem::CPULocalMemory, arr::AbstractArray)
+    CPULocalArray(mem.device, [copy(arr) for _ in 1:(mem.device.n)])
 end
 function transfer(task::CPUThread, arr::CPULocalArray)
     if get_device(task) === arr.device

src/postprocess.jl

@@ -323,6 +323,7 @@ end
 
 iseffectful(ex) = false
 function iseffectful(ex::Expr)
+    ex.head == :break && return true
     @capture(ex, :call(~f::(!ispure), ~args...)) && return true
     @capture(ex, :(=)(~lhs::issymbol, ~rhs)) && lhs != :_ && return true
     @capture(ex, :for(~lhs, ~rhs)) && return iseffectful(rhs) #TODO this could be handled better if we desugared :for

src/tensors/levels/sparse_bytemap_levels.jl

@@ -254,8 +254,8 @@ function distribute_level(
         distribute_buffer(ctx, lvl.tbl, arch, style),
         distribute_buffer(ctx, lvl.srt, arch, style),
         lvl.shape,
-        lvl.qos_fill,
-        lvl.qos_stop,
+        distribute_buffer(ctx, lvl.qos_fill, arch, style),
+        distribute_buffer(ctx, lvl.qos_stop, arch, style),
     )
 end